[feat] eval: async VideoPool + GPU-side common metrics + safe optical-flow chunks

Pipelines path → tensor decode behind GPU metric compute via a new
VideoPool, so multi-sample eval runs no longer serialize disk I/O and
metric work. The Evaluator owns one pool per evaluate(samples=...)
call; each EvalWorker is a single-GPU consumer that grabs decoded
samples from the shared queue (work-stealing across replicas when
num_gpus > 1).

Worker pre-uploads video/reference to its device once per sample so
every metric in the loop consumes the same GPU-resident tensor (no
per-metric .to(device) traffic).

SSIM and PSNR move to the GPU — at 1080p × 121 frames the CPU path
was both slow (5–10 s/pair) and contended with the loader thread for
DDR bandwidth. LPIPS gains a chunk_size knob (default 8) that caps
peak from ~60 GB to ~5 GB with bit-identical output. Optical-flow
metrics drop chunk_size to 1 because DPFlow's cost volume is ~4 GB
per frame pair at 1080p (matches mhuo/ptlflow upstream).

physics_iq and a handful of vbench metrics drop their list-batch
shims — the per-sample contract is now uniform across the suite.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.agents/scripts/sync-skills.sh

@@ -0,0 +1,96 @@
+#!/usr/bin/env bash
+# Sync .agents/skills/ into .claude/skills/ via per-skill symlinks.
+#
+# Why: Claude Code only scans .claude/skills/ and ~/.claude/skills/ for
+# user-invocable skills (no skillsPath config exists — see
+# https://code.claude.com/docs/en/skills.md). This repo's skills live
+# in .agents/skills/ so they travel with the repo and stay under git.
+# Run this once after cloning (or after adding/removing a skill) to
+# expose them to Claude Code without maintaining a parallel tree.
+#
+# Usage:
+#   .agents/scripts/sync-skills.sh
+#
+# Idempotent and safe to re-run. Prunes stale symlinks whose source
+# has been removed from .agents/skills/. Leaves hand-written
+# .claude/skills/<name>/ directories untouched (only symlinks are
+# managed).
+
+set -euo pipefail
+
+REPO_ROOT="$(git -C "$(dirname "$0")" rev-parse --show-toplevel)"
+SRC_DIR="$REPO_ROOT/.agents/skills"
+DST_DIR="$REPO_ROOT/.claude/skills"
+
+if [[ ! -d "$SRC_DIR" ]]; then
+    echo "Error: $SRC_DIR does not exist." >&2
+    exit 1
+fi
+
+mkdir -p "$DST_DIR"
+
+linked=0
+unchanged=0
+skipped=0
+pruned=0
+
+link_skill() {
+    local name="$1"
+    local src="$SRC_DIR/$name"
+    local dst="$DST_DIR/$name"
+    # Relative target keeps symlinks portable across clones.
+    local rel="../../.agents/skills/$name"
+
+    if [[ -L "$dst" ]]; then
+        if [[ "$(readlink "$dst")" == "$rel" ]]; then
+            unchanged=$((unchanged + 1))
+            return
+        fi
+        rm "$dst"
+    elif [[ -e "$dst" ]]; then
+        echo "Skipped (not a symlink): .claude/skills/$name" >&2
+        skipped=$((skipped + 1))
+        return
+    fi
+
+    ln -s "$rel" "$dst"
+    echo "Linked: .claude/skills/$name -> $rel"
+    linked=$((linked + 1))
+}
+
+prune_stale() {
+    local link="$1"
+    local target
+    target="$(readlink "$link")"
+    case "$target" in
+        ../../.agents/skills/*) ;;
+        *) return ;;
+    esac
+    local name="${target##*/}"
+    if [[ ! -d "$SRC_DIR/$name" ]]; then
+        rm "$link"
+        echo "Pruned stale: .claude/skills/$(basename "$link")"
+        pruned=$((pruned + 1))
+    fi
+}
+
+for src in "$SRC_DIR"/*/; do
+    [[ -d "$src" ]] || continue
+    name="$(basename "$src")"
+    # Only treat directories that actually contain a SKILL.md as skills.
+    [[ -f "$src/SKILL.md" ]] || continue
+    link_skill "$name"
+done
+
+shopt -s nullglob
+for link in "$DST_DIR"/*; do
+    [[ -L "$link" ]] || continue
+    prune_stale "$link"
+done
+shopt -u nullglob
+
+printf "\nSummary: %d linked, %d unchanged, %d pruned" "$linked" "$unchanged" "$pruned"
+if [[ "$skipped" -gt 0 ]]; then
+    printf ", %d skipped (non-symlink collision)" "$skipped"
+fi
+printf "\n"

.agents/skills/index.jsonl

@@ -5,3 +5,4 @@
 {"name": "evaluate-video-quality", "description": "Evaluate generated video quality using available metrics (SSIM, loss trajectory, caption consistency)", "path": "evaluate-video-quality/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "index-related-work", "description": "Ingest a paper or repository into the related work index", "path": "index-related-work/SKILL.md", "status": "draft", "trust": "low"}
 {"name": "search-related-work", "description": "Query the related work index for relevant papers, repos, or comparisons", "path": "search-related-work/SKILL.md", "status": "draft", "trust": "low"}
+{"name": "seed-ssim-references", "description": "Run a new or updated fastvideo/tests/ssim/ test on Modal, pull generated videos, and upload them to FastVideo/ssim-reference-videos so the test has a regression baseline", "path": "seed-ssim-references/SKILL.md", "status": "draft", "trust": "low"}

.agents/skills/seed-ssim-references/SKILL.md

@@ -0,0 +1,250 @@
+---
+name: seed-ssim-references
+description: Seed HF reference videos for a single newly-added SSIM test. Runs the test on Modal L40S, downloads the generated mp4s via `modal volume get`, pauses for the user to eyeball quality, then uploads only that test's files to `FastVideo/ssim-reference-videos`. Use when a new `fastvideo/tests/ssim/test_*_similarity.py` has just been added and has no references on HF yet.
+---
+
+# Seed SSIM Reference Videos
+
+## Purpose
+
+A brand-new SSIM test in `fastvideo/tests/ssim/` fails forever until its
+reference videos exist on the HF dataset (`FastVideo/ssim-reference-videos`).
+This skill:
+
+1. Runs the test on Modal's L40S pool to generate the videos.
+2. Downloads them to the local repo via `modal volume get`.
+3. Pauses so the user can eyeball the mp4s and confirm quality.
+4. Uploads only the new test's files to HF, with a guard that refuses to
+   overwrite anything already present.
+
+The skill is run **manually**, once per new test. Before invoking it, the user
+has already sanity-tested the new test locally — it launches `VideoGenerator`
+and writes an mp4 without crashing. The skill does not re-test locally; it
+goes straight to Modal L40S (which is what CI uses).
+
+## When to use
+
+- A new `test_*_similarity.py` file has been added in `fastvideo/tests/ssim/`
+  and the HF dataset has no `reference_videos/default/L40S_reference_videos/<model_id>/`
+  subtree for it yet.
+
+## When not to use
+
+- Regular CI runs — once refs exist, `pytest fastvideo/tests/ssim/` downloads
+  them automatically.
+- Re-seeding an existing test. That requires `--force` on the upload step, and
+  is out of scope here; treat as a separate, deliberate operation.
+
+## Inputs
+
+The skill has **one required input**: the path to the new SSIM test file.
+Prompt the user for it if they didn't supply it.
+
+| Parameter | Required | Description |
+|-----------|----------|-------------|
+| `test_file` | Yes | e.g. `fastvideo/tests/ssim/test_ltx2_similarity.py`. The skill's first action is to ask for this if missing. |
+
+Everything else is fixed:
+
+- Modal runner GPU: **L40S** (hardcoded in `fastvideo/tests/modal/ssim_test.py`).
+- Device folder: `L40S_reference_videos`.
+- Quality tier: `default` (the tier CI runs). The `full_quality` tier is not
+  seeded by this skill.
+- HF repo: `FastVideo/ssim-reference-videos` (dataset).
+- Multi-model test files: all model ids in `*_MODEL_TO_PARAMS` are seeded
+  together; the Modal run produces one mp4 per (model, prompt, backend) and
+  the upload scopes by `--model-id`, looping if there is more than one.
+
+## Prerequisites
+
+The user has confirmed:
+
+- `modal` CLI authenticated.
+- `HF_API_KEY` (or `HUGGINGFACE_HUB_TOKEN` / `HF_TOKEN`) exported with write
+  access to `FastVideo/ssim-reference-videos`.
+- The test file runs locally end-to-end (generates an mp4; SSIM assertion
+  failure due to missing reference is expected and fine).
+
+Fail fast if the token env var is missing.
+
+## Steps
+
+### 1. Ask for the test file
+
+If the user didn't name one, ask: *"Which SSIM test file do you want to seed
+references for? (e.g. `fastvideo/tests/ssim/test_ltx2_similarity.py`)"*.
+
+Validate:
+
+- Path exists and matches `fastvideo/tests/ssim/test_*_similarity.py`.
+- File defines a `*_MODEL_TO_PARAMS` dict — grep it to extract the set of
+  model ids. Those ids drive step 5.
+
+If either check fails, stop and tell the user what's wrong.
+
+### 2. Run the test on Modal L40S
+
+Pick a subdir name so repeated runs don't collide:
+
+```bash
+SHORT_COMMIT=$(git rev-parse --short=12 HEAD)
+TIMESTAMP=$(date -u +%Y%m%d_%H%M%S)
+SUBDIR="${TIMESTAMP}_${SHORT_COMMIT}"
+```
+
+Then launch the Modal run:
+
+```bash
+modal run fastvideo/tests/modal/ssim_test.py \
+    --git-repo="$(git config --get remote.origin.url)" \
+    --git-commit="$(git rev-parse HEAD)" \
+    --hf-api-key="$HF_API_KEY" \
+    --test-files="<test_file>" \
+    --sync-generated-to-volume \
+    --generated-volume-subdir="$SUBDIR" \
+    --skip-reference-download \
+    --no-fail-fast
+```
+
+Flag rationale:
+- `--skip-reference-download`: no refs exist yet, so conftest must not try to
+  pull them.
+- `--no-fail-fast`: lets the test finish generation before `_assert_similarity`
+  raises `FileNotFoundError: Reference video folder does not exist`. The
+  expected failure is what we want — the mp4 has already been written.
+- `--sync-generated-to-volume` + `--generated-volume-subdir`: copies the
+  generated mp4s to the `hf-model-weights` Modal volume under
+  `ssim_generated_videos/default/<SUBDIR>/generated_videos/` so we can pull
+  them locally.
+
+The Modal run will end with a nonzero exit (expected) and print a
+`modal volume get hf-model-weights ssim_generated_videos/default/<SUBDIR>/generated_videos ./generated_videos_modal/default`
+command. Capture that `<SUBDIR>` — you need it for step 3.
+
+### 3. Download generated videos locally
+
+```bash
+modal volume get --force hf-model-weights \
+    ssim_generated_videos/default/"$SUBDIR"/generated_videos \
+    ./generated_videos_modal/default
+```
+
+`--force` is required when the parent `./generated_videos_modal/default`
+already exists; without it, `modal volume get` errors with `[Errno 21] Is a
+directory`. Safe to pass on the first run too.
+
+After this, the mp4s live at
+`./generated_videos_modal/default/generated_videos/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
+The extra `generated_videos/` level comes from the volume layout in
+`_sync_generated_videos_to_volume` (`ssim_test.py`) — the command copies
+`<repo>/fastvideo/tests/ssim/generated_videos/<tier>` to
+`ssim_generated_videos/<tier>/<SUBDIR>/generated_videos/`, and `modal volume
+get` preserves that trailing `generated_videos/` segment.
+
+### 4. PAUSE — user reviews quality
+
+Print the list of downloaded mp4s and their paths, then stop. Tell the user:
+
+> "Generated videos downloaded to `./generated_videos_modal/default/generated_videos/L40S_reference_videos/`. Please open them and confirm the quality looks correct. Reply **`upload`** to continue, or anything else to abort."
+
+Do not proceed until the user explicitly says `upload`. If they abort, leave
+everything on disk so they can inspect further — no cleanup.
+
+### 5. Copy into the local reference layout
+
+Scoped copy — only the new test's mp4s. Loop over each `<model_id>` extracted
+in step 1:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py copy-local \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --generated-dir ./generated_videos_modal/default/generated_videos/L40S_reference_videos
+```
+
+(The `--generated-dir` points at the device-folder root inside the
+downloaded tree; `copy-local` walks all `<model>/<backend>/*.mp4`
+underneath it. Since the Modal run was scoped to a single test file via
+`--test-files`, only that test's model(s) are present — so the copy is
+implicitly per-test.)
+
+Result: `fastvideo/tests/ssim/reference_videos/default/L40S_reference_videos/<model_id>/<backend>/<prompt>.mp4`.
+
+### 6. Upload to HF — scoped per model_id, with overwrite guard
+
+For each `<model_id>`:
+
+```bash
+python fastvideo/tests/ssim/reference_videos_cli.py upload \
+    --quality-tier default \
+    --device-folder L40S_reference_videos \
+    --model-id "<model_id>"
+```
+
+The upload command:
+
+- Uploads **only** `reference_videos/default/L40S_reference_videos/<model_id>/`.
+- **Refuses** if any file already exists at that path on HF (this is the
+  guard — seeding a new test should never clobber existing refs). To override,
+  the user must re-run with `--force`. If the guard fires, stop and report
+  exactly which files exist; do not silently `--force`.
+
+Reads the HF token from `HF_API_KEY` / `HUGGINGFACE_HUB_TOKEN` / `HF_TOKEN`.
+
+### 7. Report success
+
+List what was uploaded (paths in repo) and remind the user to push any
+related code changes. Do **not** auto-verify by re-running Modal — the user
+can run `pytest fastvideo/tests/ssim/<test_file>` later to confirm end-to-end;
+it will auto-download the refs they just uploaded.
+
+## Failure modes and how to handle them
+
+- **`HF_API_KEY` unset.** Stop before step 2. The Modal run needs it (passed
+  via `--hf-api-key`), and step 6 needs it for upload.
+- **Modal run fails before generation.** No mp4s on the volume — nothing to
+  download. Fix the test locally (`pytest fastvideo/tests/ssim/<test_file>`)
+  and retry from step 2.
+- **`./generated_videos_modal/default/L40S_reference_videos/` missing after
+  `modal volume get`.** The run didn't produce videos (most likely the test
+  crashed before writing, or `REQUIRED_GPUS` exceeded the partition capacity
+  — see Modal logs).
+- **Upload guard fires (files already exist).** The test name / model id
+  collides with something already on HF. Verify the user actually wants to
+  replace existing refs; if so, re-run the upload with `--force`. If not,
+  rename the model id in `*_MODEL_TO_PARAMS` and re-seed.
+- **Quality looks wrong in step 4.** Abort. The mp4s stay on disk for
+  inspection. The fix is usually in the test's params (resolution, steps,
+  seed) — edit the test, then re-run the skill.
+
+## Design notes (for future skill maintainers)
+
+- The skill deliberately runs on Modal, **not** locally, because the CI
+  runner is L40S. Seeding from a different GPU SKU produces refs that CI's
+  L40S runs can't match (SSIM drifts across SKUs).
+- The skill is default-tier only. `full_quality` refs are seeded by a
+  separate, deliberate operation — they double runtime and aren't what CI
+  gates on.
+- The overwrite guard in `reference_videos_cli.py upload` is default-on
+  specifically because this skill exists. Re-seeding is a distinct operation
+  that requires explicit `--force`.
+
+## References
+
+- `fastvideo/tests/modal/ssim_test.py` — Modal orchestrator; see
+  `--sync-generated-to-volume`, `--generated-volume-subdir`,
+  `--skip-reference-download`, `--no-fail-fast`.
+- `fastvideo/tests/ssim/reference_videos_cli.py` — `copy-local`, `upload`
+  (with `--model-id`, `--force`), `download`, `ensure` subcommands.
+- `fastvideo/tests/ssim/README.md` — reference layout, HF repo conventions.
+- `fastvideo/tests/ssim/inference_similarity_utils.py` —
+  `run_text_to_video_similarity_test` + `_build_init_kwargs`: what each test
+  config passes to `VideoGenerator.from_pretrained`.
+
+## Changelog
+
+| Date | Change |
+|------|--------|
+| 2026-04-17 | Initial version (Modal sync-to-volume flow). |
+| 2026-04-21 | Rewrite: single-test scope, explicit user-review pause, per-`model_id` upload, HF overwrite guard. Dropped `scripts/seed_ssim.sh`. |
+| 2026-04-21 | Post-first-run fixes: `modal volume get` needs `--force` when parent exists; download tree has an extra `generated_videos/` level so `--generated-dir` must reflect it. |

.agents/workflows/evaluation-development.md

@@ -4,21 +4,24 @@ description: How to develop, validate, and register a new evaluation metric
 
 # Evaluation Development SOP
 
-Standard procedure for adding new video quality evaluation metrics to the
-FastVideo agent toolkit.
+Standard procedure for adding new video quality evaluation metrics to
+the FastVideo agent toolkit.
 
-## When to Use
+## When to use
 
-- You need a metric that doesn't exist in `.agents/memory/evaluation-registry/README.md`.
+- You need a metric that does not exist in
+  `.agents/memory/evaluation-registry/README.md`.
 - An existing metric needs significant changes to its methodology.
-- You're exploring a new evaluation approach.
+- You are exploring a new evaluation approach.
 
 ## Steps
 
 ### 1. Research
 
-- Search `.agents/memory/related-work/` for existing evaluation approaches.
-- Check the `evaluation_registry.md` for current metrics and their limitations.
+- Search `.agents/memory/related-work/` for existing evaluation
+  approaches.
+- Check `.agents/memory/evaluation-registry/README.md` for current
+  metrics and their limitations.
 - Review literature: FVD, CLIP-Score, human preference, etc.
 
 ### 2. Prototype
@@ -29,21 +32,25 @@ FastVideo agent toolkit.
 
 ### 3. Validate
 
-- **Known-good test**: Metric should score high on reference-quality videos.
-- **Known-bad test**: Metric should score low on degraded/unrelated videos.
-- **Sensitivity test**: Small quality differences should produce meaningful
-  score differences.
+- **Known-good test**: metric should score high on reference-quality
+  videos.
+- **Known-bad test**: metric should score low on degraded or unrelated
+  videos.
+- **Sensitivity test**: small quality differences should produce
+  meaningful score differences.
 - Document thresholds and their justification.
 
 ### 4. Register
 
 Update `.agents/memory/evaluation-registry/README.md`:
+
 - Add the metric with status `Active`.
 - Document location, thresholds, and trust level.
 
 ### 5. Integrate
 
-Update `.agents/skills/evaluate-video-quality.md`:
+Update `.agents/skills/evaluate-video-quality/SKILL.md`:
+
 - Add the new metric as a section.
 - Include code examples and interpretation guide.
 
@@ -52,3 +59,35 @@ Update `.agents/skills/evaluate-video-quality.md`:
 - Move the exploration log content into the skill.
 - Clean up the exploration file or mark it as `promoted`.
 - If anything went wrong during development, create a lesson.
+
+## Where the metrics live
+
+The eval suite is `fastvideo/eval/`. New metrics register themselves
+via `@register("<group>.<name>")` and are auto-discovered when
+`fastvideo.eval.metrics` is imported.
+
+- **Native metrics** (SSIM, PSNR, LPIPS, optical flow, VLM): add a
+  file under the appropriate group dir
+  (`fastvideo/eval/metrics/common/`, `optical_flow/`, `videoscore2/`,
+  `physics_iq/`).
+- **Metrics that wrap upstream research code**: follow the vbench
+  pattern in `fastvideo/eval/metrics/vbench/`. The contract is:
+  - Upstream lives as a git submodule under
+    `fastvideo/third_party/eval/<bench>/`, pinned to a SHA in repo-root
+    `.gitmodules`.
+  - The metric package's `__init__.py` inserts the submodule path on
+    `sys.path` and installs runtime compat shims (attribute-level
+    monkey-patches) for any modern-dep drift. Do not modify upstream
+    files on disk, and do not ship a `setup.sh`.
+  - See `fastvideo/eval/README.md` for the worked vbench example.
+  - Full porting guide:
+    [`docs/contributing/eval-metrics.md`](../../docs/contributing/eval-metrics.md).
+
+## Out of scope of the initial eval port
+
+The following land in follow-up PRs:
+
+- **MIND** metrics (depends on a separate `vipe` submodule).
+- **VBench-2.0** sibling package.
+- Native conversion of **FVD** under `fastvideo/eval/metrics/fvd/`.
+- The training-time `EvalCallback`.

.gitmodules

@@ -4,3 +4,6 @@
 [submodule "fastvideo-kernel/include/cutlass"]
 	path = fastvideo-kernel/include/cutlass
 	url = https://github.com/NVIDIA/cutlass.git
+[submodule "fastvideo/third_party/eval/vbench"]
+	path = fastvideo/third_party/eval/vbench
+	url = https://github.com/Vchitect/VBench.git

docs/contributing/eval-metrics.md

@@ -0,0 +1,559 @@
+# Porting Eval Metrics into `fastvideo.eval`
+
+This guide is for contributors adding new evaluation metrics to
+FastVideo's eval suite. To run the existing metrics, see
+[`fastvideo/eval/README.md`](../../fastvideo/eval/README.md).
+
+## When to use this guide
+
+Use this guide when you are:
+
+- Adding a new metric (native or wrapping a third-party library).
+- Porting a benchmark (e.g. VBench, MIND, EvalCrafter) whose Python
+  code needs to be importable from a pinned upstream.
+- Adding a new metric group (audio, vlm, etc.).
+
+## TL;DR
+
+Metrics are auto-discovered from
+`fastvideo/eval/metrics/<group>/<name>/metric.py`. Each declares itself
+with `@register("<group>.<name>")` and subclasses `BaseMetric`. Three
+recipes:
+
+1. **Native metric** (pure-PyTorch, no submodule). Drop a file,
+   declare deps, implement `compute(sample)`.
+2. **Library-wrapped metric** (CLIP, torch.hub, transformers, pyiqa).
+   Same as above, plus route the library's cache through
+   `get_cache_dir()` if it has a `download_root=` / `cache_dir=`
+   kwarg.
+3. **Upstream-submodule-wrapped metric** (vbench-style). Pin upstream
+   as a git submodule under `fastvideo/third_party/eval/<bench>/`. The
+   adapter `__init__.py` does the `sys.path` insert and any runtime
+   compat shims for modern dep versions. Patches live as Python in
+   that file rather than as on-disk patches to the submodule.
+
+The full recipes are below.
+
+---
+
+## 0) Layout and auto-discovery
+
+```
+fastvideo/eval/metrics/
+├── base.py                  # BaseMetric + lifecycle contract
+├── common/                  # group: SSIM, PSNR, LPIPS
+├── optical_flow/            # group: gt_optical_flow, synthetic_optical_flow
+├── vlm/                     # group: VideoScore-2
+├── physics_iq/              # group + sub-metrics
+└── vbench/                  # group: 16 sub-metrics
+    ├── __init__.py          # sys.path bootstrap + runtime compat shims
+    ├── _grit_helper.py      # shared upstream-touching helpers
+    └── <sub_metric>/metric.py
+```
+
+Auto-discovery (`fastvideo/eval/metrics/__init__.py`) walks each group
+dir and imports every `metric.py` it finds, which fires the
+`@register` decorators. Names starting with `_` are skipped. Use that
+prefix for shared helpers or vendored code that should not register
+itself.
+
+---
+
+## 1) The `BaseMetric` contract
+
+Every metric subclasses `fastvideo.eval.metrics.base.BaseMetric` and
+declares:
+
+```python
+class YourMetric(BaseMetric):
+    name: str = "common.your_metric"          # must match @register
+    requires_reference: bool = True           # needs sample["reference"]
+    higher_is_better: bool = True             # for ranking / aggregates
+    dependencies: list[str] = []              # importable module names;
+                                              # registry surfaces a clean
+                                              # ImportError if missing
+    needs_gpu: bool = False
+    backbone: str | None = None               # e.g. "clip_vit_l14"
+```
+
+You must implement:
+
+```python
+def compute(self, sample: dict) -> list[MetricResult]:
+    """sample['video'] is (1, T, C, H, W). Return a one-element list.
+
+    The leading 1 is preserved for forward-compat with batched eval;
+    today :class:`EvalWorker` always invokes metrics with B=1.
+    """
+```
+
+You may override:
+
+- `setup(self) -> None`. Eager model loading. Called once by
+  `create_evaluator`. Idempotent (re-entrant). Use the `if self._model
+  is not None: return` pattern.
+- `to(self, device)`. Move the metric and its submodels to `device`.
+
+If a required input is missing (e.g. an fps-aware metric called
+without `fps`), return `self._skip(sample, reason)` instead of
+raising.
+
+---
+
+## 2) Recipe A: native metric (no external deps)
+
+Smallest case. Pixel math, simple closed-form.
+
+```python
+# fastvideo/eval/metrics/common/your_metric/metric.py
+from __future__ import annotations
+import torch
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("common.your_metric")
+class YourMetric(BaseMetric):
+    name = "common.your_metric"
+    requires_reference = True
+    higher_is_better = True
+    needs_gpu = False
+    dependencies: list[str] = []  # nothing extra
+
+    def compute(self, sample: dict) -> list[MetricResult]:
+        gen, ref = sample["video"], sample["reference"]   # (B,T,C,H,W) each
+        per_video = ((gen - ref) ** 2).mean(dim=(1, 2, 3, 4)).sqrt()
+        return [
+            MetricResult(name=self.name, score=float(s), details={})
+            for s in per_video
+        ]
+```
+
+That is the whole recipe. Drop the file and the registry picks it up.
+
+---
+
+## 3) Recipe B: library-wrapped metric (CLIP, torch.hub, transformers, pyiqa)
+
+If your metric loads a backbone from a Python package, route the
+library at the eval cache so users get one knob (`FASTVIDEO_EVAL_CACHE`)
+to redirect everything.
+
+### Cache routing rules
+
+| Library | How to route | Location after redirect |
+|---|---|---|
+| `clip.load("ViT-X")` | pass `download_root=str(get_cache_dir() / "clip")` | `${FASTVIDEO_EVAL_CACHE}/clip/` |
+| `torch.hub.load(...)` | nothing; `TORCH_HOME` is redirected at `fastvideo.eval` import time | `${FASTVIDEO_EVAL_CACHE}/torch/hub/` |
+| `transformers.from_pretrained(...)` | nothing; leave HF's default cache (`~/.cache/huggingface/hub/`) so users dedupe with other ML projects | `~/.cache/huggingface/hub/` |
+| `huggingface_hub.snapshot_download` / `hf_hub_download` | use `ensure_checkpoint(...)` (it wraps these with filelock) | same as above |
+| `pyiqa.create_metric(...)` | no env var or kwarg honored; document in metric docstring | pyiqa-internal |
+| `lpips`, `ptlflow` | torch.hub-based, auto-redirected | `${FASTVIDEO_EVAL_CACHE}/torch/hub/` |
+| Raw URL (no HF Hub) | use `ensure_checkpoint(name, source="https://...")` | `${FASTVIDEO_EVAL_CACHE}/models/<name>` |
+| Dataset asset (raw video/mask/image) auto-fetched from a public bucket | download into `get_cache_dir() / "datasets" / "<bench>"`, mirroring upstream's relative layout. Vendor any small manifest (CSV/JSON ≤1 MB) under the metric folder so the dataset can be used without external setup. | `${FASTVIDEO_EVAL_CACHE}/datasets/<bench>/` |
+
+### Dataset assets: vendor the manifest, auto-fetch the rest
+
+If your metric ships with its own paired-reference dataset (Physics-IQ
+is the canonical example), follow this layout:
+
+- **Manifest** (CSV/JSON ≤1 MB): vendor it under
+  `fastvideo/eval/metrics/<bench>/_vendored/<manifest>.<ext>`, with a
+  sibling `_vendored/LICENSE` recording attribution and provenance.
+  The `_vendored/` subdir is the project-wide convention for
+  upstream-provenance files: it is auto-skipped by metric discovery
+  (the `_` prefix) and by codespell (one `*/_vendored/*` glob in
+  `[tool.codespell].skip`), so dropping in a new vendored file
+  requires no further config. Read the manifest from the dataset
+  module via a `Path(__file__)`-relative resolver. Mirror
+  `_VENDORED_DESCRIPTIONS_CSV` in
+  `fastvideo/eval/datasets/physics_iq.py`.
+- **Heavy assets** (videos, masks, images): do not vendor. Auto-fetch
+  on first miss into `get_cache_dir() / "datasets" / "<bench>"`,
+  mirroring upstream's relative directory layout one-for-one so a
+  pre-downloaded mirror at any path works as a drop-in
+  `dataset_root=`. Use atomic `.part` then final-rename to be safe
+  under concurrent SLURM ranks.
+- **Bucket override**: expose `FASTVIDEO_<BENCH>_BUCKET_URL` so users
+  with internal mirrors can redirect.
+- **Opt-out**: accept `auto_download: bool = True` in the dataset
+  constructor; on `False`, raise `FileNotFoundError` instead of
+  fetching. This covers air-gapped runs and CI.
+
+The end-state is `get_dataset("<bench>")` with no kwargs.
+
+### Example: CLIP backbone + LAION head
+
+```python
+# fastvideo/eval/metrics/your_group/your_metric/metric.py
+from __future__ import annotations
+import torch
+import torch.nn as nn
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("your_group.your_metric")
+class YourMetric(BaseMetric):
+    name = "your_group.your_metric"
+    requires_reference = False
+    needs_gpu = True
+    dependencies = ["clip"]      # "openai-clip" PyPI; importable as `clip`
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._clip = None
+        self._head = None
+
+    def setup(self) -> None:
+        if self._clip is not None:
+            return
+        import clip
+        from fastvideo.eval.models import ensure_checkpoint, get_cache_dir
+
+        # Backbone: route CLIP's cache through our root.
+        self._clip, _ = clip.load(
+            "ViT-L/14",
+            device=self.device,
+            download_root=str(get_cache_dir() / "clip"),
+        )
+        self._clip.eval()
+
+        # URL-fetched head: ensure_checkpoint downloads to
+        # ${FASTVIDEO_EVAL_CACHE}/models/ with filelock + atomic rename.
+        ckpt = ensure_checkpoint(
+            "your_head.pth",
+            source="https://example.com/path/to/your_head.pth",
+        )
+        self._head = nn.Linear(768, 1)
+        self._head.load_state_dict(
+            torch.load(ckpt, map_location="cpu", weights_only=True)
+        )
+        self._head.to(self.device).eval()
+
+    def to(self, device):
+        super().to(device)
+        if self._clip is not None:
+            self._clip = self._clip.to(self.device)
+        if self._head is not None:
+            self._head = self._head.to(self.device)
+        return self
+
+    def compute(self, sample: dict) -> list[MetricResult]:
+        ...
+```
+
+### Do not redirect other `~/.cache/...` dirs
+
+If a third-party library hard-codes `~/.cache/<lib>/` and offers no
+override, document the exception in the metric's docstring. Forcing
+redirection by setting `os.environ` or patching `os.path.expanduser`
+is fragile and breaks user expectations of where the library's cache
+lives.
+
+---
+
+## 4) Recipe C: upstream-submodule-wrapped metric (vbench pattern)
+
+Use this when the upstream benchmark ships Python code (`vbench/`,
+`MIND/`, etc.) that is not pip-installable cleanly. See
+`fastvideo/eval/metrics/vbench/__init__.py` for the worked example.
+
+### 4.1 Pin the upstream as a submodule
+
+```bash
+git submodule add <upstream-url> fastvideo/third_party/eval/<bench>
+cd fastvideo/third_party/eval/<bench>
+git checkout <pinned-sha>
+cd -
+git add .gitmodules fastvideo/third_party/eval/<bench>
+```
+
+The submodule pulls under the standard `git submodule update --init
+--recursive` flow that users already run for kernel deps.
+
+### 4.2 Bootstrap on `sys.path`
+
+```python
+# fastvideo/eval/metrics/<bench>/__init__.py
+from __future__ import annotations
+import sys
+from pathlib import Path
+
+# fastvideo/eval/metrics/<bench>/__init__.py → ../../../../third_party/eval/<bench>
+_UPSTREAM = Path(__file__).resolve().parents[3] / "third_party" / "eval" / "<bench>"
+if _UPSTREAM.is_dir() and str(_UPSTREAM) not in sys.path:
+    sys.path.insert(0, str(_UPSTREAM))
+```
+
+We do not `pip install` the upstream because its egg-link/.pth would
+just re-do this `sys.path.insert`, and skipping the install also skips
+the upstream's `setup.py` (which often gates on a specific CUDA
+version).
+
+### 4.3 Modern-dep compat: runtime shims
+
+Upstream code pinned to e.g. `transformers==4.33.2`, `numpy<2`
+typically breaks against modern versions in 3-4 known places (API
+renames). Fix those at import time, in the same `__init__.py`:
+
+```python
+def _install_compat_shims() -> None:
+    # Example: transformers.modeling_utils API moved.
+    try:
+        import transformers.modeling_utils as _mu
+        import transformers.pytorch_utils as _pu
+        for _n in ("apply_chunking_to_forward",
+                   "find_pruneable_heads_and_indices",
+                   "prune_linear_layer"):
+            if not hasattr(_mu, _n) and hasattr(_pu, _n):
+                setattr(_mu, _n, getattr(_pu, _n))
+    except ImportError:
+        pass
+
+    # Example: numpy.lib.function_base.disp removed in numpy>=2.
+    try:
+        import types, numpy.lib as _nl
+        if not hasattr(_nl, "function_base"):
+            _stub = types.ModuleType("numpy.lib.function_base")
+            _stub.disp = lambda *a, **k: None
+            sys.modules["numpy.lib.function_base"] = _stub
+            _nl.function_base = _stub
+    except ImportError:
+        pass
+
+_install_compat_shims()
+```
+
+For function-level patches that cannot be expressed as attribute
+writes (e.g. wrapping a model factory function), use a
+`sys.meta_path` finder that wraps the loader. See
+`_install_modeling_finetune_hook()` in
+`fastvideo/eval/metrics/vbench/__init__.py` for the pattern (about 30
+lines).
+
+Why shims rather than `git apply` patches: patches go stale when the
+upstream SHA changes; shims are versioned Python code in our repo,
+they are grep-able, and they only run if the targeted module is
+imported.
+
+### 4.4 Per-sub-metric files
+
+Each sub-metric is a normal `BaseMetric` subclass that imports from
+the upstream:
+
+```python
+# fastvideo/eval/metrics/<bench>/<sub>/metric.py
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("<bench>.<sub>")
+class YourSubMetric(BaseMetric):
+    ...
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        # The sys.path bootstrap fired when fastvideo.eval.metrics.<bench>
+        # was imported (which auto-discovery does before importing this
+        # sub-package). Upstream imports just work:
+        from <bench>.something import SomeModel
+        ...
+```
+
+### 4.5 Conditional registration when the upstream is missing
+
+If a user installed `fastvideo[eval]` but did not run `git submodule
+update --init`, `<bench>.*` metrics should not register. The
+auto-discovery walker imports each sub-package's `metric` module; have
+that import bail out cleanly:
+
+```python
+# fastvideo/eval/metrics/<bench>/__init__.py — at the bottom
+_AVAILABLE = (_UPSTREAM / "<bench>" / "__init__.py").is_file()
+```
+
+```python
+# fastvideo/eval/metrics/<bench>/<sub>/__init__.py
+from fastvideo.eval.metrics.<bench> import _AVAILABLE
+if _AVAILABLE:
+    from .metric import YourSubMetric  # noqa
+```
+
+`fastvideo eval list` then reflects what the user actually has rather
+than what they could have.
+
+### 4.6 Leave the upstream alone unless it blocks the metric
+
+The upstream is pinned. If a metric works against the pinned SHA,
+leave the upstream files untouched. If it actively breaks against
+modern deps (the import-drift cases above), shim it. Avoid
+fastvideo-side forks of upstream code; they make patches go stale and
+parity drift.
+
+---
+
+## 5) Model checkpoints: `ensure_checkpoint`
+
+Use `ensure_checkpoint(name, source, filename=None)` for any
+non-package weights. It resolves a local path, downloading on miss,
+with filelock safety across processes and SLURM ranks.
+
+| `source` form | What happens |
+|---|---|
+| `"/abs/path/to/file.pth"` | passthrough, returned unchanged |
+| `"https://..."` | downloaded to `${FASTVIDEO_EVAL_CACHE}/models/<name>` via `huggingface_hub.http_get`, atomic rename, filelock |
+| `"org/repo"` (no `filename`) | `snapshot_download(repo_id)` → `~/.cache/huggingface/hub/` |
+| `"org/repo"` (with `filename`) | `hf_hub_download(repo_id, filename)` → `~/.cache/huggingface/hub/` |
+
+`name` is only used as the local filename for URL sources. HF sources
+ignore it (HF manages its own cache key by content hash).
+
+```python
+from fastvideo.eval.models import ensure_checkpoint
+
+# URL: name matters
+ckpt = ensure_checkpoint(
+    "amt-s.pth",
+    source="https://huggingface.co/lalala125/AMT/resolve/main/amt-s.pth",
+)
+
+# HF single file: name is decorative
+ckpt = ensure_checkpoint(
+    "raft-things.pth",                      # ignored; HF cache uses repo+sha
+    source="OpenGVLab/VBench_Used_Models",
+    filename="raft-things.pth",
+)
+```
+
+---
+
+## 6) Declaring `dependencies`
+
+Set `dependencies = ["pkg1", "pkg2"]` on your metric class with
+importable module names (not PyPI distribution names). The registry
+checks each via `importlib.util.find_spec` at instantiation time and
+raises a clean `ImportError` pointing the user at the right install
+extra:
+
+```python
+class YourMetric(BaseMetric):
+    dependencies = ["clip", "timm"]   # importable as `import clip`, `import timm`
+```
+
+If a dep is in `[project.optional-dependencies.eval-<group>]`, you do
+not need to do anything more. If it is a new dep, add it to that
+group in `pyproject.toml`.
+
+---
+
+## 7) Common gotchas
+
+- The standard `git submodule update --init --recursive` is enough
+  for a benchmark; do not write a `setup.sh`. Modern-dep compat goes
+  into your `__init__.py` as runtime shims.
+- Do not modify upstream files on disk. The submodule should always
+  match its pinned SHA. Compat lives in our `__init__.py`.
+- Do not pip-install the upstream. The egg-link is a glorified
+  `sys.path.insert`, which we do directly in `__init__.py`.
+- Do not call `torch.hub.set_dir(...)` from your metric. It is done
+  globally in `fastvideo/eval/__init__.py`.
+- Do not put cache-redirection env vars in your metric's `setup()`.
+  By the time `setup()` runs, the library has likely already cached
+  the default-location decision. Set env vars at package-init time.
+- Skip rather than raise when an input is missing. Use
+  `self._skip(sample, reason)` for any expected-missing input. It
+  returns a list of `MetricResult(score=None)` so other metrics in
+  the same evaluator continue.
+- Watch for upstream re-registration conflicts. If the upstream uses
+  a global registry (detectron2's `META_ARCH_REGISTRY`, MMCV, etc.),
+  loading the same model twice in the same process will throw. The
+  evaluator already loads each metric once; if you write a custom
+  setup-then-call pattern, mirror that single-load discipline.
+
+---
+
+## 8) Training-time eval: keep evaluators hot, free caches between calls
+
+When wiring eval into a training loop, the working pattern is:
+
+1. Construct the `Evaluator` once and attach it to the pipeline
+   (`self._eval = create_evaluator(...)`). Do not recreate it per
+   validation round; that re-pays the model load cost.
+2. Save validation videos to disk (the diffusion path already does
+   this). Pass paths to `evaluator.evaluate`, not in-memory tensors
+   that share GPU memory with the training model.
+3. Run validation only on rank 0 of each sequence-parallel group.
+   Gather paths from other ranks and let rank 0 score everything.
+4. After every `evaluate(...)` call, call
+   `evaluator.release_cuda_memory()` in a `finally` block. That runs
+   `gc.collect()` + `torch.cuda.empty_cache()` +
+   `torch.cuda.ipc_collect()`. The eval model stays loaded; only
+   transient activation buffers from the just-finished call get
+   freed:
+
+   ```python
+   for video_path in batch:
+       try:
+           scores = self._eval.evaluate(video=load_video(video_path))
+       finally:
+           self._eval.release_cuda_memory()
+   ```
+
+5. If memory pressure spikes (rare on H200), call
+   `evaluator.unload()` to drop every metric reference and let the
+   GPU memory be GC'd. `unload` is reversible:
+   `evaluator.reload()` rebuilds the same metrics with the original
+   config (re-paying the model load cost). Calling `evaluate`
+   between `unload` and `reload` raises a clear `RuntimeError`.
+
+For most metrics (sub-1 GB backbones, e.g. CLIP/DINO/RAFT/AMT) the
+eval model can stay co-resident with the training model in
+`transformer.eval()` mode without any swap. For larger ones
+(VideoScore2 at 14 GB), measure first; if it fits on the rank-0 GPU
+during validation (training model in eval mode means no
+grads/optimizer updates), keep it hot. If not, `unload` between
+rounds.
+
+## 9) Local verification
+
+Native and library-wrapped metrics: a single-GPU smoke is enough.
+
+```python
+import torch
+from fastvideo.eval import create_evaluator
+
+ev = create_evaluator(metrics=["<group>.<your_metric>"], device="cuda")
+video = torch.randn(1, 49, 3, 256, 256, device="cuda").clamp(0, 1)
+print(ev.evaluate(video=video))
+```
+
+Submodule-wrapped metrics: also do a parity check against the
+upstream once. Clone upstream into a separate venv, run the same
+video through both, and expect an exact match on bit-deterministic
+metrics and ≤1% drift on backbone-heavy ones (driven by
+transformers/torch version differences).
+
+For quick parity in CI: pin a tiny test video, record expected
+scores ± tolerance, and add a calibration test under
+`fastvideo/tests/eval/`.
+
+---
+
+## 10) When not to add a metric
+
+- **Set-vs-set distribution metrics** (FVD, FID-style) do not fit
+  `BaseMetric.compute(sample)` cleanly; they need a population.
+  Adding them requires a stateful accumulator interface that does
+  not exist yet. Open an issue first.
+- **Metrics requiring a single-GPU model larger than available
+  memory.** Eval is not the place for tensor-parallel sharding;
+  metrics are expected to fit on one GPU.
+- **Metrics that need `mmcv` with a conflicting CUDA ABI.** Document
+  the affected sub-metrics as unsupported and skip them. Building
+  isolation infrastructure (subprocess engine, per-metric venv) is
+  out of scope.

docs/design/inference_schema_parity_inventory.yaml

@@ -29,7 +29,7 @@ surfaces:
       vae_cpu_offload: generator.engine.offload.vae
       pin_cpu_memory: generator.engine.offload.pin_cpu_memory
       enable_torch_compile: generator.engine.compile.enabled
-      torch_compile_kwargs: generator.engine.compile.kwargs
+      torch_compile_kwargs: generator.engine.compile.backend,fullgraph,mode,dynamic,extras
       disable_autocast: generator.engine.disable_autocast
       enable_stage_verification: generator.engine.enable_stage_verification
       prompt_txt: request.inputs.prompt_path
@@ -40,8 +40,8 @@ surfaces:
       init_weights_from_safetensors_2: generator.pipeline.components.transformer_2_weights
       override_pipeline_cls_name: generator.pipeline.components.override_pipeline_cls_name
       boundary_ratio: request.sampling.boundary_ratio
+      ltx2_vae_tiling: generator.pipeline.vae_tiling
     preset_owned:
-      ltx2_vae_tiling: generator.pipeline.preset_overrides.ltx2.vae_tiling
       ltx2_vae_spatial_tile_size_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_size_in_pixels
       ltx2_vae_spatial_tile_overlap_in_pixels: generator.pipeline.preset_overrides.ltx2.vae.spatial_tile_overlap_in_pixels
       ltx2_vae_temporal_tile_size_in_frames: generator.pipeline.preset_overrides.ltx2.vae.temporal_tile_size_in_frames
@@ -354,6 +354,8 @@ surfaces:
       return_frames: request.output.return_frames
       return_trajectory_latents: request.runtime.return_trajectory_latents
       return_trajectory_decoded: request.runtime.return_trajectory_decoded
+      continuation_state: request.state
+      return_continuation_state: request.output.return_state
     preset_owned:
       t_thresh: request.stage_overrides.refine.t_thresh
       spatial_refine_only: request.stage_overrides.refine.spatial_refine_only

docs/design/server_contracts/streaming.md

@@ -0,0 +1,177 @@
+# Streaming WebSocket Server Contract
+
+The streaming server (`fastvideo/entrypoints/streaming/server.py`) speaks
+a JSON-over-WebSocket protocol with binary fMP4 chunks for media. This
+document is the authoritative spec for the message catalogue and the
+session state machine. Any change to either must update this document
+in the same PR that touches `protocol.py` or `session.py`.
+
+## Endpoint
+
+| Path | Protocol | Purpose |
+|---|---|---|
+| `WS /v1/stream` | WebSocket (JSON + binary) | Per-session realtime streaming |
+| `GET /health` | HTTP | Liveness probe (`status`, `stream_mode`, active `sessions`) |
+
+The server is launched by `fastvideo serve --config <serve.yaml>` when
+the config carries a `streaming:` block. Without that block the same CLI
+launches the OpenAI stateless HTTP server instead.
+
+## Connection lifecycle
+
+Every WebSocket connection holds exactly one `Session`. Sessions move
+through the states in `SessionState` (`fastvideo/entrypoints/streaming/session.py`).
+
+```
+                    ┌──────────────┐
+                    │ INITIALIZING │  ← WebSocket accepted, before init frame
+                    └──────┬───────┘
+                           │ session_init_v2 received
+            ┌──────────────┼──────────────┐
+            ▼              ▼              ▼
+         QUEUED      GPU_BINDING       REJECTED
+            │              │              ↑
+            │ slot ready   │              │ max-sessions hit
+            ▼              ▼              │ or invalid init
+                       ┌────────┐         │
+                       │ ACTIVE │ ────────┘
+                       └────┬───┘
+              segment loop  │
+                            │
+                ┌───────────┼───────────┐
+                ▼           ▼           ▼
+            COMPLETE      ERROR      TIMEOUT
+        (clean leave)  (any failure)  (idle / segment_cap reached)
+```
+
+Terminal states (`COMPLETE`, `ERROR`, `TIMEOUT`, `REJECTED`) are sinks —
+no transitions out. The transition matrix is enforced in
+`session.py::_VALID_TRANSITIONS`; bad transitions raise.
+
+`SessionManager` enforces the per-process budgets pulled from
+`StreamingConfig`:
+
+- `session_timeout_seconds` — idle reaper drops sessions that haven't
+  advanced; non-terminal sessions transition to `TIMEOUT`.
+- `generation_segment_cap` — a session that hits the cap transitions to
+  `COMPLETE` after the last segment ships.
+
+## Message catalogue
+
+Every JSON frame carries `{"type": <str>, ...}`. Pydantic models in
+`protocol.py` are the source of truth; this table is the human-readable
+view.
+
+### Client → server
+
+| `type` | Required fields | Purpose |
+|---|---|---|
+| `session_init_v2` | — | Opening frame. Carries preset, curated prompts, optional initial image, feature toggles, optional `continuation_state` to resume from a snapshot. |
+| `segment_prompt_source` | `prompt` | Request the next segment using the supplied prompt; optional sampling overrides (`seed`, `num_inference_steps`, `guidance_scale`, `negative_prompt`). |
+| `seed_prompts_updated` | `seed_prompts` | Replace the session's seed-prompt list; takes effect on the next segment. |
+| `enhancement_updated` | `enabled` | Toggle prompt enhancement for subsequent segments. |
+| `auto_extension_updated` | `enabled` | Toggle automatic per-segment prompt extension. |
+| `loop_generation_updated` | `enabled` | Toggle loop-generation mode. |
+| `generation_paused_updated` | `paused` | Pause/resume segment generation; queued requests defer. |
+| `snapshot_state` | — | Request the current `ContinuationState` for export; server replies with `continuation_state_snapshot`. |
+
+The opening frame must be `session_init_v2`. Any other first frame is
+rejected with an `error` (code `invalid_message`) and the WebSocket is
+closed.
+
+### Server → client
+
+| `type` | Carries | When emitted |
+|---|---|---|
+| `queue_status` | `position`, `queue_depth` | After `session_init_v2` accepted, before GPU binding. |
+| `gpu_assigned` | GPU id, model id | Once a generator slot is bound. |
+| `ltx2_stream_start` | session-level metadata | Once the session enters `ACTIVE`. |
+| `ltx2_segment_start` | `segment_idx`, `prompt`, prompt source | When a `segment_prompt_source` request begins generation. |
+| `step_complete` | `segment_idx`, denoise timings | After the segment's denoising loop finishes (before media emission). |
+| `media_init` | `segment_idx`, mime, stream id | First frame of fMP4 output for the segment. |
+| binary frame | fMP4 fragment bytes | Subsequent media chunks; the protocol enforces that `media_init` precedes any binary frames. |
+| `media_segment_complete` | `segment_idx`, chunk count, byte count | Last media chunk for the segment. |
+| `ltx2_segment_complete` | `segment_idx`, segment summary | Segment fully shipped; ready for the next `segment_prompt_source`. |
+| `ltx2_stream_complete` | session summary | Session reached `generation_segment_cap` or client requested clean shutdown. |
+| `session_timeout` | reason | Session hit `session_timeout_seconds`; immediately followed by close. |
+| `continuation_state_snapshot` | `kind`, `payload` | Reply to `snapshot_state`. The payload is the same shape produced by `LTX2ContinuationState.to_continuation_state(...)`. |
+| `error` | `code`, `message` | Any validation/runtime error. Non-fatal errors keep the connection open; fatal errors precede a `close`. |
+
+## Continuation state
+
+The session optionally accepts a `continuation_state` dict inside the
+opening `session_init_v2` frame. When present, the server hydrates it
+into a `ContinuationState(kind, payload)` envelope and feeds it as the
+`request.state` on the first segment's `GenerationRequest` — letting a
+client resume after a disconnect, migrate sessions across processes,
+or replay a prior session.
+
+After every segment, if the runtime returns a fresh state, the server
+persists it to the `SessionStore` so a `snapshot_state` request can
+export it. The store and serialization contracts live with the model
+family (e.g. `fastvideo/pipelines/basic/ltx2/continuation.py` for LTX-2).
+
+## Example flow
+
+```
+client                                                    server
+──────                                                    ──────
+WS /v1/stream  ─────── connect ─────────────────────────►
+               ◄────── (accept)
+
+{"type": "session_init_v2",
+ "preset": "ltx2_two_stage",
+ "curated_prompts": ["a fox in snow", "the fox jumps"],
+ "initial_image": {...},
+ "stream_mode": "av_fmp4"} ─────────────────────────────►
+
+                                                          (validate, queue, bind)
+               ◄──── {"type": "queue_status",
+                       "position": 0, "queue_depth": 0}
+               ◄──── {"type": "gpu_assigned",
+                       "gpu_id": 0, "model_id": "..."}
+               ◄──── {"type": "ltx2_stream_start", ...}
+
+{"type": "segment_prompt_source",
+ "prompt": "a fox in snow",
+ "source": "curated"} ───────────────────────────────────►
+                                                          (run pipeline)
+               ◄──── {"type": "ltx2_segment_start",
+                       "segment_idx": 1, ...}
+               ◄──── {"type": "step_complete",
+                       "segment_idx": 1, "timings": {...}}
+               ◄──── {"type": "media_init",
+                       "segment_idx": 1,
+                       "mime": "video/mp4", ...}
+               ◄──── <binary fMP4 init segment>
+               ◄──── <binary fMP4 fragment>
+               ◄──── <binary fMP4 fragment>
+               ◄──── {"type": "media_segment_complete",
+                       "segment_idx": 1, "chunks": 12}
+               ◄──── {"type": "ltx2_segment_complete",
+                       "segment_idx": 1, ...}
+
+{"type": "segment_prompt_source",
+ "prompt": "the fox jumps"} ─────────────────────────────►
+                                                          (segment 2 …)
+
+{"type": "snapshot_state"} ──────────────────────────────►
+               ◄──── {"type": "continuation_state_snapshot",
+                       "kind": "ltx2.v1",
+                       "payload": {"schema_version": 1, ...}}
+
+(close) ──────────────────────────────────────────────────►
+                                                          (session → COMPLETE)
+```
+
+## Backward / forward compatibility
+
+- Adding a new client message: append a Pydantic model to `protocol.py`
+  with a unique `type`; add the discriminator entry to `ClientMessage`;
+  add a row to the table above. Old clients that don't send the new
+  message remain compatible.
+- Adding a new server message: emit only when a new feature flag is
+  enabled (or always emit, since clients ignore unknown types).
+- Changing an existing message: bump the `type` (e.g. `session_init_v2`
+  → `session_init_v3`) and accept both for one release cycle. Never
+  silently change field semantics under the same `type`.

docs/distillation/dmd.md

@@ -86,3 +86,25 @@ sbatch examples/distill/Wan2.2-TI2V-5B-Diffusers/Data-free/distill_dmd_t2v_5B.sh
 - Learning rate: 2e-5
 - Training steps: 3000 (~12 hours)
 - HSDP shard dim: 1
+
+## 🧭 Note on `real_score_guidance_scale`
+
+The teacher CFG used inside the DMD loss follows the DMD2 reference
+implementation and uses the parameterization
+
+```
+x = x_cond + w * (x_cond - x_uncond)
+```
+
+rather than the Ho & Salimans form `x_uncond + w * (x_cond - x_uncond)`. The
+two are mathematically equivalent up to a constant offset:
+
+| `real_score_guidance_scale` (`w`) | Equivalent standard CFG (`w + 1`) | Output                |
+|-----------------------------------|-----------------------------------|-----------------------|
+| `-1`                              | `0`                               | unconditional         |
+| `0`                               | `1`                               | conditional           |
+| `3.5` (default)                   | `4.5`                             | strong guidance       |
+
+So `real_score_guidance_scale` should be read as the **extra** guidance
+strength added on top of the conditional prediction. When porting values
+from a paper that uses the Ho & Salimans form, subtract 1.

examples/inference/eval/basic_ltx2_eval.py

@@ -0,0 +1,100 @@
+"""Generate one LTX2 video and score it with VBench metrics.
+
+The generation block is the same as
+``examples/inference/basic/basic_ltx2.py`` — same prompt, same model,
+same shape, same num_frames. After ``shutdown()`` the script loads the
+mp4 back, builds a single :class:`fastvideo.eval.Evaluator`, and runs
+the prompt-aware VBench subset that's meaningful for an arbitrary
+text→video sample.
+
+The first run downloads CLIP / DINO / RAFT / AMT / ViCLIP / MUSIQ
+weights to ``~/.cache/fastvideo/eval/`` (~few GB total).
+
+GPU memory caveat
+-----------------
+Scoring 1088×1920×121 with all 8 metrics needs a dedicated GPU (~80 GB).
+On a shared GPU, ``vbench.motion_smoothness`` (AMT correlation volume)
+will OOM — its memory autoscale reads ``total_memory`` rather than
+``mem_get_info()`` free memory and therefore underestimates the
+required scale-down. Drop ``motion_smoothness`` from ``METRICS`` if
+sharing, or run on a smaller-resolution generation.
+"""
+import torch
+
+from fastvideo import VideoGenerator
+from fastvideo.eval import Evaluator
+from fastvideo.eval.io import build_eval_kwargs
+
+PROMPT = (
+    "A warm sunny backyard. The camera starts in a tight cinematic close-up "
+    "of a woman and a man in their 30s, facing each other with serious "
+    "expressions. The woman, emotional and dramatic, says softly, \"That's "
+    "it... Dad's lost it. And we've lost Dad.\" The man exhales, slightly "
+    "annoyed: \"Stop being so dramatic, Jess.\" A beat. He glances aside, "
+    "then mutters defensively, \"He's just having fun.\" The camera slowly "
+    "pans right, revealing the grandfather in the garden wearing enormous "
+    "butterfly wings, waving his arms in the air like he's trying to take "
+    "off. He shouts, \"Wheeeew!\" as he flaps his wings with full commitment. "
+    "The woman covers her face, on the verge of tears. The tone is deadpan, "
+    "absurd, and quietly tragic."
+)
+
+# VBench sub-metrics meaningful for an arbitrary text→video sample
+# (just the generated frames, optionally fps + the source prompt).
+# Structured-prompt metrics (vbench.color, vbench.multiple_objects,
+# vbench.scene, ...) are excluded — they need prompts built to a
+# specific schema.
+METRICS = [
+    "vbench.aesthetic_quality",       # CLIP + LAION aesthetic head
+    "vbench.subject_consistency",     # DINO frame-to-first cosine
+    "vbench.background_consistency",  # DINO on background patches
+    "vbench.imaging_quality",         # pyiqa MUSIQ
+    "vbench.temporal_flickering",     # pixel-wise frame deltas
+    "vbench.motion_smoothness",       # AMT frame interpolator residual
+    "vbench.dynamic_degree",          # RAFT optical-flow magnitude (needs fps)
+    "vbench.overall_consistency",     # ViCLIP video↔prompt similarity
+]
+
+
+def main() -> None:
+    # ----- generation (matches examples/inference/basic/basic_ltx2.py) -----
+    generator = VideoGenerator.from_pretrained(
+        "Davids048/LTX2-Base-Diffusers",
+        num_gpus=1,
+    )
+
+    output_path = "outputs_video/ltx2_basic/output_ltx2_base_t2v_1088_1920_1.1.mp4"
+    generator.generate_video(
+        prompt=PROMPT,
+        output_path=output_path,
+        save_video=True,
+        num_frames=121,
+        height=1088,
+        width=1920,
+    )
+    generator.shutdown()
+    # Free residual CUDA memory the generator left behind so the
+    # evaluator can grab the largest possible workspace for AMT/RAFT.
+    torch.cuda.empty_cache()
+
+    # ----- scoring -----
+    print(f"\n[eval] building evaluator: {METRICS}")
+    evaluator = Evaluator(metrics=METRICS)
+
+    # LTX2 outputs at 24 fps by default.
+    sample = build_eval_kwargs({"prompt": PROMPT}, output_path, fps=24.0)
+    print(f"[eval] running ({sample['video'].shape[1]} frames @ 24 fps)...")
+    results = evaluator.evaluate(**sample)
+
+    print("\n=== VBench scores ===")
+    for name in METRICS:
+        r = results[name]
+        if r.score is None:
+            reason = r.details.get("skipped", "no score")
+            print(f"  {name}: SKIPPED ({reason})")
+        else:
+            print(f"  {name}: {r.score:.4f}")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/eval/bench_physics_iq.py

@@ -0,0 +1,157 @@
+"""End-to-end Physics-IQ: dataset → generate → score → aggregate.
+
+Generates one video per take-1 scenario with LTX2 (using the scenario
+caption as the prompt), scores each generated video against the take-1
+reference and the take-2 "physical-variance" reference, and prints
+aggregate scores using :meth:`PhysicsIQMetric.aggregate_components` —
+the official scoring recipe from the upstream benchmark.
+
+Reference videos / masks / switch-frames auto-fetch on first miss into
+``${FASTVIDEO_EVAL_CACHE}/datasets/physics_iq/``; pass ``--dataset-root``
+to point at a pre-downloaded mirror instead.
+
+Quick smoke run on 4 scenarios across 2 GPUs::
+
+    python examples/inference/eval/bench_physics_iq.py \\
+        --limit 4 --num-gpus 2 \\
+        --videos-dir outputs_video/physics_iq_smoke
+
+Re-score existing generations without regenerating::
+
+    python examples/inference/eval/bench_physics_iq.py \\
+        --videos-dir outputs_video/physics_iq_smoke \\
+        --skip-generation
+"""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from fastvideo.eval import create_evaluator, get_metric
+from fastvideo.eval.datasets import get_dataset
+
+
+def _expected_filename(row: dict) -> str:
+    """Filename Physics-IQ expects for the generated video for *row*.
+
+    Uses the dataset's own ``expected_gen_filename`` annotation so the
+    output filenames match the benchmark's manifest convention.
+    """
+    return row["auxiliary_info"]["expected_gen_filename"]
+
+
+def _generate_videos(rows: list[dict], videos_dir: Path,
+                     model: str, num_gpus: int,
+                     num_frames: int, height: int, width: int) -> None:
+    from fastvideo import VideoGenerator
+
+    videos_dir.mkdir(parents=True, exist_ok=True)
+    todo = [(row, videos_dir / _expected_filename(row)) for row in rows]
+    todo = [(row, out) for (row, out) in todo if not out.is_file()]
+    if not todo:
+        print(f"[gen] all {len(rows)} videos already present; skipping.")
+        return
+
+    print(f"[gen] {len(todo)}/{len(rows)} scenarios to render with {model} "
+          f"({num_frames}x{height}x{width})...")
+    gen = VideoGenerator.from_pretrained(model, num_gpus=num_gpus)
+    try:
+        for row, out_path in todo:
+            gen.generate_video(
+                prompt=row["prompt"], output_path=str(out_path), save_video=True,
+                num_frames=num_frames, height=height, width=width,
+            )
+    finally:
+        gen.shutdown()
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__,
+                                formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--dataset-root", type=Path, default=None,
+                   help="Path to a pre-downloaded Physics-IQ release. "
+                        "Defaults to ${FASTVIDEO_EVAL_CACHE}/datasets/physics_iq, "
+                        "auto-fetching missing assets from the public bucket.")
+    p.add_argument("--videos-dir", type=Path,
+                   default=Path("outputs_video/bench_physics_iq"),
+                   help="Where to read/write generated videos.")
+    p.add_argument("--limit", type=int, default=None,
+                   help="Truncate to first N scenarios for smoke runs.")
+    p.add_argument("--num-gpus", type=int, default=1)
+    p.add_argument("--model", default="Davids048/LTX2-Base-Diffusers",
+                   help="HF repo id of the text→video generator to use.")
+    p.add_argument("--num-frames", type=int, default=121)
+    p.add_argument("--height", type=int, default=1088)
+    p.add_argument("--width", type=int, default=1920)
+    p.add_argument("--skip-generation", action="store_true",
+                   help="Re-score existing videos under --videos-dir.")
+    p.add_argument("--scores-out", type=Path, default=None,
+                   help="Where to write per-scenario scores (JSON). "
+                        "Defaults to <videos-dir>/scores.json.")
+    args = p.parse_args()
+
+    # 1. Walk the Physics-IQ corpus. Pass --limit to the dataset
+    # constructor so auto-download only fetches the assets we'll use.
+    ds = get_dataset("physics_iq", dataset_root=args.dataset_root, limit=args.limit)
+    rows = list(ds)
+    print(f"[load] Physics-IQ: {len(rows)} scenarios from {ds.dataset_dir}")
+
+    # 2. Generate (or reuse) one mp4 per scenario.
+    if not args.skip_generation:
+        _generate_videos(
+            rows, args.videos_dir, args.model, args.num_gpus,
+            args.num_frames, args.height, args.width,
+        )
+
+    # 3. Score each scenario. The metric reads file paths directly out
+    #    of the row dict (reference, reference_take2, masks), so we
+    #    just attach the generated video path and forward.
+    evaluator = create_evaluator(metrics=["physics_iq"], num_gpus=args.num_gpus)
+
+    samples: list[dict] = []
+    matched: list[dict] = []
+    for row in rows:
+        video_path = args.videos_dir / _expected_filename(row)
+        if not video_path.is_file():
+            print(f"[eval] missing {video_path}; skipping.")
+            continue
+        # The physics_iq metric accepts file paths via its polymorphic
+        # input handling — no need to load the tensors here.
+        samples.append({"video": str(video_path), **row})
+        matched.append(row)
+
+    all_results = evaluator.evaluate(samples=samples)
+    evaluator.shutdown()
+
+    # 4. Aggregate per the upstream scoring recipe.
+    metric = get_metric("physics_iq")
+    components = metric.aggregate_components(
+        [r["physics_iq"] for r in all_results]
+    )
+
+    print()
+    print("=== Physics-IQ aggregate ===")
+    for name, value in components.items():
+        print(f"  {name:24s}  {value:.4f}")
+
+    detailed = [
+        {
+            "scenario": row["auxiliary_info"]["scenario_id"],
+            "view": row["view"],
+            "scenario_name": row["auxiliary_info"]["scenario_name"],
+            "score": results["physics_iq"].score,
+        }
+        for row, results in zip(matched, all_results)
+    ]
+    out = args.scores_out or (args.videos_dir / "scores.json")
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(
+        {"aggregate": components, "per_scenario": detailed},
+        indent=2,
+    ))
+    print(f"\n[done] per-scenario scores → {out}")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/eval/bench_vbench.py

@@ -0,0 +1,155 @@
+"""End-to-end VBench: dataset → generate → score → aggregate.
+
+Iterates the VBench prompt corpus, generates one video per prompt with
+LTX2, scores each generated video against the requested ``vbench.*``
+sub-metrics, and prints per-metric averages over the run.
+
+Re-running with ``--skip-generation`` reuses any mp4 already on disk
+under ``--videos-dir``, so you can iterate on metric selection without
+re-paying the generation cost.
+
+Example — quick smoke run on 4 prompts from the ``aesthetic_quality``
+dimension across 2 GPUs::
+
+    python examples/inference/eval/bench_vbench.py \\
+        --dimensions aesthetic_quality \\
+        --limit 4 --num-gpus 2 \\
+        --videos-dir outputs_video/vbench_smoke
+
+Full benchmark on a single dimension::
+
+    python examples/inference/eval/bench_vbench.py \\
+        --dimensions subject_consistency --num-gpus 8
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import re
+from collections import defaultdict
+from pathlib import Path
+
+from fastvideo.eval import create_evaluator
+from fastvideo.eval.datasets import get_dataset
+
+
+def _slugify(prompt: str, max_len: int = 100) -> str:
+    """Filesystem-safe filename stem; mirrors VBench's official convention."""
+    s = re.sub(r'[\\/:*?"<>|]', "", prompt[:max_len]).strip().strip(".")
+    return re.sub(r"\s+", " ", s) or "output"
+
+
+def _generate_videos(prompts: list[str], videos_dir: Path,
+                     model: str, num_gpus: int,
+                     num_frames: int, height: int, width: int) -> None:
+    from fastvideo import VideoGenerator
+
+    videos_dir.mkdir(parents=True, exist_ok=True)
+    todo = [(p, videos_dir / f"{_slugify(p)}.mp4") for p in prompts]
+    todo = [(p, out) for (p, out) in todo if not out.is_file()]
+    if not todo:
+        print(f"[gen] all {len(prompts)} videos already present; skipping.")
+        return
+
+    print(f"[gen] {len(todo)}/{len(prompts)} prompts to render with {model} "
+          f"({num_frames}x{height}x{width})...")
+    gen = VideoGenerator.from_pretrained(model, num_gpus=num_gpus)
+    try:
+        for prompt, out_path in todo:
+            gen.generate_video(
+                prompt=prompt, output_path=str(out_path), save_video=True,
+                num_frames=num_frames, height=height, width=width,
+            )
+    finally:
+        gen.shutdown()
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__,
+                                formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--dimensions", default="aesthetic_quality,subject_consistency",
+                   help="Comma-separated VBench dimensions (or 'all').")
+    p.add_argument("--limit", type=int, default=None,
+                   help="Truncate to first N prompts for smoke runs.")
+    p.add_argument("--videos-dir", type=Path,
+                   default=Path("outputs_video/bench_vbench"))
+    p.add_argument("--num-gpus", type=int, default=1)
+    p.add_argument("--model", default="Davids048/LTX2-Base-Diffusers",
+                   help="HF repo id of the text→video generator to use.")
+    p.add_argument("--num-frames", type=int, default=121)
+    p.add_argument("--height", type=int, default=1088)
+    p.add_argument("--width", type=int, default=1920)
+    p.add_argument("--fps", type=float, default=24.0,
+                   help="Frame-rate annotation passed to fps-aware metrics.")
+    p.add_argument("--skip-generation", action="store_true",
+                   help="Re-score existing videos under --videos-dir without "
+                        "regenerating.")
+    p.add_argument("--scores-out", type=Path, default=None,
+                   help="Where to dump per-prompt scores as JSON. "
+                        "Defaults to <videos-dir>/scores.json.")
+    args = p.parse_args()
+
+    # 1. Pull prompts from VBench.
+    dims_arg: list[str] | str = (
+        args.dimensions if args.dimensions == "all"
+        else [d.strip() for d in args.dimensions.split(",") if d.strip()]
+    )
+    ds = get_dataset("vbench", dimensions=dims_arg)
+    rows = list(ds)[: args.limit]
+    print(f"[load] VBench: {len(rows)} prompts across {ds.dimensions}")
+
+    # 2. Generate (or reuse) one mp4 per prompt.
+    if not args.skip_generation:
+        _generate_videos(
+            [row["prompt"] for row in rows],
+            args.videos_dir, args.model, args.num_gpus,
+            args.num_frames, args.height, args.width,
+        )
+
+    # 3. Score each video against the requested vbench sub-metrics.
+    metric_names = sorted(set(f"vbench.{d}" for d in ds.dimensions))
+    print(f"[eval] metrics: {metric_names}")
+    evaluator = create_evaluator(metrics=metric_names, num_gpus=args.num_gpus)
+
+    samples: list[dict] = []
+    matched_rows: list[dict] = []
+    for row in rows:
+        video_path = args.videos_dir / f"{_slugify(row['prompt'])}.mp4"
+        if not video_path.is_file():
+            print(f"[eval] missing {video_path}; skipping this row.")
+            continue
+        # Pass the path; the worker decodes lazily so memory stays bounded.
+        samples.append({
+            "video": str(video_path),
+            "fps": args.fps,
+            **row,                                       # prompt / aux / dims
+        })
+        matched_rows.append(row)
+
+    all_results = evaluator.evaluate(samples=samples)
+    evaluator.shutdown()
+
+    # 4. Aggregate per-metric.
+    by_metric: dict[str, list[float]] = defaultdict(list)
+    detailed: list[dict] = []
+    for row, results in zip(matched_rows, all_results):
+        scores = {name: r.score for name, r in results.items()}
+        detailed.append({"prompt": row["prompt"], "scores": scores})
+        for name, score in scores.items():
+            if score is not None:
+                by_metric[name].append(score)
+
+    print()
+    print("=== per-metric averages ===")
+    for name in sorted(by_metric):
+        avg = sum(by_metric[name]) / len(by_metric[name])
+        print(f"  {name:42s}  {avg:.4f}   (n={len(by_metric[name])})")
+
+    out = args.scores_out or (args.videos_dir / "scores.json")
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(detailed, indent=2))
+    print(f"\n[done] per-prompt scores → {out}")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/eval/eval_ltx2_vbench.py

@@ -0,0 +1,167 @@
+"""End-to-end: generate a video with LTX2 and score it with VBench.
+
+Pipeline:
+    prompt → LTX2-Base → mp4 → fastvideo.eval → vbench scores
+
+Run::
+
+    pip install -e .[eval]
+    git submodule update --init fastvideo/third_party/eval/vbench
+
+    python examples/inference/eval/eval_ltx2_vbench.py
+    # or with 4 GPUs and the distilled checkpoint:
+    python examples/inference/eval/eval_ltx2_vbench.py \
+        --model FastVideo/LTX2-Distilled-Diffusers --num-gpus 4
+
+The default metric set covers the vbench sub-metrics that are
+meaningful for an arbitrary text→video sample — i.e. those that need
+only the generated video (and optionally fps + the source prompt).
+Structured-prompt metrics like ``vbench.color``, ``vbench.scene``,
+``vbench.multiple_objects`` etc. are *not* on by default — they only
+make sense when the prompt is built to a specific schema, and they
+require GRiT/detectron2 setup. Pass them via ``--metrics`` if you have
+a matching prompt.
+
+First-time runs download CLIP, DINO, RAFT, AMT, ViCLIP, and MUSIQ
+weights to ``~/.cache/fastvideo/eval/models/`` and
+``~/.cache/torch/hub/`` (~few GB total). Subsequent runs are fast.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from fastvideo import VideoGenerator
+from fastvideo.eval import create_evaluator
+from fastvideo.eval.io import load_video
+
+
+PROMPT = (
+    "A warm sunny backyard. The camera starts in a tight cinematic close-up "
+    "of a woman and a man in their 30s, facing each other with serious "
+    "expressions. The woman, emotional and dramatic, says softly, \"That's "
+    "it... Dad's lost it. And we've lost Dad.\" The man exhales, slightly "
+    "annoyed: \"Stop being so dramatic, Jess.\" A beat. He glances aside, "
+    "then mutters defensively, \"He's just having fun.\" The camera slowly "
+    "pans right, revealing the grandfather in the garden wearing enormous "
+    "butterfly wings, waving his arms in the air like he's trying to take "
+    "off. He shouts, \"Wheeeew!\" as he flaps his wings with full commitment. "
+    "The woman covers her face, on the verge of tears. The tone is deadpan, "
+    "absurd, and quietly tragic."
+)
+
+DEFAULT_METRICS = [
+    # No-input metrics: just need the generated frames.
+    "vbench.aesthetic_quality",       # CLIP + LAION aesthetic head
+    "vbench.subject_consistency",     # DINO frame-to-first cosine
+    "vbench.background_consistency",  # DINO on background patches
+    "vbench.imaging_quality",         # pyiqa MUSIQ
+    "vbench.temporal_flickering",     # pixel-wise frame deltas
+    "vbench.motion_smoothness",       # AMT frame interpolator residual
+    # Need fps annotation:
+    "vbench.dynamic_degree",          # RAFT optical-flow magnitude
+    # Need the source prompt:
+    "vbench.overall_consistency",     # ViCLIP video↔prompt similarity
+]
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description=__doc__,
+                                formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--model", default="Davids048/LTX2-Base-Diffusers",
+                   help="HF repo id of the LTX2 checkpoint.")
+    p.add_argument("--num-gpus", type=int, default=1)
+    p.add_argument("--output", default="outputs_video/ltx2_eval/clip.mp4",
+                   help="Where to save the generated mp4.")
+    p.add_argument("--num-frames", type=int, default=121)
+    p.add_argument("--height", type=int, default=1088)
+    p.add_argument("--width", type=int, default=1920)
+    p.add_argument("--prompt", default=PROMPT)
+    p.add_argument("--fps", type=float, default=24.0,
+                   help="Frame-rate annotation passed to fps-aware metrics "
+                        "(e.g. vbench.dynamic_degree). LTX2 outputs at 24 fps "
+                        "by default.")
+    p.add_argument("--metrics", default=",".join(DEFAULT_METRICS),
+                   help="Comma-separated metric names. Pass 'all' for every "
+                        "registered metric, or e.g. 'vbench' for the whole group.")
+    p.add_argument("--scores-out", default="outputs_video/ltx2_eval/scores.json")
+    p.add_argument("--skip-generation", action="store_true",
+                   help="Reuse an existing --output video instead of regenerating.")
+    return p.parse_args()
+
+
+def generate(args: argparse.Namespace) -> Path:
+    out = Path(args.output)
+    if args.skip_generation and out.is_file():
+        print(f"[gen] reusing existing video at {out}")
+        return out
+    out.parent.mkdir(parents=True, exist_ok=True)
+
+    print(f"[gen] loading {args.model} ({args.num_gpus} GPU)...")
+    generator = VideoGenerator.from_pretrained(args.model, num_gpus=args.num_gpus)
+    try:
+        print(f"[gen] generating to {out}...")
+        generator.generate_video(
+            prompt=args.prompt,
+            output_path=str(out),
+            save_video=True,
+            num_frames=args.num_frames,
+            height=args.height,
+            width=args.width,
+        )
+    finally:
+        generator.shutdown()
+    return out
+
+
+def evaluate_video(video_path: Path, prompt: str, fps: float,
+                   metric_names) -> dict:
+    print(f"[eval] loading video from {video_path}...")
+    video = load_video(str(video_path))                  # (T, C, H, W) in [0, 1]
+    video = video.unsqueeze(0)                            # → (1, T, C, H, W)
+
+    print(f"[eval] building evaluator: {metric_names}")
+    evaluator = create_evaluator(metrics=metric_names, device="cuda")
+
+    print(f"[eval] running ({video.shape[1]} frames @ {fps} fps)...")
+    results = evaluator.evaluate(
+        video=video,
+        text_prompt=[prompt],
+        fps=fps,
+    )
+
+    if isinstance(results, list):
+        results = results[0]                              # batch of 1
+
+    return {
+        name: {"score": r.score, "details": r.details}
+        for name, r in results.items()
+    }
+
+
+def main() -> None:
+    args = parse_args()
+    if args.metrics.strip() == "all":
+        metric_names = "all"
+    else:
+        metric_names = [m.strip() for m in args.metrics.split(",") if m.strip()]
+
+    video_path = generate(args)
+    scores = evaluate_video(video_path, args.prompt, args.fps, metric_names)
+
+    print("\n=== VBench scores ===")
+    for name, payload in scores.items():
+        print(f"  {name}: {payload['score']}")
+
+    out = Path(args.scores_out)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(
+        {"video": str(video_path), "prompt": args.prompt, "scores": scores},
+        indent=2,
+    ))
+    print(f"[done] scores written to {out}")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/eval/score_folder.py

@@ -0,0 +1,95 @@
+"""Score a folder of videos in parallel across multiple GPUs.
+
+Uses :meth:`Evaluator.evaluate(samples=[...])`, which round-robins each
+sample dict across the GPU replicas the evaluator was built with.
+
+Example::
+
+    python examples/inference/eval/score_folder.py \\
+        --videos generated/ \\
+        --metrics vbench.aesthetic_quality,vbench.subject_consistency \\
+        --num-gpus 4 \\
+        --output scores.json
+
+Pair each generated video with a same-name reference video (e.g.
+``ref/<stem>.mp4``) by passing ``--reference-dir``::
+
+    python examples/inference/eval/score_folder.py \\
+        --videos generated/ --reference-dir ref/ \\
+        --metrics common.psnr,common.ssim,common.lpips \\
+        --num-gpus 4
+"""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from fastvideo.eval import create_evaluator
+
+
+def _list_videos(directory: Path) -> list[Path]:
+    exts = {".mp4", ".avi", ".mov", ".mkv", ".gif"}
+    return sorted(p for p in directory.iterdir() if p.suffix.lower() in exts)
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__,
+                                formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--videos", type=Path, required=True,
+                   help="Directory of generated videos.")
+    p.add_argument("--reference-dir", type=Path, default=None,
+                   help="Directory of reference videos with matching stems.")
+    p.add_argument("--metrics", default="vbench.aesthetic_quality")
+    p.add_argument("--num-gpus", type=int, default=1)
+    p.add_argument("--fps", type=float, default=None,
+                   help="Frame-rate annotation for fps-aware metrics.")
+    p.add_argument("--output", type=Path, default=Path("scores.json"))
+    args = p.parse_args()
+
+    video_paths = _list_videos(args.videos)
+    if not video_paths:
+        raise SystemExit(f"No videos under {args.videos}")
+    print(f"Found {len(video_paths)} videos in {args.videos}")
+
+    metrics: list[str] | str = (
+        args.metrics if args.metrics == "all"
+        else [m.strip() for m in args.metrics.split(",") if m.strip()]
+    )
+    evaluator = create_evaluator(metrics=metrics, num_gpus=args.num_gpus)
+
+    # Build per-video sample dicts holding *paths*, not pre-loaded
+    # tensors. Each path is decoded inside the worker thread that picks
+    # up its sample, so peak resident memory is bounded by num_gpus
+    # rather than scaling with the size of the folder.
+    samples: list[dict] = []
+    for vp in video_paths:
+        sample: dict = {"video": str(vp)}
+        if args.reference_dir is not None:
+            ref_path = args.reference_dir / vp.name
+            if not ref_path.is_file():
+                raise FileNotFoundError(f"Missing reference for {vp.name} at {ref_path}")
+            sample["reference"] = str(ref_path)
+        if args.fps is not None:
+            sample["fps"] = args.fps
+        samples.append(sample)
+
+    print(f"Scoring with {len(evaluator.metric_names)} metric(s) "
+          f"on {evaluator.num_gpus} GPU(s)...")
+    all_results = evaluator.evaluate(samples=samples)
+    evaluator.shutdown()
+
+    payload = [
+        {
+            "video": str(vp),
+            "scores": {name: r.score for name, r in results.items()},
+        }
+        for vp, results in zip(video_paths, all_results)
+    ]
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(payload, indent=2))
+    print(f"Wrote {args.output}")
+
+
+if __name__ == "__main__":
+    main()

examples/inference/eval/score_video.py

@@ -0,0 +1,68 @@
+"""Score one video on one GPU.
+
+Smallest possible use of ``fastvideo.eval``: load an mp4, build an
+:class:`Evaluator` for the requested metric set, run it.
+
+Examples::
+
+    # Reference-free (just the generated video):
+    python examples/inference/eval/score_video.py \\
+        --video clip.mp4 \\
+        --metrics vbench.aesthetic_quality,vbench.imaging_quality
+
+    # Reference-paired (compare against ground truth):
+    python examples/inference/eval/score_video.py \\
+        --video gen.mp4 --reference ref.mp4 \\
+        --metrics common.psnr,common.ssim,common.lpips
+"""
+from __future__ import annotations
+
+import argparse
+import json
+
+from fastvideo.eval import create_evaluator
+from fastvideo.eval.io import load_video
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__,
+                                formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--video", required=True, help="Path to the generated mp4.")
+    p.add_argument("--reference", default=None,
+                   help="Optional path to a reference mp4 (for paired metrics).")
+    p.add_argument("--metrics", default="common.psnr,common.ssim",
+                   help="Comma-separated metric names, or a group name like 'vbench'.")
+    p.add_argument("--device", default="cuda:0")
+    p.add_argument("--text-prompt", default=None,
+                   help="Text prompt for prompt-aware metrics "
+                        "(vbench.overall_consistency, etc.).")
+    p.add_argument("--fps", type=float, default=None,
+                   help="Frame-rate annotation for fps-aware metrics "
+                        "(vbench.dynamic_degree, etc.).")
+    args = p.parse_args()
+
+    metrics: list[str] | str = (
+        args.metrics if args.metrics in ("all",)
+        else [m.strip() for m in args.metrics.split(",") if m.strip()]
+    )
+    evaluator = create_evaluator(metrics=metrics, device=args.device)
+
+    sample: dict = {"video": load_video(args.video)}
+    if args.reference is not None:
+        sample["reference"] = load_video(args.reference)
+    if args.text_prompt is not None:
+        sample["text_prompt"] = args.text_prompt
+    if args.fps is not None:
+        sample["fps"] = args.fps
+
+    results = evaluator.evaluate(**sample)
+    evaluator.shutdown()
+
+    print(json.dumps(
+        {name: r.score for name, r in results.items()},
+        indent=2,
+    ))
+
+
+if __name__ == "__main__":
+    main()

fastvideo-kernel/build.sh

@@ -61,7 +61,16 @@ has_cmake_arg() {
 }
 
 detect_with_torch() {
-    uv run --active --no-project python -c "import torch
+    # Prefer the active venv's python directly over `uv run --active --no-project`,
+    # which on some uv versions provisions its own interpreter and misses packages
+    # installed into VIRTUAL_ENV.
+    local py
+    if [[ -n "${VIRTUAL_ENV:-}" && -x "${VIRTUAL_ENV}/bin/python" ]]; then
+        py="${VIRTUAL_ENV}/bin/python"
+    else
+        py="$(command -v python3 || command -v python)"
+    fi
+    "${py}" -c "import torch
 if not torch.cuda.is_available():
     raise RuntimeError('torch.cuda.is_available() is false')
 mj, mn = torch.cuda.get_device_capability(0)

fastvideo-kernel/python/fastvideo_kernel/__init__.py

@@ -5,6 +5,11 @@ from fastvideo_kernel.ops import (
     video_sparse_attn,
 )
 
+from fastvideo_kernel.block_sparse_attn import (
+    block_sparse_attn,
+    block_sparse_attn_from_indices,
+)
+
 from fastvideo_kernel.vmoba import (
     moba_attn_varlen,
     process_moba_input,
@@ -22,6 +27,8 @@ from fastvideo_kernel.turbodiffusion_ops import (
 __all__ = [
     "sliding_tile_attention",
     "video_sparse_attn",
+    "block_sparse_attn",
+    "block_sparse_attn_from_indices",
     "moba_attn_varlen",
     "process_moba_input",
     "process_moba_output",

fastvideo-kernel/python/fastvideo_kernel/block_sparse_attn.py

@@ -1,3 +1,5 @@
+"""Autograd-enabled block-sparse attention. Index-native ops with a bool-mask compat shim."""
+
 from __future__ import annotations
 
 import os
@@ -6,6 +8,11 @@ from typing import Tuple
 import torch
 
 
+# ---------------------------------------------------------------------------
+# Backend selection helpers
+# ---------------------------------------------------------------------------
+
+
 def _get_sm90_ops():
     try:
         from fastvideo_kernel._C import fastvideo_kernel_ops  # type: ignore
@@ -25,38 +32,66 @@ def _is_sm90() -> bool:
 
 
 def _force_triton() -> bool:
-    # Force Triton even on SM90 and even if the compiled extension is available.
-    # Useful for CI / debugging / parity testing.
     return os.environ.get("FASTVIDEO_KERNEL_VSA_FORCE_TRITON", "0") == "1"
 
 
-def _map_to_index(block_map: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-    """
-    Preferred map->index conversion used by the wrapper.
+# ---------------------------------------------------------------------------
+# Index helpers
+# ---------------------------------------------------------------------------
 
-    This wrapper **requires** the Triton implementation.
-    If Triton (or the Triton map_to_index module) is not available, it raises.
-    """
+
+def _map_to_index(block_map: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Compact a bool block_map to (q2k_idx, q2k_num). Legacy path only."""
     if block_map.dim() == 3:
         block_map = block_map.unsqueeze(0)
     if block_map.dim() != 4:
-        raise ValueError(f"block_map must be [B,H,Q,KV] (or [H,Q,KV]), got shape={tuple(block_map.shape)}")
+        raise ValueError(
+            f"block_map must be [B,H,Q,KV] (or [H,Q,KV]), "
+            f"got shape={tuple(block_map.shape)}"
+        )
     if block_map.dtype != torch.bool:
         block_map = block_map.to(torch.bool)
-
     if not block_map.is_cuda:
-        raise RuntimeError("block_map must be a CUDA tensor (Triton map_to_index required).")
+        raise RuntimeError(
+            "block_map must be a CUDA tensor (Triton map_to_index required)."
+        )
 
     try:
-        from fastvideo_kernel.triton_kernels.index import map_to_index as triton_map_to_index  # local import
-    except Exception as e:
+        from fastvideo_kernel.triton_kernels.index import map_to_index as triton_map_to_index
+    except Exception as e:  # pragma: no cover - environment issue
         raise ImportError(
             "Triton map_to_index is required but not available. "
-            "Ensure Triton is installed and fastvideo_kernel.triton_kernels.index is importable."
+            "Ensure Triton is installed and "
+            "fastvideo_kernel.triton_kernels.index is importable."
         ) from e
     return triton_map_to_index(block_map)
 
 
+def _invert_indices_for_backward(
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
+    num_kv_blocks: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    from fastvideo_kernel.triton_kernels.index import invert_indices
+    return invert_indices(q2k_idx, q2k_num, num_kv_blocks=num_kv_blocks)
+
+
+def _as_int32_contig(t: torch.Tensor, name: str) -> torch.Tensor:
+    """Return `t` as a contiguous int32 tensor, raising a clear error on CPU input."""
+    if not t.is_cuda:
+        raise RuntimeError(f"{name} must be a CUDA tensor, got device={t.device}")
+    if t.dtype != torch.int32:
+        t = t.to(torch.int32)
+    if not t.is_contiguous():
+        t = t.contiguous()
+    return t
+
+
+# ---------------------------------------------------------------------------
+# Triton backend custom ops (index-native)
+# ---------------------------------------------------------------------------
+
+
 @torch.library.custom_op(
     "fastvideo_kernel::block_sparse_attn_triton",
     mutates_args=(),
@@ -66,34 +101,40 @@ def block_sparse_attn_triton(
     q: torch.Tensor,
     k: torch.Tensor,
     v: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
-    q = q.contiguous()
-    k = k.contiguous()
-    v = v.contiguous()
-    block_map = block_map.to(torch.bool)
-    q2k_idx, q2k_num = _map_to_index(block_map)
-
-    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (  # local import
+    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (
         triton_block_sparse_attn_forward,
     )
 
-    o, M = triton_block_sparse_attn_forward(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
+    o, M = triton_block_sparse_attn_forward(
+        q.contiguous(),
+        k.contiguous(),
+        v.contiguous(),
+        q2k_idx,
+        q2k_num,
+        variable_block_sizes,
+    )
     return o, M
 
 
-
 @torch.library.register_fake("fastvideo_kernel::block_sparse_attn_triton")
 def _block_sparse_attn_triton_fake(
     q: torch.Tensor,
     k: torch.Tensor,
     v: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     o = torch.empty_like(q)
-    M = torch.empty((q.shape[0], q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
+    M = torch.empty(
+        (q.shape[0], q.shape[1], q.shape[2]),
+        device=q.device,
+        dtype=torch.float32,
+    )
     return o, M
 
 
@@ -109,20 +150,32 @@ def block_sparse_attn_backward_triton(
     v: torch.Tensor,
     o: torch.Tensor,
     M: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    grad_output = grad_output.contiguous()
-    block_map = block_map.to(torch.bool)
-    q2k_idx, q2k_num = _map_to_index(block_map)
-    k2q_idx, k2q_num = _map_to_index(block_map.transpose(-1, -2).contiguous())
-
-    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (  # local import
+    from fastvideo_kernel.triton_kernels.block_sparse_attn_triton import (
         triton_block_sparse_attn_backward,
     )
 
+    num_kv_blocks = int(variable_block_sizes.numel())
+    k2q_idx, k2q_num = _invert_indices_for_backward(
+        q2k_idx, q2k_num, num_kv_blocks
+    )
+    # q/k/v are saved from the user-facing inputs and may be non-contiguous;
+    # o/M are kernel outputs so are already contiguous.
     dq, dk, dv = triton_block_sparse_attn_backward(
-        grad_output, q, k, v, o, M, q2k_idx, q2k_num, k2q_idx, k2q_num, variable_block_sizes
+        grad_output.contiguous(),
+        q.contiguous(),
+        k.contiguous(),
+        v.contiguous(),
+        o,
+        M,
+        q2k_idx,
+        q2k_num,
+        k2q_idx,
+        k2q_num,
+        variable_block_sizes,
     )
     return dq, dk, dv
 
@@ -135,7 +188,8 @@ def _block_sparse_attn_backward_triton_fake(
     v: torch.Tensor,
     o: torch.Tensor,
     M: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     dq = torch.empty_like(q)
@@ -144,19 +198,28 @@ def _block_sparse_attn_backward_triton_fake(
     return dq, dk, dv
 
 
-def _backward_triton(ctx, grad_o, grad_M):
-    q, k, v, o, M, block_map, variable_block_sizes = ctx.saved_tensors
-    dq, dk, dv = block_sparse_attn_backward_triton(grad_o, q, k, v, o, M, block_map, variable_block_sizes)
-    return dq, dk, dv, None, None
-
-
 def _setup_context_triton(ctx, inputs, output):
-    q, k, v, block_map, variable_block_sizes = inputs
+    q, k, v, q2k_idx, q2k_num, variable_block_sizes = inputs
     o, M = output
-    ctx.save_for_backward(q, k, v, o, M, block_map, variable_block_sizes)
+    ctx.save_for_backward(q, k, v, o, M, q2k_idx, q2k_num, variable_block_sizes)
 
 
-block_sparse_attn_triton.register_autograd(_backward_triton, setup_context=_setup_context_triton)
+def _backward_triton(ctx, grad_o, grad_M):
+    q, k, v, o, M, q2k_idx, q2k_num, variable_block_sizes = ctx.saved_tensors
+    dq, dk, dv = block_sparse_attn_backward_triton(
+        grad_o, q, k, v, o, M, q2k_idx, q2k_num, variable_block_sizes
+    )
+    return dq, dk, dv, None, None, None
+
+
+block_sparse_attn_triton.register_autograd(
+    _backward_triton, setup_context=_setup_context_triton
+)
+
+
+# ---------------------------------------------------------------------------
+# SM90 backend custom ops (index-native)
+# ---------------------------------------------------------------------------
 
 
 @torch.library.custom_op(
@@ -168,21 +231,21 @@ def block_sparse_attn_sm90(
     q_padded: torch.Tensor,
     k_padded: torch.Tensor,
     v_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     block_sparse_fwd, _ = _get_sm90_ops()
     if block_sparse_fwd is None:
         raise ImportError("fastvideo_kernel_ops.block_sparse_fwd is not available")
 
-    q_padded = q_padded.contiguous()
-    k_padded = k_padded.contiguous()
-    v_padded = v_padded.contiguous()
-    block_map = block_map.to(torch.bool)
-    q2k_idx, q2k_num = _map_to_index(block_map)
-
     o_padded, lse_padded = block_sparse_fwd(
-        q_padded, k_padded, v_padded, q2k_idx, q2k_num, variable_block_sizes.int()
+        q_padded.contiguous(),
+        k_padded.contiguous(),
+        v_padded.contiguous(),
+        q2k_idx,
+        q2k_num,
+        variable_block_sizes,
     )
     return o_padded, lse_padded
 
@@ -192,11 +255,16 @@ def _block_sparse_attn_sm90_fake(
     q_padded: torch.Tensor,
     k_padded: torch.Tensor,
     v_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     o = torch.empty_like(q_padded)
-    lse = torch.empty((q_padded.shape[0], q_padded.shape[1], q_padded.shape[2], 1), device=q_padded.device, dtype=torch.float32)
+    lse = torch.empty(
+        (q_padded.shape[0], q_padded.shape[1], q_padded.shape[2], 1),
+        device=q_padded.device,
+        dtype=torch.float32,
+    )
     return o, lse
 
 
@@ -212,30 +280,34 @@ def block_sparse_attn_backward_sm90(
     v_padded: torch.Tensor,
     o_padded: torch.Tensor,
     lse_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     _, block_sparse_bwd = _get_sm90_ops()
     if block_sparse_bwd is None:
         raise ImportError("fastvideo_kernel_ops.block_sparse_bwd is not available")
 
-    grad_output_padded = grad_output_padded.contiguous()
-    block_map = block_map.to(torch.bool)
-    k2q_idx, k2q_num = _map_to_index(block_map.transpose(-1, -2).contiguous())
+    num_kv_blocks = int(variable_block_sizes.numel())
+    k2q_idx, k2q_num = _invert_indices_for_backward(
+        q2k_idx, q2k_num, num_kv_blocks
+    )
 
+    # q/k/v are saved from user-facing inputs; o/lse are kernel outputs.
     dq, dk, dv = block_sparse_bwd(
-        q_padded,
-        k_padded,
-        v_padded,
+        q_padded.contiguous(),
+        k_padded.contiguous(),
+        v_padded.contiguous(),
         o_padded,
         lse_padded,
-        grad_output_padded,
+        grad_output_padded.contiguous(),
         k2q_idx,
         k2q_num,
-        variable_block_sizes.int(),
+        variable_block_sizes,
     )
-    # C++ kernel returns fp32 grads; cast back to match PyTorch convention if needed
-    return dq.to(grad_output_padded.dtype), dk.to(grad_output_padded.dtype), dv.to(grad_output_padded.dtype)
+    # C++ kernel returns fp32 grads; cast back to the input dtype.
+    out_dtype = grad_output_padded.dtype
+    return dq.to(out_dtype), dk.to(out_dtype), dv.to(out_dtype)
 
 
 @torch.library.register_fake("fastvideo_kernel::block_sparse_attn_backward_sm90")
@@ -246,7 +318,8 @@ def _block_sparse_attn_backward_sm90_fake(
     v_padded: torch.Tensor,
     o_padded: torch.Tensor,
     lse_padded: torch.Tensor,
-    block_map: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     dq = torch.empty_like(q_padded)
@@ -255,21 +328,57 @@ def _block_sparse_attn_backward_sm90_fake(
     return dq, dk, dv
 
 
-def _backward_sm90(ctx, grad_o, grad_lse):
-    q, k, v, o, lse, block_map, variable_block_sizes = ctx.saved_tensors
-    dq, dk, dv = block_sparse_attn_backward_sm90(
-        grad_o, q, k, v, o, lse, block_map, variable_block_sizes
-    )
-    return dq, dk, dv, None, None
-
-
 def _setup_context_sm90(ctx, inputs, output):
-    q, k, v, block_map, variable_block_sizes = inputs
+    q, k, v, q2k_idx, q2k_num, variable_block_sizes = inputs
     o, lse = output
-    ctx.save_for_backward(q, k, v, o, lse, block_map, variable_block_sizes)
+    ctx.save_for_backward(q, k, v, o, lse, q2k_idx, q2k_num, variable_block_sizes)
 
 
-block_sparse_attn_sm90.register_autograd(_backward_sm90, setup_context=_setup_context_sm90)
+def _backward_sm90(ctx, grad_o, grad_lse):
+    q, k, v, o, lse, q2k_idx, q2k_num, variable_block_sizes = ctx.saved_tensors
+    dq, dk, dv = block_sparse_attn_backward_sm90(
+        grad_o, q, k, v, o, lse, q2k_idx, q2k_num, variable_block_sizes
+    )
+    return dq, dk, dv, None, None, None
+
+
+block_sparse_attn_sm90.register_autograd(
+    _backward_sm90, setup_context=_setup_context_sm90
+)
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+def block_sparse_attn_from_indices(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
+    variable_block_sizes: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Block-sparse attention with autograd, taking compact per-row KV indices."""
+    # Normalize index tensors once at the public boundary so the custom ops
+    # and their fakes can assume int32/contiguous. No-op on well-formed input.
+    q2k_idx = _as_int32_contig(q2k_idx, "q2k_idx")
+    q2k_num = _as_int32_contig(q2k_num, "q2k_num")
+    variable_block_sizes = _as_int32_contig(variable_block_sizes, "variable_block_sizes")
+
+    block_sparse_fwd, block_sparse_bwd = _get_sm90_ops()
+    use_sm90 = (
+        (not _force_triton())
+        and _is_sm90()
+        and block_sparse_fwd is not None
+        and block_sparse_bwd is not None
+    )
+    if use_sm90:
+        return block_sparse_attn_sm90(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
+    # Triton path: supports q_seq_len != kv_seq_len as long as both are padded
+    # to a multiple of the block size (64 tokens).
+    return block_sparse_attn_triton(q, k, v, q2k_idx, q2k_num, variable_block_sizes)
 
 
 def block_sparse_attn(
@@ -279,16 +388,8 @@ def block_sparse_attn(
     block_map: torch.Tensor,
     variable_block_sizes: torch.Tensor,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
-    """
-    Unified block-sparse attention op with autograd support.
-    - On SM90 with compiled extension present: uses fastvideo_kernel_ops.block_sparse_fwd/bwd.
-    - Otherwise: uses Triton implementation (requires q/k/v to have same padded length today).
-    """
-    block_sparse_fwd, block_sparse_bwd = _get_sm90_ops()
-    if (not _force_triton()) and _is_sm90() and (block_sparse_fwd is not None) and (block_sparse_bwd is not None):
-        return block_sparse_attn_sm90(q, k, v, block_map, variable_block_sizes)
-    # Triton path: supports q_seq_len != kv_seq_len as long as both are padded
-    # to a multiple of the block size (64 tokens).
-    return block_sparse_attn_triton(q, k, v, block_map, variable_block_sizes)
-
-
+    """Bool-mask compat wrapper; prefer block_sparse_attn_from_indices."""
+    q2k_idx, q2k_num = _map_to_index(block_map)
+    return block_sparse_attn_from_indices(
+        q, k, v, q2k_idx, q2k_num, variable_block_sizes
+    )

fastvideo-kernel/python/fastvideo_kernel/ops.py

@@ -1,6 +1,6 @@
 import math
 import torch
-from .block_sparse_attn import block_sparse_attn
+from .block_sparse_attn import block_sparse_attn, block_sparse_attn_from_indices
 from .triton_kernels.st_attn_triton import sliding_tile_attention_triton
 
 # Try to load the C++ extension
@@ -125,13 +125,18 @@ def video_sparse_attn(
     out_c = out_c.repeat(1, 1, 1, block_elements,
                          1).view(batch, heads, q_seq_len, dim)
 
-    # Sparse branch
+    # Sparse branch: feed top-k indices directly, skipping the bool-mask round-trip.
     topk_idx = torch.topk(scores, topk, dim=-1).indices
-    mask = torch.zeros_like(scores,
-                            dtype=torch.bool).scatter_(-1, topk_idx, True)
-
-    # out_s = block_sparse_attn(q, k, v, mask, variable_block_sizes)[0]
-    out_s = block_sparse_attn(q, k, v, mask, variable_block_sizes)[0]
+    q2k_idx = topk_idx.to(torch.int32).contiguous()
+    q2k_num = torch.full(
+        (batch, heads, q_num_blocks),
+        topk,
+        dtype=torch.int32,
+        device=q.device,
+    )
+    out_s = block_sparse_attn_from_indices(
+        q, k, v, q2k_idx, q2k_num, variable_block_sizes
+    )[0]
 
     if compress_attn_weight is not None:
         return out_c * compress_attn_weight + out_s

fastvideo-kernel/python/fastvideo_kernel/triton_kernels/index.py

@@ -1,9 +1,10 @@
 ## pytorch sdpa version of block sparse ##
+from typing import Tuple
+
 import triton
 import triton.language as tl
 import torch
 
-
 @triton.jit
 def topk_index_to_map_kernel(
     map_ptr,
@@ -153,3 +154,114 @@ def map_to_index(block_map: torch.Tensor):
     )
 
     return index, index_num
+
+
+@triton.jit
+def _invert_indices_kernel(
+    q2k_idx_ptr,
+    q2k_num_ptr,
+    k2q_idx_ptr,
+    k2q_num_ptr,
+    q2k_idx_b, q2k_idx_h, q2k_idx_q, q2k_idx_k,
+    q2k_num_b, q2k_num_h, q2k_num_q,
+    k2q_idx_b, k2q_idx_h, k2q_idx_k, k2q_idx_q,
+    k2q_num_b, k2q_num_h, k2q_num_k,
+    MAX_KV_PER_Q: tl.constexpr,
+):
+    # One program per (b, h, q): reserve a slot in k2q via atomicAdd, write q.
+    pid_b = tl.program_id(0)
+    pid_h = tl.program_id(1)
+    pid_q = tl.program_id(2)
+
+    n = tl.load(
+        q2k_num_ptr
+        + pid_b * q2k_num_b
+        + pid_h * q2k_num_h
+        + pid_q * q2k_num_q
+    )
+
+    q2k_row = (
+        q2k_idx_ptr
+        + pid_b * q2k_idx_b
+        + pid_h * q2k_idx_h
+        + pid_q * q2k_idx_q
+    )
+
+    for i in tl.range(0, MAX_KV_PER_Q):
+        if i < n:
+            kv = tl.load(q2k_row + i * q2k_idx_k)
+            count_ptr = (
+                k2q_num_ptr
+                + pid_b * k2q_num_b
+                + pid_h * k2q_num_h
+                + kv * k2q_num_k
+            )
+            pos = tl.atomic_add(count_ptr, 1)
+            tl.store(
+                k2q_idx_ptr
+                + pid_b * k2q_idx_b
+                + pid_h * k2q_idx_h
+                + kv * k2q_idx_k
+                + pos * k2q_idx_q,
+                pid_q,
+            )
+
+
+def invert_indices(
+    q2k_idx: torch.Tensor,
+    q2k_num: torch.Tensor,
+    num_kv_blocks: int,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Transpose a Q->KV index list into a K->Q one via atomic compaction (GPU)."""
+    if q2k_idx.dim() != 4:
+        raise ValueError(
+            f"q2k_idx must be [B, H, Nq, Mk], got shape={tuple(q2k_idx.shape)}"
+        )
+    if q2k_num.dim() != 3:
+        raise ValueError(
+            f"q2k_num must be [B, H, Nq], got shape={tuple(q2k_num.shape)}"
+        )
+    if not q2k_idx.is_cuda or not q2k_num.is_cuda:
+        raise RuntimeError("invert_indices requires CUDA tensors.")
+
+    B, H, Nq, Mk = q2k_idx.shape
+    if q2k_num.shape != (B, H, Nq):
+        raise ValueError(
+            f"q2k_num shape {tuple(q2k_num.shape)} does not match q2k_idx "
+            f"[B, H, Nq] = {(B, H, Nq)}"
+        )
+
+    q2k_idx = q2k_idx.contiguous()
+    q2k_num = q2k_num.contiguous()
+    if q2k_idx.dtype != torch.int32:
+        q2k_idx = q2k_idx.to(torch.int32)
+    if q2k_num.dtype != torch.int32:
+        q2k_num = q2k_num.to(torch.int32)
+
+    # Any KV block is attended by at most Nq Q blocks (one per Q row), so
+    # `Nq` is a tight upper bound on the compacted K->Q slots.
+    k2q_idx = torch.empty(
+        (B, H, num_kv_blocks, Nq),
+        dtype=torch.int32,
+        device=q2k_idx.device,
+    )
+    k2q_num = torch.zeros(
+        (B, H, num_kv_blocks),
+        dtype=torch.int32,
+        device=q2k_idx.device,
+    )
+
+    grid = (B, H, Nq)
+    _invert_indices_kernel[grid](
+        q2k_idx,
+        q2k_num,
+        k2q_idx,
+        k2q_num,
+        q2k_idx.stride(0), q2k_idx.stride(1), q2k_idx.stride(2), q2k_idx.stride(3),
+        q2k_num.stride(0), q2k_num.stride(1), q2k_num.stride(2),
+        k2q_idx.stride(0), k2q_idx.stride(1), k2q_idx.stride(2), k2q_idx.stride(3),
+        k2q_num.stride(0), k2q_num.stride(1), k2q_num.stride(2),
+        MAX_KV_PER_Q=Mk,
+    )
+
+    return k2q_idx, k2q_num

fastvideo/api/compat.py

@@ -16,6 +16,8 @@ from fastvideo.api.request_metadata import (
     reset_tracking_roots,
 )
 from fastvideo.api.schema import (
+    CompileConfig,
+    ContinuationState,
     GenerationRequest,
     GeneratorConfig,
     InputConfig,
@@ -25,6 +27,10 @@ from fastvideo.api.schema import (
 )
 from fastvideo.api.sampling_param import SamplingParam
 from fastvideo.fastvideo_args import FastVideoArgs
+from fastvideo.pipelines.basic.ltx2.stage_overrides import (
+    refine_preset_override_fields,
+    refine_stage_override_fields,
+)
 from fastvideo.utils import shallow_asdict
 
 _INPUT_FIELD_NAMES = {field.name for field in fields(InputConfig)}
@@ -38,6 +44,10 @@ _LEGACY_REQUEST_ALIASES = {
 _REQUEST_PIPELINE_OVERRIDE_FIELDS = frozenset({
     "embedded_cfg_scale",
 })
+# torch.compile kwargs that map to first-class CompileConfig fields.
+_COMPILE_TYPED_KEYS = ("backend", "fullgraph", "mode", "dynamic")
+# LTX-2 refine flat kwargs (init + per-request) known to FastVideoArgs.
+_LTX2_REFINE_FLAT_KEYS = (refine_preset_override_fields() | refine_stage_override_fields())
 
 
 def normalize_generator_config(config: GeneratorConfig | Mapping[str, Any], ) -> GeneratorConfig:
@@ -80,6 +90,8 @@ def legacy_from_pretrained_to_config(
     components: dict[str, Any] = {}
     quantization: dict[str, Any] = {}
     experimental: dict[str, Any] = {}
+    preset_overrides: dict[str, Any] = {}
+    preset_refine: dict[str, Any] = {}
 
     for key, value in kwargs.items():
         if key == "revision":
@@ -106,8 +118,33 @@ def legacy_from_pretrained_to_config(
             offload["pin_cpu_memory"] = value
         elif key == "enable_torch_compile":
             compile_config["enabled"] = value
+        elif key == "enable_torch_compile_text_encoder":
+            compile_config["text_encoder_enabled"] = value
         elif key == "torch_compile_kwargs":
-            compile_config["kwargs"] = deepcopy(value)
+            remaining: dict[str, Any] = (dict(deepcopy(value)) if isinstance(value, Mapping) else {})
+            for first_class in _COMPILE_TYPED_KEYS:
+                if first_class in remaining:
+                    compile_config[first_class] = remaining.pop(first_class)
+            if remaining:
+                compile_config["extras"] = remaining
+        elif key == "ltx2_vae_tiling":
+            pipeline["vae_tiling"] = value
+        elif key == "config_model_path":
+            components["config_root"] = value
+        elif key == "ltx2_refine_enabled":
+            preset_refine["enabled"] = value
+        elif key == "ltx2_refine_upsampler_path":
+            # Empty string means "no upsampler"; keep typed None.
+            components["upsampler_weights"] = value or None
+        elif key == "ltx2_refine_lora_path":
+            # Empty string means "no refine LoRA"; keep typed None.
+            components["lora_path"] = value or None
+        elif key == "ltx2_refine_add_noise":
+            preset_refine["add_noise"] = value
+        elif key == "ltx2_refine_num_inference_steps":
+            preset_refine["num_inference_steps"] = value
+        elif key == "ltx2_refine_guidance_scale":
+            preset_refine["guidance_scale"] = value
         elif key in {"enable_stage_verification", "use_fsdp_inference", "disable_autocast"}:
             engine[key] = value
         elif key == "override_text_encoder_quant":
@@ -147,6 +184,10 @@ def legacy_from_pretrained_to_config(
 
     if components:
         pipeline["components"] = components
+    if preset_refine:
+        preset_overrides["refine"] = preset_refine
+    if preset_overrides:
+        pipeline["preset_overrides"] = preset_overrides
     if experimental:
         pipeline["experimental"] = experimental
     if pipeline:
@@ -162,12 +203,8 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         unsupported.append("pipeline.preset")
     if normalized.pipeline.preset_version is not None:
         unsupported.append("pipeline.preset_version")
-    if normalized.pipeline.components.config_root is not None:
-        unsupported.append("pipeline.components.config_root")
     if normalized.pipeline.components.vae_weights is not None:
         unsupported.append("pipeline.components.vae_weights")
-    if normalized.pipeline.components.upsampler_weights is not None:
-        unsupported.append("pipeline.components.upsampler_weights")
     if unsupported:
         joined = ", ".join(unsupported)
         raise NotImplementedError(f"VideoGenerator compatibility adapter does not support {joined} yet")
@@ -191,13 +228,21 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         "vae_cpu_offload": engine.offload.vae,
         "pin_cpu_memory": engine.offload.pin_cpu_memory,
         "enable_torch_compile": engine.compile.enabled,
-        "torch_compile_kwargs": deepcopy(engine.compile.kwargs),
+        "torch_compile_kwargs": _compile_config_to_torch_kwargs(engine.compile),
         "enable_stage_verification": engine.enable_stage_verification,
         "use_fsdp_inference": engine.use_fsdp_inference,
         "disable_autocast": engine.disable_autocast,
     }
     if normalized.pipeline.workload_type is not None:
         kwargs["workload_type"] = normalized.pipeline.workload_type
+    if normalized.pipeline.vae_tiling is not None:
+        kwargs["ltx2_vae_tiling"] = normalized.pipeline.vae_tiling
+    if engine.compile.text_encoder_enabled is not None:
+        # ``FastVideoArgs.from_kwargs`` filters to declared fields, so
+        # this is a no-op on the current legacy path. Emit anyway so the
+        # realtime runtime (PR 7.6) — which reads from the kwargs dict
+        # before FastVideoArgs filtering — can pick it up once wired.
+        kwargs["enable_torch_compile_text_encoder"] = (engine.compile.text_encoder_enabled)
 
     quantization = engine.quantization
     if quantization is not None and quantization.text_encoder_quant is not None:
@@ -220,8 +265,18 @@ def generator_config_to_fastvideo_args(config: GeneratorConfig | Mapping[str, An
         kwargs["init_weights_from_safetensors"] = components.transformer_weights
     if components.transformer_2_weights is not None:
         kwargs["init_weights_from_safetensors_2"] = components.transformer_2_weights
+    if components.config_root is not None:
+        kwargs["config_model_path"] = components.config_root
+    if components.upsampler_weights is not None:
+        kwargs["ltx2_refine_upsampler_path"] = components.upsampler_weights
 
-    kwargs.update(deepcopy(normalized.pipeline.preset_overrides))
+    preset_overrides = deepcopy(normalized.pipeline.preset_overrides)
+    refine = preset_overrides.pop("refine", None)
+    if isinstance(refine, Mapping):
+        for key in _LTX2_REFINE_FLAT_KEYS:
+            if key in refine:
+                kwargs[f"ltx2_refine_{key}"] = refine[key]
+    kwargs.update(preset_overrides)
     kwargs.update(deepcopy(normalized.pipeline.experimental))
     return FastVideoArgs.from_kwargs(**kwargs)
 
@@ -271,10 +326,13 @@ def request_to_sampling_param(
 ) -> SamplingParam:
     if request.plan is not None:
         raise NotImplementedError("GenerationRequest.plan is not wired into VideoGenerator yet")
-    if request.state is not None:
-        raise NotImplementedError("GenerationRequest.state is not wired into VideoGenerator yet")
 
     sampling_param = SamplingParam.from_pretrained(model_path)
+    if request.state is not None:
+        _validate_continuation_state(request.state)
+        sampling_param.continuation_state = request.state
+    if request.output.return_state:
+        sampling_param.return_continuation_state = True
     updates = explicit_request_updates(request)
 
     for key, value in updates.items():
@@ -316,6 +374,25 @@ def _looks_like_run_or_serve_config(raw: Mapping[str, Any]) -> bool:
     return isinstance(raw.get("generator"), Mapping)
 
 
+def _compile_config_to_torch_kwargs(compile_config: CompileConfig, ) -> dict[str, Any]:
+    """Flatten typed ``CompileConfig`` back to a ``torch_compile_kwargs``
+    dict that the legacy ``FastVideoArgs`` path still expects.
+
+    Typed first-class fields (:attr:`backend`, :attr:`fullgraph`,
+    :attr:`mode`, :attr:`dynamic`) are only emitted when the user set
+    them explicitly (non-``None``). ``extras`` is merged on top for any
+    uncommon kwargs.
+    """
+    out: dict[str, Any] = {}
+    for key in _COMPILE_TYPED_KEYS:
+        value = getattr(compile_config, key)
+        if value is not None:
+            out[key] = value
+    if compile_config.extras:
+        out.update(deepcopy(compile_config.extras))
+    return out
+
+
 def _sampling_param_to_request_raw(sampling_param: SamplingParam | None, ) -> dict[str, Any]:
     if sampling_param is None:
         return {}
@@ -476,6 +553,37 @@ def _serialize_generation_request(request: GenerationRequest) -> dict[str, Any]:
 
 _SCHEMA_DEFAULT_UPDATES = _extract_request_updates(config_to_dict(GenerationRequest()))
 
+_KNOWN_CONTINUATION_KINDS: set[str] = set()
+
+
+def register_continuation_kind(kind: str) -> None:
+    """Register a :class:`ContinuationState.kind` as recognized.
+
+    PR 7 wires the envelope through; per-kind payload deserializers live
+    with each model family (e.g. ``fastvideo.pipelines.basic.ltx2.
+    continuation.LTX2ContinuationState``). The registry lets the
+    public-API compat layer validate the kind early, before the state
+    reaches the pipeline.
+    """
+    if not isinstance(kind, str) or not kind:
+        raise ValueError("ContinuationState kind must be a non-empty string")
+    _KNOWN_CONTINUATION_KINDS.add(kind)
+
+
+def _validate_continuation_state(state: ContinuationState) -> None:
+    if not isinstance(state.kind, str) or not state.kind:
+        raise ValueError("GenerationRequest.state.kind must be a non-empty string; got "
+                         f"{state.kind!r}")
+    if not isinstance(state.payload, Mapping):
+        raise ValueError(f"GenerationRequest.state.payload must be a mapping; got "
+                         f"{type(state.payload).__name__}")
+    if state.kind not in _KNOWN_CONTINUATION_KINDS:
+        known = sorted(_KNOWN_CONTINUATION_KINDS)
+        raise ValueError(f"Unknown ContinuationState kind {state.kind!r}; registered "
+                         f"kinds: {known}. Import the model family that owns this kind "
+                         "(e.g. `import fastvideo.pipelines.basic.ltx2.continuation`) "
+                         "to register it, or drop the state field.")
+
 
 def _fan_out_batched_input_value(
     source_request: GenerationRequest,
@@ -509,6 +617,7 @@ __all__ = [
     "load_generator_config_from_file",
     "normalize_generation_request",
     "normalize_generator_config",
+    "register_continuation_kind",
     "request_to_pipeline_overrides",
     "request_to_sampling_param",
 ]

fastvideo/api/sampling_param.py

@@ -1,11 +1,16 @@
 # SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
 import copy
 from dataclasses import dataclass, field, fields
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from fastvideo.logger import init_logger
 from fastvideo.utils import StoreBoolean
 
+if TYPE_CHECKING:
+    from fastvideo.api.schema import ContinuationState
+
 logger = init_logger(__name__)
 
 
@@ -92,9 +97,13 @@ class SamplingParam:
     movement_distance: float | None = None
     camera_rotation: str | None = None
 
-    # LTX2 multi-modal CFG and STG
-    ltx2_cfg_scale_video: float = 3.0
-    ltx2_cfg_scale_audio: float = 7.0
+    # LTX-2 multi-modal CFG and STG.
+    # cfg_scale defaults are 1.0 (CFG off) so ``ForwardBatch.__post_init__``
+    # doesn't force ``do_classifier_free_guidance`` on non-LTX-2 models that
+    # never override these fields. LTX-2 presets that need text-CFG on set
+    # them in their ``defaults`` dict (e.g. ``ltx2_base``).
+    ltx2_cfg_scale_video: float = 1.0
+    ltx2_cfg_scale_audio: float = 1.0
     ltx2_modality_scale_video: float = 3.0
     ltx2_modality_scale_audio: float = 3.0
     ltx2_rescale_scale: float = 0.7
@@ -103,6 +112,12 @@ class SamplingParam:
     ltx2_stg_blocks_video: list[int] = field(default_factory=lambda: [29])
     ltx2_stg_blocks_audio: list[int] = field(default_factory=lambda: [29])
 
+    # Continuation state carried across streaming/multi-segment calls.
+    continuation_state: ContinuationState | None = None
+    # When True, the pipeline returns a ContinuationState on the result so
+    # the caller can resume from the generated segment.
+    return_continuation_state: bool = False
+
     # Misc
     save_video: bool = True
     return_frames: bool = True
@@ -127,7 +142,7 @@ class SamplingParam:
         self.__post_init__()
 
     @classmethod
-    def from_pretrained(cls, model_path: str) -> "SamplingParam":
+    def from_pretrained(cls, model_path: str) -> SamplingParam:
         sampling_param = cls._from_preset(model_path)
         if sampling_param is not None:
             return sampling_param
@@ -143,7 +158,7 @@ class SamplingParam:
     def _from_preset(
         cls,
         model_path: str,
-    ) -> "SamplingParam | None":
+    ) -> SamplingParam | None:
         """Build a SamplingParam from preset defaults.
 
         Returns ``None`` when no preset is configured for

fastvideo/api/schema.py

@@ -33,8 +33,25 @@ class OffloadConfig:
 
 @dataclass
 class CompileConfig:
+    """Typed ``torch.compile`` configuration.
+
+    ``backend``/``fullgraph``/``mode``/``dynamic`` are the four most
+    common ``torch.compile`` knobs. ``extras`` holds any remaining
+    ``torch.compile`` kwargs (e.g. ``options``, ``disable``).
+    """
+
     enabled: bool = False
-    kwargs: dict[str, Any] = field(default_factory=dict)
+    text_encoder_enabled: bool | None = None
+    """Whether ``torch.compile`` is applied to the text encoder. ``None``
+    keeps the runtime default. The public ``FastVideoArgs`` adapter does
+    not yet consume this flag; reserved so the realtime runtime upstream
+    (PR 7.6) has a typed home for its ``enable_torch_compile_text_encoder``
+    kwarg without routing through ``pipeline.experimental``."""
+    backend: str | None = None
+    fullgraph: bool | None = None
+    mode: str | None = None
+    dynamic: bool | None = None
+    extras: dict[str, Any] = field(default_factory=dict)
 
 
 @dataclass
@@ -76,6 +93,8 @@ class PipelineSelection:
     preset: str | None = None
     preset_version: int | None = None
     components: ComponentConfig = field(default_factory=ComponentConfig)
+    vae_tiling: bool | None = None
+    """Tile-based VAE decode. ``None`` keeps the model's default."""
     preset_overrides: dict[str, Any] = field(default_factory=dict)
     experimental: dict[str, Any] = field(default_factory=dict)
 

fastvideo/configs/pipelines/__init__.py

@@ -5,7 +5,7 @@ from fastvideo.configs.pipelines.hunyuan import FastHunyuanConfig, HunyuanConfig
 from fastvideo.configs.pipelines.hunyuan15 import Hunyuan15T2V480PConfig, Hunyuan15T2V720PConfig
 from fastvideo.configs.pipelines.hunyuangamecraft import HunyuanGameCraftPipelineConfig
 from fastvideo.configs.pipelines.hyworld import HYWorldConfig
-from fastvideo.configs.pipelines.ltx2 import LTX2T2VConfig
+from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig
 from fastvideo.registry import get_pipeline_config_cls_from_name
 from fastvideo.configs.pipelines.wan import (SelfForcingWanT2V480PConfig, WanI2V480PConfig, WanI2V720PConfig,
                                              WanT2V480PConfig, WanT2V720PConfig)

fastvideo/configs/pipelines/longcat.py

@@ -11,10 +11,34 @@ import torch
 from fastvideo.configs.models import DiTConfig, VAEConfig
 from fastvideo.configs.models.dits.base import DiTArchConfig
 from fastvideo.configs.models.encoders import BaseEncoderOutput, T5Config
+from fastvideo.configs.models.encoders.base import TextEncoderArchConfig
+from fastvideo.configs.models.encoders.t5 import T5ArchConfig
 from fastvideo.configs.models.vaes import WanVAEConfig
 from fastvideo.configs.pipelines.base import PipelineConfig
 
 
+@dataclass
+class LongCatT5ArchConfig(T5ArchConfig):
+    """T5 arch that pads tokenizer output to ``max_length``.
+
+    LongCat's denoising stage concatenates positive and negative
+    attention masks along the batch dimension for CFG, which requires
+    uniform seq length. The shared :class:`T5ArchConfig` dropped the
+    ``"padding": "max_length"`` tokenizer kwarg so other DiTs could run
+    with variable-length masks; LongCat still needs the uniform
+    contract.
+    """
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.tokenizer_kwargs["padding"] = "max_length"
+
+
+@dataclass
+class LongCatT5Config(T5Config):
+    arch_config: TextEncoderArchConfig = field(default_factory=LongCatT5ArchConfig)
+
+
 @dataclass
 class LongCatDiTArchConfig(DiTArchConfig):
     """Extended DiTArchConfig with LongCat-specific fields."""
@@ -103,8 +127,9 @@ class LongCatT2V480PConfig(PipelineConfig):
     vae_precision: str = "bf16"
     text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16", ))
 
-    # Text encoding (UMT5 uses T5-like config; postprocess to fixed 512)
-    text_encoder_configs: tuple[T5Config, ...] = field(default_factory=lambda: (T5Config(), ))
+    # UMT5 uses T5-like config; postprocess pads to 512. LongCatT5Config
+    # restores ``padding="max_length"`` for the CFG concat contract.
+    text_encoder_configs: tuple[T5Config, ...] = field(default_factory=lambda: (LongCatT5Config(), ))
     preprocess_text_funcs: tuple[Callable[[str], str], ...] = field(default_factory=lambda: (longcat_preprocess_text, ))
     postprocess_text_funcs: tuple[Callable[[BaseEncoderOutput], torch.Tensor],
                                   ...] = field(default_factory=lambda: (umt5_postprocess_text, ))

fastvideo/entrypoints/cli/eval.py

@@ -0,0 +1,216 @@
+"""``fastvideo eval`` CLI: list registered eval metrics and run them
+against a set of videos.
+
+This is a thin wrapper around :mod:`fastvideo.eval`. Heavy lifting
+(metric loading, GPU handling, batching) lives in
+:func:`fastvideo.eval.create_evaluator`.
+
+Examples::
+
+    fastvideo eval list
+    fastvideo eval list --group vbench
+    fastvideo eval run --videos path/to/videos/*.mp4 \\
+        --metrics common.ssim --reference path/to/refs/
+    fastvideo eval run --videos clip.mp4 --metrics vbench.aesthetic_quality \\
+        --output scores.json
+"""
+
+from __future__ import annotations
+
+import argparse
+import glob
+import json
+from pathlib import Path
+from typing import cast
+
+from fastvideo.entrypoints.cli.cli_types import CLISubcommand
+from fastvideo.logger import init_logger
+from fastvideo.utils import FlexibleArgumentParser
+
+logger = init_logger(__name__)
+
+
+class EvalSubcommand(CLISubcommand):
+    """The ``eval`` subcommand — entry point for the eval suite."""
+
+    def __init__(self) -> None:
+        self.name = "eval"
+        super().__init__()
+
+    def cmd(self, args: argparse.Namespace) -> None:
+        action = getattr(args, "eval_action", None)
+        if action == "list":
+            _cmd_list(args)
+        elif action == "run":
+            _cmd_run(args)
+        else:
+            # Re-print help if no action was given.
+            self._parser.print_help()  # type: ignore[attr-defined]
+
+    def validate(self, args: argparse.Namespace) -> None:
+        action = getattr(args, "eval_action", None)
+        if action == "run" and not args.videos:
+            raise SystemExit("`fastvideo eval run` requires --videos")
+
+    def subparser_init(self, subparsers: argparse._SubParsersAction) -> FlexibleArgumentParser:
+        eval_parser = subparsers.add_parser(
+            "eval",
+            help="Run video-gen evaluation metrics",
+            usage="fastvideo eval {list,run} [...]",
+        )
+        sub = eval_parser.add_subparsers(dest="eval_action", required=False)
+
+        # `eval list`
+        list_p = sub.add_parser("list", help="List registered metrics")
+        list_p.add_argument("--group", type=str, default=None, help="Filter to a metric group (e.g. 'vbench').")
+
+        # `eval run`
+        run_p = sub.add_parser("run", help="Evaluate videos against one or more metrics")
+        run_p.add_argument("--videos",
+                           type=str,
+                           nargs="+",
+                           required=False,
+                           help="Path, glob, or directory of generated videos.")
+        run_p.add_argument("--reference",
+                           type=str,
+                           default=None,
+                           help="Path / glob / dir of reference videos (for paired metrics).")
+        run_p.add_argument("--metrics", type=str, default="all", help="Comma-separated metric names, or 'all'.")
+        run_p.add_argument("--device", type=str, default="cuda", help="Torch device (e.g. 'cuda', 'cuda:0', 'cpu').")
+        run_p.add_argument("--text-prompt",
+                           type=str,
+                           nargs="*",
+                           default=None,
+                           help="Prompt(s) for text-conditioned metrics. One per video.")
+        run_p.add_argument("--fps", type=float, default=None, help="Frame-rate annotation passed to fps-aware metrics.")
+        run_p.add_argument("--output",
+                           type=str,
+                           default=None,
+                           help="Write results as JSON to this path (default: stdout).")
+
+        # Stash the parser so cmd() can re-print help on no-action.
+        self._parser = eval_parser  # type: ignore[attr-defined]
+        return cast(FlexibleArgumentParser, eval_parser)
+
+
+def _cmd_list(args: argparse.Namespace) -> None:
+    from fastvideo.eval import list_metrics
+    names = list_metrics()
+    if args.group:
+        prefix = args.group.rstrip(".") + "."
+        names = [n for n in names if n == args.group or n.startswith(prefix)]
+    if not names:
+        print(f"(no metrics matched group {args.group!r})")
+        return
+    for name in names:
+        print(name)
+
+
+def _cmd_run(args: argparse.Namespace) -> None:
+    from fastvideo.eval import create_evaluator
+    from fastvideo.eval.io import load_video
+
+    video_paths = _expand_paths(args.videos)
+    if not video_paths:
+        raise SystemExit(f"No videos matched: {args.videos}")
+    ref_paths = _expand_paths([args.reference]) if args.reference else None
+
+    metrics_arg: list[str] | str = ("all" if args.metrics == "all" else
+                                    [m.strip() for m in args.metrics.split(",") if m.strip()])
+
+    evaluator = create_evaluator(metrics=metrics_arg, device=args.device)
+
+    all_results: list[dict] = []
+    for i, vp in enumerate(video_paths):
+        logger.info("Evaluating %s (%d/%d)", vp, i + 1, len(video_paths))
+        kwargs: dict = {"video": load_video(vp)}
+        if ref_paths is not None:
+            ref = ref_paths[i] if i < len(ref_paths) else ref_paths[0]
+            kwargs["reference"] = load_video(ref)
+        if args.text_prompt is not None:
+            prompt = (args.text_prompt[i] if i < len(args.text_prompt) else args.text_prompt[0])
+            kwargs["text_prompt"] = [prompt]
+        if args.fps is not None:
+            kwargs["fps"] = args.fps
+
+        results = evaluator.evaluate(**kwargs)
+        all_results.append({
+            "video": str(vp),
+            "scores": _serialize_results(results),
+        })
+
+    payload = json.dumps(all_results, indent=2, default=_jsonable)
+    if args.output:
+        Path(args.output).write_text(payload)
+        logger.info("Wrote results to %s", args.output)
+    else:
+        print(payload)
+
+
+def _expand_paths(patterns: list[str]) -> list[str]:
+    out: list[str] = []
+    for pat in patterns:
+        p = Path(pat)
+        if p.is_dir():
+            for ext in (".mp4", ".avi", ".mov", ".mkv", ".gif"):
+                out.extend(sorted(str(f) for f in p.iterdir() if f.suffix.lower() == ext))
+        elif any(c in pat for c in "*?["):
+            out.extend(sorted(glob.glob(pat)))
+        else:
+            out.append(pat)
+    # de-dup, preserve order
+    seen: set[str] = set()
+    deduped: list[str] = []
+    for x in out:
+        if x not in seen:
+            seen.add(x)
+            deduped.append(x)
+    return deduped
+
+
+def _serialize_results(results) -> dict | list:
+    """Turn evaluator output (dict or list-of-dicts) into JSON-friendly form."""
+    if isinstance(results, list):
+        return [_serialize_results(r) for r in results]
+    if isinstance(results, dict):
+        return {k: _serialize_metric_result(v) for k, v in results.items()}
+    return _serialize_metric_result(results)
+
+
+def _serialize_metric_result(mr) -> dict:
+    return {
+        "name": getattr(mr, "name", None),
+        "score": getattr(mr, "score", None),
+        "details": getattr(mr, "details", None),
+    }
+
+
+def _jsonable(obj):
+    """``json.dumps(default=...)`` coercer for metric outputs.
+
+    Metrics frequently land numpy scalars / arrays, torch tensors, and
+    pathlib paths inside ``MetricResult.details`` (e.g. ``optical_flow``
+    populates ``per_frame_metrics`` with numpy floats). The stdlib JSON
+    encoder rejects all of those by default — this callback walks the
+    leaves and coerces them to native Python types.
+    """
+    import numpy as np
+    import torch
+
+    if isinstance(obj, np.integer):
+        return int(obj)
+    if isinstance(obj, np.floating):
+        return float(obj)
+    if isinstance(obj, np.bool_):
+        return bool(obj)
+    if isinstance(obj, np.ndarray):
+        return obj.tolist()
+    if isinstance(obj, torch.Tensor):
+        return obj.detach().cpu().tolist()
+    if isinstance(obj, Path):
+        return str(obj)
+    raise TypeError(f"Object of type {type(obj).__name__} is not JSON serializable")
+
+
+def cmd_init() -> list[CLISubcommand]:
+    return [EvalSubcommand()]

fastvideo/entrypoints/cli/main.py

@@ -5,6 +5,7 @@ from fastvideo.entrypoints.cli.generate import cmd_init as generate_cmd_init
 from fastvideo.utils import FlexibleArgumentParser
 from fastvideo.entrypoints.cli.serve import cmd_init as serve_cmd_init
 from fastvideo.entrypoints.cli.bench import cmd_init as bench_cmd_init
+from fastvideo.entrypoints.cli.eval import cmd_init as eval_cmd_init
 
 
 def cmd_init() -> list[CLISubcommand]:
@@ -13,6 +14,7 @@ def cmd_init() -> list[CLISubcommand]:
     commands.extend(generate_cmd_init())
     commands.extend(serve_cmd_init())
     commands.extend(bench_cmd_init())
+    commands.extend(eval_cmd_init())
     return commands
 
 

fastvideo/entrypoints/streaming/__init__.py

@@ -1,4 +1,31 @@
 # SPDX-License-Identifier: Apache-2.0
-from fastvideo.entrypoints.streaming.server import run_server
+from fastvideo.entrypoints.streaming.server import build_app, run_server
+from fastvideo.entrypoints.streaming.session import (
+    Session,
+    SessionManager,
+    SessionState,
+)
+from fastvideo.entrypoints.streaming.session_store import (
+    BlobStore,
+    InMemoryBlobStore,
+    InMemorySessionStore,
+    SessionStore,
+)
+from fastvideo.entrypoints.streaming.stream import (
+    FragmentedMP4Chunk,
+    FragmentedMP4Encoder,
+)
 
-__all__ = ["run_server"]
+__all__ = [
+    "BlobStore",
+    "FragmentedMP4Chunk",
+    "FragmentedMP4Encoder",
+    "InMemoryBlobStore",
+    "InMemorySessionStore",
+    "Session",
+    "SessionManager",
+    "SessionState",
+    "SessionStore",
+    "build_app",
+    "run_server",
+]

fastvideo/entrypoints/streaming/protocol.py

@@ -0,0 +1,252 @@
+# SPDX-License-Identifier: Apache-2.0
+"""JSON WebSocket protocol schemas for the streaming server.
+
+Every control message shares the envelope ``{"type": <str>, ...}``.
+Pydantic models live here so the server can parse / validate incoming
+frames and emit well-typed outgoing frames without hand-rolled dicts.
+
+The message catalogue matches the contract in
+``docs/design/server_contracts/streaming.md``; additions must land in
+both places in the same PR.
+"""
+from __future__ import annotations
+
+from typing import Annotated, Any, Literal, Union
+
+from pydantic import BaseModel, ConfigDict, Field
+
+# ---------------------------------------------------------------------------
+# Client → server
+# ---------------------------------------------------------------------------
+
+
+class SessionInitV2(BaseModel):
+    """Opening frame the client sends after the WebSocket handshake."""
+
+    model_config = ConfigDict(extra="allow")
+
+    type: Literal["session_init_v2"]
+    client_id: str | None = None
+    preset: str | None = None
+    preset_label: str | None = None
+    curated_prompts: list[str] = Field(default_factory=list)
+    initial_image: dict[str, Any] | None = None
+    enhancement_enabled: bool = False
+    auto_extension_enabled: bool = False
+    loop_generation_enabled: bool = False
+    single_clip_mode: bool = False
+    stream_mode: Literal["av_fmp4", "legacy_jpeg"] = "av_fmp4"
+    continuation_state: dict[str, Any] | None = None
+    """Optional ``{kind, payload}`` dict; hydrated into
+    :class:`fastvideo.api.ContinuationState` server-side."""
+
+
+class SegmentPromptSource(BaseModel):
+    """Request a new segment using a specific prompt."""
+
+    type: Literal["segment_prompt_source"]
+    prompt: str
+    negative_prompt: str | None = None
+    source: Literal["curated", "enhanced", "user", "auto_extension"] = "user"
+    seed: int | None = None
+    num_inference_steps: int | None = None
+    guidance_scale: float | None = None
+
+
+class SeedPromptsUpdated(BaseModel):
+    type: Literal["seed_prompts_updated"]
+    seed_prompts: list[str] = Field(default_factory=list)
+
+
+class EnhancementUpdated(BaseModel):
+    type: Literal["enhancement_updated"]
+    enabled: bool
+
+
+class AutoExtensionUpdated(BaseModel):
+    type: Literal["auto_extension_updated"]
+    enabled: bool
+
+
+class LoopGenerationUpdated(BaseModel):
+    type: Literal["loop_generation_updated"]
+    enabled: bool
+
+
+class GenerationPausedUpdated(BaseModel):
+    type: Literal["generation_paused_updated"]
+    paused: bool
+
+
+class SnapshotState(BaseModel):
+    """Request the current ``ContinuationState`` for export."""
+
+    type: Literal["snapshot_state"]
+
+
+ClientMessage = Annotated[
+    Union[  # noqa: UP007 - Annotated requires Union for discriminator
+        SessionInitV2,
+        SegmentPromptSource,
+        SeedPromptsUpdated,
+        EnhancementUpdated,
+        AutoExtensionUpdated,
+        LoopGenerationUpdated,
+        GenerationPausedUpdated,
+        SnapshotState,
+    ],
+    Field(discriminator="type"),
+]
+
+# ---------------------------------------------------------------------------
+# Server → client
+# ---------------------------------------------------------------------------
+
+
+class QueueStatus(BaseModel):
+    type: Literal["queue_status"] = "queue_status"
+    position: int
+    queue_depth: int
+
+
+class GpuAssigned(BaseModel):
+    type: Literal["gpu_assigned"] = "gpu_assigned"
+    gpu_id: int
+    session_timeout: int
+
+
+class Ltx2StreamStart(BaseModel):
+    type: Literal["ltx2_stream_start"] = "ltx2_stream_start"
+    preset: str | None = None
+    width: int
+    height: int
+    fps: int
+    num_frames: int
+
+
+class Ltx2SegmentStart(BaseModel):
+    type: Literal["ltx2_segment_start"] = "ltx2_segment_start"
+    segment_idx: int
+    prompt: str
+    total_steps: int
+
+
+class StepComplete(BaseModel):
+    type: Literal["step_complete"] = "step_complete"
+    segment_idx: int
+    step: int
+    total_steps: int
+    stage: str = "denoise"
+
+
+class MediaInit(BaseModel):
+    """Descriptor for the fMP4 initialization segment that follows."""
+
+    type: Literal["media_init"] = "media_init"
+    segment_idx: int
+    mime: str = "video/mp4; codecs=\"avc1.64001f, mp4a.40.2\""
+    stream_id: str
+    mode: Literal["av_fmp4"] = "av_fmp4"
+
+
+class MediaSegmentComplete(BaseModel):
+    type: Literal["media_segment_complete"] = "media_segment_complete"
+    segment_idx: int
+    stream_id: str
+    chunks: int
+    duration_ms: float | None = None
+    pts_base_ms: float | None = None
+
+
+class Ltx2SegmentComplete(BaseModel):
+    type: Literal["ltx2_segment_complete"] = "ltx2_segment_complete"
+    segment_idx: int
+    generation_time_ms: float
+    e2e_latency_ms: float | None = None
+
+
+class Ltx2StreamComplete(BaseModel):
+    type: Literal["ltx2_stream_complete"] = "ltx2_stream_complete"
+    reason: Literal["segment_cap", "stop_requested", "error"] = "stop_requested"
+
+
+class SessionTimeout(BaseModel):
+    type: Literal["session_timeout"] = "session_timeout"
+    timeout_seconds: int
+
+
+class ContinuationStateSnapshot(BaseModel):
+    type: Literal["continuation_state_snapshot"] = "continuation_state_snapshot"
+    state: dict[str, Any]
+    """``{kind, payload}`` dict matching
+    :class:`fastvideo.api.ContinuationState`."""
+
+
+class ErrorMessage(BaseModel):
+    type: Literal["error"] = "error"
+    code: Literal[
+        "session_rejected",
+        "invalid_message",
+        "preset_mismatch",
+        "gpu_unavailable",
+        "worker_failed",
+        "upstream_timeout",
+        "internal_error",
+    ] = "internal_error"
+    message: str
+    retryable: bool = False
+
+
+ServerMessage = Union[  # noqa: UP007 - pydantic Union handling
+    QueueStatus,
+    GpuAssigned,
+    Ltx2StreamStart,
+    Ltx2SegmentStart,
+    StepComplete,
+    MediaInit,
+    MediaSegmentComplete,
+    Ltx2SegmentComplete,
+    Ltx2StreamComplete,
+    SessionTimeout,
+    ContinuationStateSnapshot,
+    ErrorMessage,
+]
+
+
+def parse_client_message(raw: dict[str, Any]) -> ClientMessage:
+    """Parse an incoming WebSocket dict into a typed client message.
+
+    Unknown ``type`` values raise :class:`pydantic.ValidationError`; the
+    server handler turns that into an ``error`` frame with
+    ``code="invalid_message"``.
+    """
+    from pydantic import TypeAdapter
+
+    return TypeAdapter(ClientMessage).validate_python(raw)
+
+
+__all__ = [
+    "AutoExtensionUpdated",
+    "ClientMessage",
+    "ContinuationStateSnapshot",
+    "EnhancementUpdated",
+    "ErrorMessage",
+    "GenerationPausedUpdated",
+    "GpuAssigned",
+    "Ltx2SegmentComplete",
+    "Ltx2SegmentStart",
+    "Ltx2StreamComplete",
+    "Ltx2StreamStart",
+    "LoopGenerationUpdated",
+    "MediaInit",
+    "MediaSegmentComplete",
+    "QueueStatus",
+    "SeedPromptsUpdated",
+    "SegmentPromptSource",
+    "ServerMessage",
+    "SessionInitV2",
+    "SessionTimeout",
+    "SnapshotState",
+    "StepComplete",
+    "parse_client_message",
+]

fastvideo/entrypoints/streaming/server.py

@@ -1,15 +1,531 @@
 # SPDX-License-Identifier: Apache-2.0
+"""Single-generator FastAPI + WebSocket streaming server."""
 from __future__ import annotations
 
-from fastvideo.api.schema import ServeConfig
+import asyncio
+import contextlib
+import os
+import time
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+from fastapi import FastAPI, WebSocket, WebSocketDisconnect
+from fastapi.responses import JSONResponse
+
+from fastvideo.api.schema import (
+    ContinuationState,
+    GenerationRequest,
+    InputConfig,
+    OutputConfig,
+    SamplingConfig,
+    ServeConfig,
+)
+from fastvideo.entrypoints.streaming.protocol import (
+    AutoExtensionUpdated,
+    ContinuationStateSnapshot,
+    EnhancementUpdated,
+    ErrorMessage,
+    GenerationPausedUpdated,
+    GpuAssigned,
+    LoopGenerationUpdated,
+    Ltx2SegmentComplete,
+    Ltx2SegmentStart,
+    Ltx2StreamComplete,
+    Ltx2StreamStart,
+    MediaInit,
+    MediaSegmentComplete,
+    QueueStatus,
+    SeedPromptsUpdated,
+    SegmentPromptSource,
+    SessionInitV2,
+    SnapshotState,
+    StepComplete,
+    parse_client_message,
+)
+from fastvideo.entrypoints.streaming.session import (
+    InvalidSessionTransition,
+    Session,
+    SessionManager,
+    SessionRejected,
+    SessionState,
+)
+from fastvideo.entrypoints.streaming.session_init_image import (
+    persist_session_init_image, )
+from fastvideo.entrypoints.streaming.session_store import (
+    InMemorySessionStore,
+    SessionStore,
+)
+from fastvideo.entrypoints.streaming.stream import FragmentedMP4Encoder
 from fastvideo.logger import init_logger
 
 logger = init_logger(__name__)
 
+# RFC 6455 WebSocket close codes used by the server.
+_WS_CLOSE_UNSUPPORTED_DATA = 1003
+_WS_CLOSE_TRY_AGAIN_LATER = 1013
 
-def run_server(serve_config: ServeConfig) -> None:
-    """Launch the streaming (WebSocket / Dynamo) server."""
+
+class _GeneratorProto(Protocol):
+    """Subset of :class:`fastvideo.VideoGenerator` the server calls."""
+
+    def generate(self, request: GenerationRequest) -> Any:
+        ...
+
+
+@dataclass
+class ServerState:
+    serve_config: ServeConfig
+    generator: _GeneratorProto
+    sessions: SessionManager
+    session_store: SessionStore
+
+
+def build_app(
+    serve_config: ServeConfig,
+    generator: _GeneratorProto,
+    *,
+    session_store: SessionStore | None = None,
+) -> FastAPI:
+    """Build the FastAPI app used by :func:`run_server`.
+
+    Exposed so tests can drive the WebSocket endpoint in-process via
+    ``starlette.testclient.TestClient(app).websocket_connect(...)``.
+    """
+    if serve_config.streaming is None:
+        raise ValueError("ServeConfig.streaming must be set to launch the streaming "
+                         "server; got None. Add a `streaming:` block to your serve config.")
+
+    sessions = SessionManager(
+        segment_cap=serve_config.streaming.generation_segment_cap,
+        session_timeout_seconds=serve_config.streaming.session_timeout_seconds,
+    )
+    state = ServerState(
+        serve_config=serve_config,
+        generator=generator,
+        sessions=sessions,
+        session_store=session_store or InMemorySessionStore(),
+    )
+
+    app = FastAPI(title="FastVideo Streaming")
+
+    @app.get("/health")
+    async def _health() -> JSONResponse:
+        return JSONResponse({
+            "status": "ok",
+            "sessions": len(state.sessions),
+            "stream_mode": state.serve_config.streaming.stream_mode,
+        })
+
+    @app.websocket("/v1/stream")
+    async def _stream(websocket: WebSocket) -> None:
+        await websocket.accept()
+        try:
+            session = state.sessions.create()
+        except SessionRejected as exc:
+            await _send_error(websocket, "session_rejected", str(exc), retryable=False)
+            await websocket.close(code=_WS_CLOSE_TRY_AGAIN_LATER, reason="session_rejected")
+            return
+
+        try:
+            await _handle_session(websocket, session, state)
+        except WebSocketDisconnect:
+            logger.info("session %s: client disconnected", session.id[:8])
+        except Exception:  # pragma: no cover - defensive catch-all
+            logger.exception("session %s: unhandled error", session.id[:8])
+            with contextlib.suppress(InvalidSessionTransition):
+                session.transition(SessionState.ERROR)
+        finally:
+            _cleanup_session(session, state)
+
+    app.state.server_state = state
+    return app
+
+
+def run_server(serve_config: ServeConfig, *, generator: _GeneratorProto | None = None) -> None:
+    """Launch the streaming server.
+
+    Boots a :class:`fastvideo.VideoGenerator` from
+    ``serve_config.generator`` unless ``generator`` is provided, then
+    serves ``build_app(...)`` via uvicorn.
+    """
     if serve_config.streaming is None:
         raise ValueError("ServeConfig.streaming must be set to launch the streaming server; "
                          "got None. Add a `streaming:` block to your serve config.")
-    raise NotImplementedError("streaming server is not implemented yet")
+
+    import uvicorn
+
+    if generator is None:
+        from fastvideo import VideoGenerator  # lazy to avoid boot cost
+
+        generator = VideoGenerator.from_pretrained(config=serve_config.generator)
+    app = build_app(serve_config, generator)
+    uvicorn.run(
+        app,
+        host=serve_config.server.host,
+        port=serve_config.server.port,
+    )
+
+
+async def _handle_session(
+    websocket: WebSocket,
+    session: Session,
+    state: ServerState,
+) -> None:
+    init = await _read_init_message(websocket, session, state)
+    if init is None:
+        return
+
+    await _apply_session_init(session, init, state)
+    await _send_json(websocket, QueueStatus(position=0, queue_depth=0))
+    session.transition(SessionState.GPU_BINDING)
+    await _send_json(websocket, GpuAssigned(
+        gpu_id=0,
+        session_timeout=state.sessions.session_timeout_seconds,
+    ))
+    session.transition(SessionState.ACTIVE)
+    await _send_json(websocket, _build_stream_start(session, state))
+
+    try:
+        await _run_segment_loop(websocket, session, state)
+    finally:
+        with contextlib.suppress(RuntimeError):
+            await _send_json(websocket, Ltx2StreamComplete(reason="stop_requested"))
+
+
+async def _read_init_message(
+    websocket: WebSocket,
+    session: Session,
+    state: ServerState,
+) -> SessionInitV2 | None:
+    try:
+        raw = await asyncio.wait_for(
+            websocket.receive_json(),
+            timeout=state.sessions.session_timeout_seconds,
+        )
+    except asyncio.TimeoutError:
+        logger.info("session %s: init timeout", session.id[:8])
+        with contextlib.suppress(InvalidSessionTransition):
+            session.transition(SessionState.TIMEOUT)
+        return None
+    except WebSocketDisconnect:
+        return None
+    try:
+        parsed = parse_client_message(raw)
+    except Exception as exc:
+        await _reject_init(websocket, session, f"opening frame failed validation: {exc}", "invalid_init")
+        return None
+    if not isinstance(parsed, SessionInitV2):
+        await _reject_init(websocket, session, "first frame must be session_init_v2", "expected_session_init_v2")
+        return None
+    return parsed
+
+
+async def _reject_init(
+    websocket: WebSocket,
+    session: Session,
+    message: str,
+    close_reason: str,
+) -> None:
+    await _send_error(websocket, "invalid_message", message, retryable=False)
+    await websocket.close(code=_WS_CLOSE_UNSUPPORTED_DATA, reason=close_reason)
+    with contextlib.suppress(InvalidSessionTransition):
+        session.transition(SessionState.REJECTED)
+
+
+async def _apply_session_init(
+    session: Session,
+    init: SessionInitV2,
+    state: ServerState,
+) -> None:
+    session.client_id = init.client_id
+    session.preset = init.preset
+    session.preset_label = init.preset_label
+    session.curated_prompts = list(init.curated_prompts)
+    session.enhancement_enabled = init.enhancement_enabled
+    session.auto_extension_enabled = init.auto_extension_enabled
+    session.loop_generation_enabled = init.loop_generation_enabled
+    session.single_clip_mode = init.single_clip_mode
+    session.stream_mode = init.stream_mode
+
+    if init.initial_image is not None:
+        # Decode + disk write off the event loop; payload is up to 32 MiB.
+        image = await asyncio.to_thread(persist_session_init_image, init.initial_image)
+        if image is not None:
+            session.metadata["session_init_image"] = image.path
+
+    if init.continuation_state is not None:
+        session.continuation_state = _coerce_state(init.continuation_state)
+        if session.continuation_state is not None:
+            state.session_store.store(session.id, session.continuation_state)
+
+
+async def _run_segment_loop(
+    websocket: WebSocket,
+    session: Session,
+    state: ServerState,
+) -> None:
+    cap = state.sessions.segment_cap
+    while True:
+        if session.segment_cap_reached(cap):
+            logger.info("session %s: segment cap (%d) reached", session.id[:8], cap)
+            return
+
+        try:
+            raw = await asyncio.wait_for(
+                websocket.receive_json(),
+                timeout=state.sessions.session_timeout_seconds,
+            )
+        except asyncio.TimeoutError:
+            logger.info("session %s: idle timeout", session.id[:8])
+            with contextlib.suppress(InvalidSessionTransition):
+                session.transition(SessionState.TIMEOUT)
+            return
+        except WebSocketDisconnect:
+            return
+        session.touch()
+
+        try:
+            parsed = parse_client_message(raw)
+        except Exception as exc:
+            await _send_error(websocket, "invalid_message", str(exc), retryable=True)
+            continue
+
+        if isinstance(parsed, SnapshotState):
+            snap = state.session_store.snapshot(session.id)
+            if snap is None:
+                await _send_error(websocket,
+                                  "internal_error",
+                                  "no continuation state available for session",
+                                  retryable=False)
+                continue
+            await _send_json(websocket, ContinuationStateSnapshot(state={"kind": snap.kind, "payload": snap.payload}, ))
+            continue
+
+        if isinstance(parsed, SegmentPromptSource):
+            await _run_segment(websocket, session, state, parsed)
+            continue
+
+        # Silently ignore unknown-but-valid types (additive-evolution
+        # rule in streaming.md).
+        _apply_toggle(session, parsed)
+
+
+async def _run_segment(
+    websocket: WebSocket,
+    session: Session,
+    state: ServerState,
+    message: SegmentPromptSource,
+) -> None:
+    request = _build_generation_request(session, message, state)
+    segment_idx = session.segment_idx
+    await _send_json(
+        websocket,
+        Ltx2SegmentStart(
+            segment_idx=segment_idx,
+            prompt=message.prompt,
+            total_steps=request.sampling.num_inference_steps,
+        ))
+
+    start = time.perf_counter()
+    loop = asyncio.get_running_loop()
+    # TODO: executor-wrapped generate() cannot be cancelled, so a
+    # client disconnect mid-segment leaves the GPU work running to
+    # completion. Real cancellation needs the generate_async API.
+    try:
+        result = await loop.run_in_executor(None, state.generator.generate, request)
+    except Exception as exc:
+        logger.exception("session %s: generator failed", session.id[:8])
+        await _send_error(websocket, "worker_failed", f"generator.generate failed: {exc}", retryable=True)
+        with contextlib.suppress(InvalidSessionTransition):
+            session.transition(SessionState.ERROR)
+        return
+    elapsed_ms = (time.perf_counter() - start) * 1000.0
+
+    frames = _extract_frames(result)
+    if not frames:
+        await _send_error(websocket, "worker_failed", "generator returned no frames", retryable=True)
+        with contextlib.suppress(InvalidSessionTransition):
+            session.transition(SessionState.ERROR)
+        return
+
+    # Synchronous generator call has no per-step hook; emit one
+    # terminal StepComplete so observability wiring still sees the
+    # segment finish.
+    total = request.sampling.num_inference_steps
+    await _send_json(websocket, StepComplete(
+        segment_idx=segment_idx,
+        step=total,
+        total_steps=total,
+        stage="denoise",
+    ))
+
+    encoder = FragmentedMP4Encoder(
+        width=request.sampling.width,
+        height=request.sampling.height,
+        fps=request.sampling.fps,
+        segment_idx=segment_idx,
+    )
+    chunks_relayed = 0
+    async with encoder:
+        init_sent = False
+        async for chunk in encoder.encode(frames):
+            if chunk.kind == "init":
+                await _send_json(websocket, MediaInit(
+                    segment_idx=segment_idx,
+                    stream_id=chunk.stream_id,
+                ))
+                init_sent = True
+            await websocket.send_bytes(chunk.data)
+            if init_sent and chunk.kind == "media":
+                chunks_relayed += 1
+
+    await _send_json(
+        websocket,
+        MediaSegmentComplete(
+            segment_idx=segment_idx,
+            stream_id=encoder.stream_id,
+            chunks=chunks_relayed,
+            duration_ms=float(request.sampling.num_frames) / request.sampling.fps * 1000.0,
+        ))
+
+    new_state = _extract_state(result)
+    if new_state is not None:
+        session.continuation_state = new_state
+        state.session_store.store(session.id, new_state)
+
+    session.segment_idx += 1
+    with contextlib.suppress(InvalidSessionTransition):
+        session.transition(SessionState.ACTIVE)
+
+    await _send_json(
+        websocket,
+        Ltx2SegmentComplete(
+            segment_idx=segment_idx,
+            generation_time_ms=elapsed_ms,
+            e2e_latency_ms=elapsed_ms,
+        ))
+
+
+def _build_stream_start(
+    session: Session,
+    state: ServerState,
+) -> Ltx2StreamStart:
+    default = state.serve_config.default_request
+    return Ltx2StreamStart(
+        preset=session.preset,
+        width=default.sampling.width,
+        height=default.sampling.height,
+        fps=default.sampling.fps,
+        num_frames=default.sampling.num_frames,
+    )
+
+
+def _build_generation_request(
+    session: Session,
+    message: SegmentPromptSource,
+    state: ServerState,
+) -> GenerationRequest:
+    # Start from the operator-pinned default_request to pick up the
+    # preset-selected sampling knobs; override with per-message values.
+    base = state.serve_config.default_request
+    sampling_kwargs: dict[str, Any] = {
+        "num_videos_per_prompt":
+        base.sampling.num_videos_per_prompt,
+        "seed":
+        message.seed if message.seed is not None else base.sampling.seed,
+        "num_frames":
+        base.sampling.num_frames,
+        "height":
+        base.sampling.height,
+        "width":
+        base.sampling.width,
+        "fps":
+        base.sampling.fps,
+        "num_inference_steps":
+        (message.num_inference_steps if message.num_inference_steps is not None else base.sampling.num_inference_steps),
+        "guidance_scale":
+        (message.guidance_scale if message.guidance_scale is not None else base.sampling.guidance_scale),
+    }
+    request = GenerationRequest(
+        prompt=message.prompt,
+        negative_prompt=message.negative_prompt or base.negative_prompt,
+        inputs=InputConfig(image_path=session.metadata.get("session_init_image"), ),
+        sampling=SamplingConfig(**sampling_kwargs),
+        output=OutputConfig(save_video=False, return_frames=True, return_state=True),
+        state=session.continuation_state,
+    )
+    return request
+
+
+def _coerce_state(raw: dict[str, Any]) -> ContinuationState | None:
+    kind = raw.get("kind")
+    payload = raw.get("payload")
+    if not isinstance(kind, str) or not isinstance(payload, dict):
+        return None
+    return ContinuationState(kind=kind, payload=payload)
+
+
+def _apply_toggle(session: Session, message: Any) -> None:
+    if isinstance(message, EnhancementUpdated):
+        session.enhancement_enabled = message.enabled
+    elif isinstance(message, AutoExtensionUpdated):
+        session.auto_extension_enabled = message.enabled
+    elif isinstance(message, LoopGenerationUpdated):
+        session.loop_generation_enabled = message.enabled
+    elif isinstance(message, GenerationPausedUpdated):
+        session.generation_paused = message.paused
+    elif isinstance(message, SeedPromptsUpdated):
+        session.curated_prompts = list(message.seed_prompts)
+
+
+def _extract_frames(result: Any) -> list:
+    if hasattr(result, "frames"):
+        return list(result.frames or [])
+    if isinstance(result, dict):
+        return list(result.get("frames") or [])
+    return []
+
+
+def _extract_state(result: Any) -> ContinuationState | None:
+    state = getattr(result, "state", None)
+    if state is None and isinstance(result, dict):
+        state = result.get("state")
+    if isinstance(state, ContinuationState):
+        return state
+    if isinstance(state, dict):
+        return _coerce_state(state)
+    return None
+
+
+async def _send_json(websocket: WebSocket, message: Any) -> None:
+    payload = (message.model_dump(mode="json", exclude_none=True) if hasattr(message, "model_dump") else message)
+    await websocket.send_json(payload)
+
+
+async def _send_error(
+    websocket: WebSocket,
+    code: str,
+    message: str,
+    *,
+    retryable: bool,
+) -> None:
+    await _send_json(
+        websocket,
+        ErrorMessage(code=code, message=message, retryable=retryable),
+    )
+
+
+def _cleanup_session(session: Session, state: ServerState) -> None:
+    state.sessions.close(session.id)
+    state.session_store.drop(session.id)
+    init_image_path = session.metadata.get("session_init_image")
+    if isinstance(init_image_path, str):
+        with contextlib.suppress(FileNotFoundError):
+            os.unlink(init_image_path)
+
+
+__all__ = [
+    "ServerState",
+    "build_app",
+    "run_server",
+]

fastvideo/entrypoints/streaming/session.py

@@ -0,0 +1,214 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Per-connection session lifecycle for the streaming server.
+
+Each WebSocket opens exactly one :class:`Session`. :class:`SessionManager`
+enforces the ``generation_segment_cap`` and ``session_timeout_seconds``
+budgets from :class:`fastvideo.api.StreamingConfig`.
+"""
+from __future__ import annotations
+
+import enum
+import time
+import uuid
+from dataclasses import dataclass, field
+from typing import Any
+
+from fastvideo.api.schema import ContinuationState
+
+
+class SessionState(enum.Enum):
+    """State-machine positions for a streaming session.
+
+    Transitions are server-owned. See
+    ``docs/design/server_contracts/streaming.md`` for the full diagram.
+    """
+
+    INITIALIZING = "initializing"
+    QUEUED = "queued"
+    GPU_BINDING = "gpu_binding"
+    ACTIVE = "active"
+    COMPLETE = "complete"
+    ERROR = "error"
+    TIMEOUT = "timeout"
+    REJECTED = "rejected"
+
+
+_VALID_TRANSITIONS: dict[SessionState, frozenset[SessionState]] = {
+    SessionState.INITIALIZING:
+    frozenset({
+        SessionState.QUEUED,
+        SessionState.GPU_BINDING,
+        SessionState.REJECTED,
+        SessionState.ERROR,
+    }),
+    SessionState.QUEUED:
+    frozenset({
+        SessionState.GPU_BINDING,
+        SessionState.ERROR,
+        SessionState.TIMEOUT,
+        SessionState.REJECTED,
+    }),
+    SessionState.GPU_BINDING:
+    frozenset({
+        SessionState.ACTIVE,
+        SessionState.ERROR,
+        SessionState.TIMEOUT,
+    }),
+    SessionState.ACTIVE:
+    frozenset({
+        SessionState.ACTIVE,
+        SessionState.COMPLETE,
+        SessionState.ERROR,
+        SessionState.TIMEOUT,
+    }),
+    SessionState.COMPLETE:
+    frozenset(),
+    SessionState.ERROR:
+    frozenset(),
+    SessionState.TIMEOUT:
+    frozenset(),
+    SessionState.REJECTED:
+    frozenset(),
+}
+
+
+class InvalidSessionTransition(RuntimeError):
+    """Raised when a session is asked to transition along an illegal edge."""
+
+
+@dataclass
+class Session:
+    id: str = field(default_factory=lambda: uuid.uuid4().hex)
+    state: SessionState = SessionState.INITIALIZING
+    created_at: float = field(default_factory=time.monotonic)
+    last_activity: float = field(default_factory=time.monotonic)
+
+    client_id: str | None = None
+    preset: str | None = None
+    preset_label: str | None = None
+
+    curated_prompts: list[str] = field(default_factory=list)
+
+    segment_idx: int = 0
+
+    enhancement_enabled: bool = False
+    auto_extension_enabled: bool = False
+    loop_generation_enabled: bool = False
+    single_clip_mode: bool = False
+    generation_paused: bool = False
+
+    stream_mode: str = "av_fmp4"
+    gpu_id: int | None = None
+
+    continuation_state: ContinuationState | None = None
+
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+    def transition(self, target: SessionState) -> None:
+        """Move to ``target`` if the edge is allowed.
+
+        Raises :class:`InvalidSessionTransition` on illegal moves. The
+        self-loop on ``ACTIVE`` is legal so the server can re-assert
+        ACTIVE on segment completion without special casing.
+        """
+        allowed = _VALID_TRANSITIONS.get(self.state, frozenset())
+        if target not in allowed and target is not self.state:
+            raise InvalidSessionTransition(f"{self.state.value} -> {target.value} is not a valid "
+                                           f"session transition")
+        self.state = target
+        self.last_activity = time.monotonic()
+
+    def touch(self) -> None:
+        self.last_activity = time.monotonic()
+
+    def is_active(self) -> bool:
+        return self.state is SessionState.ACTIVE
+
+    def segment_cap_reached(self, cap: int) -> bool:
+        return self.segment_idx >= cap
+
+
+class SessionManager:
+    """Registers sessions and enforces per-server session limits."""
+
+    def __init__(
+        self,
+        *,
+        segment_cap: int,
+        session_timeout_seconds: int,
+        max_sessions: int = 1,
+    ) -> None:
+        self._segment_cap = segment_cap
+        self._session_timeout_seconds = session_timeout_seconds
+        self._max_sessions = max_sessions
+        self._sessions: dict[str, Session] = {}
+
+    @property
+    def segment_cap(self) -> int:
+        return self._segment_cap
+
+    @property
+    def session_timeout_seconds(self) -> int:
+        return self._session_timeout_seconds
+
+    def create(self) -> Session:
+        if len(self._sessions) >= self._max_sessions:
+            raise SessionRejected(f"max sessions reached ({self._max_sessions})")
+        session = Session()
+        self._sessions[session.id] = session
+        return session
+
+    def get(self, session_id: str) -> Session | None:
+        return self._sessions.get(session_id)
+
+    def close(self, session_id: str) -> None:
+        self._sessions.pop(session_id, None)
+
+    def __contains__(self, session_id: str) -> bool:
+        return session_id in self._sessions
+
+    def __len__(self) -> int:
+        return len(self._sessions)
+
+    def active_sessions(self) -> list[Session]:
+        return [s for s in self._sessions.values() if s.is_active()]
+
+    def reap_timed_out(self, now: float | None = None) -> list[str]:
+        """Return the ids of sessions that have exceeded the idle timeout.
+
+        The caller is responsible for actually closing them — this
+        method only *identifies* dead sessions so the server can emit
+        ``session_timeout`` frames before dropping the WebSocket.
+
+        TODO: unused until a background driver calls it. Per-connection
+        idle enforcement currently happens via asyncio.wait_for on
+        receive_json; this helper catches sessions stuck before any
+        receive (e.g. future QUEUED state) and is expected to be wired
+        into the GPU-pool reaper.
+        """
+        now = now if now is not None else time.monotonic()
+        dead: list[str] = []
+        for sid, session in self._sessions.items():
+            if session.state in {
+                    SessionState.COMPLETE,
+                    SessionState.ERROR,
+                    SessionState.TIMEOUT,
+                    SessionState.REJECTED,
+            }:
+                continue
+            if now - session.last_activity > self._session_timeout_seconds:
+                dead.append(sid)
+        return dead
+
+
+class SessionRejected(RuntimeError):
+    """Raised when session creation fails (queue full, auth, etc.)."""
+
+
+__all__ = [
+    "InvalidSessionTransition",
+    "Session",
+    "SessionManager",
+    "SessionRejected",
+    "SessionState",
+]

fastvideo/entrypoints/streaming/session_init_image.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Persist the initial-image blob attached to a streaming session."""
+from __future__ import annotations
+
+import base64
+import binascii
+import contextlib
+import os
+import tempfile
+from dataclasses import dataclass
+from typing import Any
+
+_ACCEPTED_MIMES = {
+    "image/png": ".png",
+    "image/jpeg": ".jpg",
+    "image/jpg": ".jpg",
+    "image/webp": ".webp",
+}
+
+_MAX_IMAGE_BYTES = 32 * 1024 * 1024  # 32 MiB cap
+
+
+@dataclass(frozen=True)
+class SessionInitImage:
+    """Location of the persisted init image.
+
+    Callers pass ``path`` to ``InputConfig.image_path``; ``display_name``
+    is only used for logs.
+    """
+
+    path: str
+    display_name: str
+    mime: str
+
+
+def persist_session_init_image(
+    payload: Any,
+    *,
+    output_dir: str | None = None,
+) -> SessionInitImage | None:
+    """Decode a client init-image blob and persist it to disk.
+
+    ``payload`` shape (matches the internal UI protocol)::
+
+        {
+            "mime": "image/png",
+            "name": "ref.png",
+            "data": "<base64 bytes>",
+        }
+
+    Returns ``None`` when ``payload`` is falsy (no init image). Raises
+    :class:`ValueError` on schema / size / decode errors so the caller
+    can surface a user-facing ``error`` frame.
+    """
+    if not payload:
+        return None
+    if not isinstance(payload, dict):
+        raise ValueError("session init image must be an object")
+
+    mime = payload.get("mime")
+    if mime not in _ACCEPTED_MIMES:
+        raise ValueError(f"session init image mime {mime!r} is not one of "
+                         f"{sorted(_ACCEPTED_MIMES)}")
+    data_b64 = payload.get("data")
+    if not isinstance(data_b64, str):
+        raise ValueError("session init image data must be a base64 string")
+    try:
+        data = base64.b64decode(data_b64, validate=True)
+    except (binascii.Error, ValueError) as exc:
+        raise ValueError(f"session init image data is not valid base64: {exc}") from exc
+    if len(data) > _MAX_IMAGE_BYTES:
+        raise ValueError(f"session init image is {len(data)} bytes; limit is "
+                         f"{_MAX_IMAGE_BYTES}")
+    if len(data) == 0:
+        raise ValueError("session init image data is empty")
+
+    ext = _ACCEPTED_MIMES[mime]
+    display_name = _sanitize_display_name(payload.get("name")) or f"init{ext}"
+    fd, path = tempfile.mkstemp(prefix="fastvideo-init-", suffix=ext, dir=output_dir)
+    try:
+        with os.fdopen(fd, "wb") as f:
+            f.write(data)
+    except Exception:
+        with contextlib.suppress(FileNotFoundError):
+            os.unlink(path)
+        raise
+    return SessionInitImage(path=path, display_name=display_name, mime=mime)
+
+
+def _sanitize_display_name(name: Any) -> str | None:
+    if not isinstance(name, str):
+        return None
+    name = name.strip()
+    if not name:
+        return None
+    # Strip any path components — we only keep the leaf for logging.
+    return os.path.basename(name)
+
+
+__all__ = [
+    "SessionInitImage",
+    "persist_session_init_image",
+]

fastvideo/entrypoints/streaming/session_store.py

@@ -0,0 +1,206 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Session state store for the FastVideo streaming server.
+
+The streaming server keeps continuation state (decoded frames + audio
+latents from the previous segment) server-side so the client doesn't
+re-upload multi-megabyte tensors each WebSocket message. Two operations
+are needed:
+
+* ``snapshot(session_id) -> ContinuationState`` — serialize the current
+  state so it can be exported (e.g. over HTTP) or migrated to a
+  different server.
+* ``hydrate(state) -> session_id`` — load a previously serialized state
+  into a new session (for resume-after-disconnect flows).
+
+The store is an ABC with an :class:`InMemorySessionStore` default; Redis
+or other backends can drop in without touching the pipeline.
+
+Large tensor payloads (video frames, audio latents) are kept out of the
+JSON payload via an accompanying :class:`BlobStore`. Both stores share a
+process today; they are separate types so that a future implementation
+can put blobs on S3 while keeping session metadata in Redis.
+"""
+from __future__ import annotations
+
+import threading
+import uuid
+from abc import ABC, abstractmethod
+from collections.abc import Iterator
+from dataclasses import dataclass
+
+from fastvideo.api.schema import ContinuationState
+
+
+class BlobStore(ABC):
+    """Opaque byte-blob storage keyed by id.
+
+    A :class:`ContinuationState` payload can reference large tensors
+    stored in a :class:`BlobStore` rather than inlining them, so the
+    JSON payload stays small when the state travels over the wire.
+    """
+
+    @abstractmethod
+    def put(self, data: bytes, *, mime: str = "application/octet-stream") -> str:
+        """Store ``data`` and return a blob id for later retrieval."""
+
+    @abstractmethod
+    def get(self, blob_id: str) -> bytes:
+        """Load a previously stored blob. Raises ``KeyError`` if absent."""
+
+    @abstractmethod
+    def drop(self, blob_id: str) -> None:
+        """Remove a blob. Missing ids are a no-op."""
+
+    @abstractmethod
+    def __contains__(self, blob_id: str) -> bool:
+        ...
+
+
+@dataclass(frozen=True)
+class _BlobRecord:
+    data: bytes
+    mime: str
+
+
+class InMemoryBlobStore(BlobStore):
+    """Thread-safe in-memory :class:`BlobStore` for single-process servers.
+
+    No eviction policy — callers are responsible for calling
+    :meth:`drop` when a blob's owning state is replaced or a session
+    ends. A redis- or filesystem-backed :class:`BlobStore` should
+    replace this when the streaming server lands as a real service
+    (PR 7.5+).
+    """
+
+    def __init__(self) -> None:
+        self._blobs: dict[str, _BlobRecord] = {}
+        self._lock = threading.Lock()
+
+    def put(self, data: bytes, *, mime: str = "application/octet-stream") -> str:
+        blob_id = uuid.uuid4().hex
+        with self._lock:
+            self._blobs[blob_id] = _BlobRecord(data=data, mime=mime)
+        return blob_id
+
+    def get(self, blob_id: str) -> bytes:
+        with self._lock:
+            record = self._blobs.get(blob_id)
+        if record is None:
+            raise KeyError(f"Unknown blob id: {blob_id}")
+        return record.data
+
+    def drop(self, blob_id: str) -> None:
+        with self._lock:
+            self._blobs.pop(blob_id, None)
+
+    def __contains__(self, blob_id: str) -> bool:
+        with self._lock:
+            return blob_id in self._blobs
+
+    def __len__(self) -> int:
+        with self._lock:
+            return len(self._blobs)
+
+
+class SessionStore(ABC):
+    """Keyed store for per-session continuation state.
+
+    Implementations own the session-id → state mapping. The streaming
+    server calls :meth:`store` after each segment and :meth:`snapshot`
+    when a client explicitly asks for an exportable state handle.
+    """
+
+    @abstractmethod
+    def store(self, session_id: str, state: ContinuationState) -> None:
+        """Persist ``state`` for ``session_id``, replacing any prior value."""
+
+    @abstractmethod
+    def snapshot(self, session_id: str) -> ContinuationState | None:
+        """Return the current state for ``session_id`` (or ``None``)."""
+
+    @abstractmethod
+    def hydrate(
+        self,
+        state: ContinuationState,
+        *,
+        session_id: str | None = None,
+    ) -> str:
+        """Install ``state`` as the starting point for a session.
+
+        When ``session_id`` is ``None`` the store allocates a fresh id
+        (UUID4); when provided the store uses it verbatim, overwriting
+        any prior state at that id.
+        """
+
+    @abstractmethod
+    def drop(self, session_id: str) -> None:
+        """Forget a session. Missing ids are a no-op."""
+
+    @abstractmethod
+    def __contains__(self, session_id: str) -> bool:
+        ...
+
+    @abstractmethod
+    def __iter__(self) -> Iterator[str]:
+        ...
+
+
+class InMemorySessionStore(SessionStore):
+    """Thread-safe in-memory :class:`SessionStore`.
+
+    Default implementation used by single-process deployments; a future
+    Redis-backed store can be dropped in without changes to the server.
+
+    No eviction / TTL / bounded capacity — sessions only leave via
+    :meth:`drop`. The live streaming server (PR 7.5+) is responsible
+    for bounding growth and for dropping any :class:`BlobStore` blobs
+    referenced by a state when that state is replaced or a session
+    ends; this class does not know about blobs.
+    """
+
+    def __init__(self) -> None:
+        self._sessions: dict[str, ContinuationState] = {}
+        self._lock = threading.Lock()
+
+    def store(self, session_id: str, state: ContinuationState) -> None:
+        with self._lock:
+            self._sessions[session_id] = state
+
+    def snapshot(self, session_id: str) -> ContinuationState | None:
+        with self._lock:
+            return self._sessions.get(session_id)
+
+    def hydrate(
+        self,
+        state: ContinuationState,
+        *,
+        session_id: str | None = None,
+    ) -> str:
+        sid = session_id or uuid.uuid4().hex
+        with self._lock:
+            self._sessions[sid] = state
+        return sid
+
+    def drop(self, session_id: str) -> None:
+        with self._lock:
+            self._sessions.pop(session_id, None)
+
+    def __contains__(self, session_id: str) -> bool:
+        with self._lock:
+            return session_id in self._sessions
+
+    def __iter__(self) -> Iterator[str]:
+        with self._lock:
+            return iter(list(self._sessions))
+
+    def __len__(self) -> int:
+        with self._lock:
+            return len(self._sessions)
+
+
+__all__ = [
+    "BlobStore",
+    "InMemoryBlobStore",
+    "InMemorySessionStore",
+    "SessionStore",
+]

fastvideo/entrypoints/streaming/stream.py

@@ -0,0 +1,213 @@
+# SPDX-License-Identifier: Apache-2.0
+"""fMP4 stream encoder used by the streaming server.
+
+The client's Media Source Extensions player needs a continuous fMP4
+byte stream: first an *initialization segment* (``ftyp`` + ``moov``),
+then one or more *media segments* (``moof`` + ``mdat``). We pipe raw
+RGB frames into an ffmpeg subprocess configured for fragmented output
+via ``-movflags empty_moov+default_base_moof+frag_keyframe+faststart``
+and stream the bytes back out.
+"""
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import subprocess
+import uuid
+from collections.abc import AsyncIterator
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Literal
+
+if TYPE_CHECKING:
+    import numpy as np
+
+
+@dataclass
+class FragmentedMP4Chunk:
+    """A single fMP4 byte chunk emitted by :class:`FragmentedMP4Encoder`.
+
+    ``kind`` identifies whether the chunk is the init segment (must be
+    fed into the client's ``SourceBuffer`` first) or a media fragment.
+    """
+
+    kind: Literal["init", "media"]
+    data: bytes
+    stream_id: str
+    segment_idx: int
+
+
+class FragmentedMP4Encoder:
+    """Stream RGB frames in, fMP4 chunks out.
+
+    One encoder covers one segment. The server creates a new encoder
+    per :class:`ltx2_segment_start`` boundary so each segment becomes
+    one media fragment the client can append independently.
+
+    Example::
+
+        encoder = FragmentedMP4Encoder(width=1024, height=576, fps=24,
+                                        segment_idx=0)
+        async with encoder:
+            async for chunk in encoder.encode(frames):
+                await websocket.send_bytes(chunk.data)
+    """
+
+    def __init__(
+        self,
+        *,
+        width: int,
+        height: int,
+        fps: int,
+        segment_idx: int,
+        stream_id: str | None = None,
+        ffmpeg_path: str = "ffmpeg",
+        preset: str = "ultrafast",
+        pixel_format_out: str = "yuv420p",
+        extra_args: list[str] | None = None,
+    ) -> None:
+        self.width = width
+        self.height = height
+        self.fps = fps
+        self.segment_idx = segment_idx
+        self.stream_id = stream_id or uuid.uuid4().hex
+        self._ffmpeg_path = ffmpeg_path
+        self._preset = preset
+        self._pixel_format_out = pixel_format_out
+        self._extra_args = list(extra_args or [])
+        self._proc: subprocess.Popen | None = None
+        self._init_emitted = False
+
+    async def __aenter__(self) -> FragmentedMP4Encoder:
+        self._spawn()
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb) -> None:
+        await self.close()
+
+    def _spawn(self) -> None:
+        args = [
+            self._ffmpeg_path,
+            "-hide_banner",
+            "-loglevel",
+            "error",
+            "-f",
+            "rawvideo",
+            "-pix_fmt",
+            "rgb24",
+            "-s",
+            f"{self.width}x{self.height}",
+            "-r",
+            str(self.fps),
+            "-i",
+            "-",
+            "-c:v",
+            "libx264",
+            "-preset",
+            self._preset,
+            "-tune",
+            "zerolatency",
+            "-pix_fmt",
+            self._pixel_format_out,
+            "-movflags",
+            "empty_moov+default_base_moof+frag_keyframe+faststart",
+            "-f",
+            "mp4",
+            *self._extra_args,
+            "-",
+        ]
+        # stderr → DEVNULL: with -loglevel error on, the only thing
+        # stderr would carry is unsolicited warnings. Piping without a
+        # reader deadlocks ffmpeg once the pipe buffer (~64 KiB) fills.
+        self._proc = subprocess.Popen(  # noqa: S603
+            args,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.DEVNULL,
+            bufsize=0,
+        )
+
+    async def encode(
+        self,
+        frames: list[np.ndarray] | AsyncIterator[np.ndarray],
+    ) -> AsyncIterator[FragmentedMP4Chunk]:
+        """Feed frames into ffmpeg and yield fMP4 chunks as they appear."""
+        if self._proc is None:
+            self._spawn()
+        assert self._proc is not None and self._proc.stdin is not None
+        proc = self._proc
+
+        loop = asyncio.get_running_loop()
+
+        async def _writer() -> None:
+            try:
+                if hasattr(frames, "__aiter__"):
+                    async for frame in frames:  # type: ignore[union-attr]
+                        await loop.run_in_executor(None, _write_frame, proc.stdin, frame)
+                else:
+                    for frame in frames:  # type: ignore[assignment]
+                        await loop.run_in_executor(None, _write_frame, proc.stdin, frame)
+            finally:
+                with contextlib.suppress(BrokenPipeError):
+                    proc.stdin.close()
+
+        writer_task = asyncio.create_task(_writer())
+        try:
+            reader = proc.stdout
+            assert reader is not None
+            # Read in reasonably-sized chunks; MSE tolerates any size
+            # but we don't want to starve the event loop.
+            chunk_size = 64 * 1024
+            while True:
+                data = await loop.run_in_executor(None, reader.read, chunk_size)
+                if not data:
+                    break
+                kind: Literal["init", "media"] = "init" if not self._init_emitted else "media"
+                self._init_emitted = True
+                yield FragmentedMP4Chunk(
+                    kind=kind,
+                    data=bytes(data),
+                    stream_id=self.stream_id,
+                    segment_idx=self.segment_idx,
+                )
+        finally:
+            await writer_task
+
+    async def close(self) -> None:
+        if self._proc is None:
+            return
+        proc = self._proc
+        self._proc = None
+        try:
+            if proc.stdin and not proc.stdin.closed:
+                proc.stdin.close()
+        except BrokenPipeError:
+            pass
+        loop = asyncio.get_running_loop()
+        try:
+            await asyncio.wait_for(
+                loop.run_in_executor(None, proc.wait),
+                timeout=5.0,
+            )
+        except asyncio.TimeoutError:
+            proc.kill()
+            await loop.run_in_executor(None, proc.wait)
+
+
+def _write_frame(stdin, frame: np.ndarray) -> None:
+    import numpy as np
+
+    if not isinstance(frame, np.ndarray):
+        raise TypeError("fMP4 encoder frames must be numpy.ndarray")
+    if frame.dtype != np.uint8:
+        frame = frame.astype(np.uint8)
+    if frame.ndim != 3 or frame.shape[-1] != 3:
+        raise ValueError("fMP4 encoder frames must be HxWx3 uint8 RGB; got "
+                         f"shape={frame.shape}, dtype={frame.dtype}")
+    with contextlib.suppress(BrokenPipeError):
+        stdin.write(frame.tobytes())
+
+
+__all__ = [
+    "FragmentedMP4Chunk",
+    "FragmentedMP4Encoder",
+]

fastvideo/eval/README.md

@@ -0,0 +1,214 @@
+# `fastvideo.eval`
+
+In-process evaluation suite for video generations. Includes pixel
+metrics (SSIM, PSNR, LPIPS), optical-flow comparisons, the full VBench
+suite, Physics-IQ, and a VLM scorer (VideoScore-2) behind a single
+registry-driven API.
+
+## Install
+
+| Use case | Install |
+|---|---|
+| Default (common, optical_flow, vbench-light, physics_iq, videoscore2) | `uv pip install -e .[eval]` |
+| Just VBench (12 of 16 sub-metrics) | `uv pip install -e .[eval-vbench]` |
+| Just Physics-IQ (covered by `[eval]`) | `uv pip install -e .[eval-physics-iq]` |
+| Plus `vbench.scene` (AVoCaDO) | `uv pip install -e .[eval-full]` |
+| Plus `vbench.{color, multiple_objects, object_class, spatial_relationship}` (GRiT) | `uv pip install -e .[eval-vbench]` then `uv pip install --no-build-isolation 'git+https://github.com/facebookresearch/detectron2.git'` |
+
+To use VBench, also pull the upstream submodule:
+
+```bash
+git submodule update --init --recursive  # fetches vbench + kernel deps
+```
+
+The submodule is a clean upstream pin. Compat with current
+transformers/numpy/timm versions is applied at import time in
+`fastvideo/eval/metrics/vbench/__init__.py` via attribute-level
+monkey-patches; the submodule files are unchanged.
+
+## Public API
+
+```python
+from fastvideo.eval import (
+    create_evaluator,        # build a reusable Evaluator
+    evaluate,                # one-shot helper
+    Evaluator,               # the class itself
+    BaseMetric, MetricResult,
+    register, list_metrics, get_metric,
+    ensure_checkpoint, get_cache_dir,
+)
+
+ev = create_evaluator(metrics=["common.ssim", "vbench.aesthetic_quality"],
+                      device="cuda")
+scores = ev.evaluate(video=tensor, reference=ref, fps=8.0)
+```
+
+`evaluate` accepts either a pre-loaded `(T, C, H, W)` tensor or a path
+string for `video` and `reference`. Paths are decoded inside the worker
+that picks up the sample, so peak memory stays bounded by `num_gpus`
+when scoring large batches.
+
+### CLI
+
+```bash
+fastvideo eval list                              # list registered metrics
+fastvideo eval list --group vbench               # filter by group
+fastvideo eval run --videos clip.mp4 \
+    --metrics vbench.aesthetic_quality \
+    --output scores.json
+fastvideo eval run --videos generated/*.mp4 \
+    --reference reference/ \
+    --metrics common.ssim,common.lpips
+```
+
+### Generate-then-score example
+
+`examples/inference/eval/eval_ltx2_vbench.py` runs an LTX2 prompt
+through `VideoGenerator` and scores the resulting mp4 with
+`vbench.aesthetic_quality` and `vbench.subject_consistency`. Use it as
+a template for end-to-end "generate then score" pipelines.
+
+## Layout
+
+```
+fastvideo/
+├── eval/
+│   ├── api.py, evaluator.py, registry.py, models.py, ...
+│   ├── io/                        # video loading helpers
+│   ├── datasets/                  # prompt corpora (vbench, physics_iq)
+│   └── metrics/
+│       ├── base.py                # BaseMetric + @register contract
+│       ├── common/                # SSIM, PSNR, LPIPS
+│       ├── optical_flow/          # gt_optical_flow, synthetic_optical_flow
+│       ├── videoscore2/           # VideoScore-2 (Qwen2.5-VL)
+│       ├── physics_iq/            # PhysicsIQ + sub-metrics
+│       └── vbench/                # adapter: sys.path bootstrap + shims
+│           ├── __init__.py
+│           └── <16 sub-metric pkgs>
+└── third_party/
+    └── eval/
+        └── vbench/                # git submodule (Vchitect/VBench)
+```
+
+### Prompt datasets
+
+```python
+from fastvideo.eval.datasets import get_dataset, list_datasets
+
+list_datasets()                    # ['physics_iq', 'vbench']
+
+ds = get_dataset("physics_iq", limit=4)   # auto-fetches assets on first miss
+for row in ds:
+    # row contains 'prompt', 'reference', 'reference_take2', and
+    # metric-specific aux fields. Drop straight into Evaluator.evaluate(**row).
+    ...
+```
+
+The Physics-IQ manifest CSV is vendored at
+`fastvideo/eval/metrics/physics_iq/_vendored/descriptions.csv`.
+Per-scenario videos, masks, and switch-frames auto-fetch on first use
+into `${FASTVIDEO_EVAL_CACHE}/datasets/physics_iq/`. For air-gapped
+runs, pass `auto_download=False` or `dataset_root=` a pre-downloaded
+copy. Set `FASTVIDEO_PHYSICS_IQ_BUCKET_URL` to redirect the fetch to
+an internal mirror.
+
+## Adding a new metric
+
+The full porting guide is at
+[`docs/contributing/eval-metrics.md`](../../docs/contributing/eval-metrics.md).
+Summary below.
+
+### Native metric (no submodule)
+
+```python
+# fastvideo/eval/metrics/common/<your_metric>/metric.py
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("common.your_metric")
+class YourMetric(BaseMetric):
+    name = "common.your_metric"
+    requires_reference = True
+    needs_gpu = False
+    dependencies: list[str] = []  # e.g. ["pyiqa"] if relevant
+
+    def compute(self, sample) -> list[MetricResult]:
+        ...
+```
+
+The metric is auto-discovered by `fastvideo/eval/metrics/__init__.py`,
+which walks all non-underscore subdirectories and imports their
+`metric` module.
+
+### Wrapping upstream code via a submodule
+
+See `fastvideo/eval/metrics/vbench/` for a worked example. The
+contract is:
+
+1. Upstream lives as a git submodule under
+   `fastvideo/third_party/eval/<bench>/`, pinned to a SHA in repo-root
+   `.gitmodules`.
+2. The metric package's `__init__.py`
+   (`fastvideo/eval/metrics/<bench>/__init__.py`) inserts that
+   submodule path on `sys.path` and installs any compat shims for
+   modern torch/transformers/numpy. Do not modify upstream files on
+   disk.
+3. Per-sub-metric `metric.py` files use `@register("<bench>.<name>")`.
+
+Patches live as Python in the metric's `__init__.py` so they are
+grep-able and reviewable.
+
+## Caches
+
+Eval cache root: `${FASTVIDEO_CACHE_ROOT}/eval/`, default
+`~/.cache/fastvideo/eval/`. Override with `FASTVIDEO_EVAL_CACHE`.
+
+```
+${FASTVIDEO_CACHE_ROOT}/eval/
+├── models/      # URL-fetched checkpoints (LAION head, AMT, GRiT)
+├── torch/       # redirected TORCH_HOME (DINO via torch.hub, lpips)
+├── clip/        # passed as download_root= to clip.load callsites
+└── datasets/    # auto-fetched dataset assets, one subdir per benchmark
+                 # (e.g. datasets/physics_iq/{split-videos,switch-frames,...})
+```
+
+HF-hosted models stay in HF's default cache
+(`~/.cache/huggingface/hub/`) so they dedupe with other ML projects on
+the same host.
+
+### Convention for new metrics
+
+If your metric wraps a third-party loader that has its own cache
+directory, route it through `get_cache_dir()` so users get one knob
+to redirect everything.
+
+```python
+# CLIP: pass download_root explicitly
+import clip
+from fastvideo.eval.models import get_cache_dir
+model, _ = clip.load("ViT-B/32", device=device,
+                     download_root=str(get_cache_dir() / "clip"))
+
+# torch.hub is already redirected by fastvideo.eval.__init__ via
+# TORCH_HOME; no per-callsite work needed.
+
+# transformers / huggingface_hub: leave alone. HF's default cache is
+# shared with other tools.
+```
+
+For libraries that do not honour any env var or kwarg (pyiqa, funasr),
+their cache lands in the library's own dir. Document the exception in
+the metric's docstring if it matters.
+
+## Out of scope (follow-up PRs)
+
+- **MIND** metrics. Depend on a separate `vipe` upstream submodule.
+- **VBench-2.0**. Sibling vbench2 package; needs its own port.
+- **FVD as a registered metric**. Currently still at `benchmarks/fvd/`.
+  FVD is a set-vs-set distribution distance and does not fit the
+  per-sample `BaseMetric.compute` API without a stateful accumulator;
+  conversion is a designed follow-up.
+- **Training-time eval callback** (`EvalCallback`) and the
+  `RolloutEvaluator` helper.

fastvideo/eval/__init__.py

@@ -0,0 +1,45 @@
+from fastvideo.eval.models import ensure_checkpoint, get_cache_dir
+
+
+def _redirect_third_party_caches() -> None:
+    """Point libraries that respect env vars at the eval cache root.
+
+    Run before metric modules import torch.hub (and friends) so the
+    redirect actually takes effect. We intentionally leave ``HF_HOME``
+    alone — HF's default cache (``~/.cache/huggingface/hub``) is widely
+    shared with other ML projects, and isolating it for eval would force
+    users to re-download already-cached transformers weights.
+
+    Other libraries with non-standard caches (CLIP, pyiqa) don't honour
+    env vars at all; their callsites in metric.py files pass
+    ``download_root=str(get_cache_dir() / "<library>")`` directly.
+    """
+    import os
+    root = get_cache_dir()
+    os.environ.setdefault("TORCH_HOME", str(root / "torch"))
+
+
+_redirect_third_party_caches()
+
+from fastvideo.eval.types import MetricResult, Video  # noqa: E402
+from fastvideo.eval.metrics.base import BaseMetric  # noqa: E402
+from fastvideo.eval.registry import register, list_metrics, get_metric  # noqa: E402
+from fastvideo.eval.api import evaluate  # noqa: E402
+from fastvideo.eval.evaluator import Evaluator, create_evaluator  # noqa: E402
+
+# Trigger metric auto-discovery
+import fastvideo.eval.metrics  # noqa: F401, E402
+
+__all__ = [
+    "evaluate",
+    "Evaluator",
+    "create_evaluator",
+    "MetricResult",
+    "Video",
+    "BaseMetric",
+    "register",
+    "list_metrics",
+    "get_metric",
+    "ensure_checkpoint",
+    "get_cache_dir",
+]

fastvideo/eval/api.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import torch
+
+from fastvideo.eval.evaluator import create_evaluator
+from fastvideo.eval.types import MetricResult
+
+
+def evaluate(
+    generated: torch.Tensor | str | Path,
+    reference: torch.Tensor | str | Path | None = None,
+    metrics: list[str] | str = "all",
+    device: str = "cuda",
+    **kwargs,
+) -> dict[str, MetricResult] | list[dict[str, MetricResult]]:
+    """One-shot evaluation. For repeated use, prefer :func:`create_evaluator`.
+
+    Parameters
+    ----------
+    generated : Tensor | str | Path
+        Generated video. Either a pre-loaded ``(T, C, H, W)`` tensor or a
+        path to an mp4/avi/etc. — paths are decoded by the worker.
+    reference : Tensor | str | Path | None
+        Reference video (same accepted shapes as *generated*).
+    metrics : list[str] | str
+        Metric names, or ``"all"``.
+    device : str
+        PyTorch device string.
+    """
+    ev = create_evaluator(metrics=metrics, device=device)
+    kw: dict = {"video": generated, **kwargs}
+    if reference is not None:
+        kw["reference"] = reference
+    return ev.evaluate(**kw)

fastvideo/eval/datasets/__init__.py

@@ -0,0 +1,51 @@
+"""Prompt-corpus datasets for end-to-end benchmark evaluation.
+
+Public API mirrors :mod:`fastvideo.eval` (metrics side):
+
+    from fastvideo.eval.datasets import (
+        PromptDataset, Sample,
+        register_dataset, get_dataset, list_datasets,
+    )
+
+A dataset is an iterable of plain dicts (one per sample). Built-in
+datasets self-register at import time. To add one, drop a module into
+this package that subclasses :class:`PromptDataset` and decorates with
+``@register_dataset("name")`` — auto-discovery picks it up.
+"""
+from fastvideo.eval.datasets.base import (BasePromptDataset, PromptDataset, Sample)
+from fastvideo.eval.datasets.registry import (get_dataset, list_datasets, register_dataset)
+
+
+def _autodiscover() -> None:
+    """Import every non-underscore .py module / subpackage in this package
+    so the ``@register_dataset`` decorators fire."""
+    import importlib
+    import os
+
+    for entry in os.listdir(os.path.dirname(__file__)):
+        if entry.startswith("_") or entry.startswith("."):
+            continue
+        if entry in {"base.py", "registry.py"}:
+            continue
+        if entry.endswith(".py"):
+            importlib.import_module(f"{__name__}.{entry[:-3]}")
+        elif os.path.isdir(os.path.join(os.path.dirname(__file__), entry)) \
+                and os.path.exists(os.path.join(
+                    os.path.dirname(__file__), entry, "__init__.py")):
+            importlib.import_module(f"{__name__}.{entry}")
+
+
+_autodiscover()
+
+# Re-export the canonical class for typed imports.
+from fastvideo.eval.datasets.vbench import VBenchPromptDataset  # noqa: E402
+
+__all__ = [
+    "PromptDataset",
+    "BasePromptDataset",
+    "Sample",
+    "register_dataset",
+    "get_dataset",
+    "list_datasets",
+    "VBenchPromptDataset",
+]

fastvideo/eval/datasets/base.py

@@ -0,0 +1,81 @@
+"""Prompt-corpus datasets.
+
+A :class:`PromptDataset` is an iterable of *sample dicts* describing the
+prompts and conditions for a benchmark. Each sample is a plain dict —
+no dataclass, no schema enforcement — that flows directly into both
+generation (``VideoGenerator.generate_video(**sample)``) and scoring
+(``Evaluator.evaluate(**eval_kwargs)``). The runner picks well-known
+keys (``prompt``, ``n_samples``, ``dimensions``, ``auxiliary_info``,
+...) and passes the rest through.
+
+This matches the surrounding FastVideo style:
+
+* :class:`fastvideo.dataset.validation_dataset.ValidationDataset` yields dicts.
+* :meth:`fastvideo.VideoGenerator.generate_video` consumes ``**kwargs``.
+* :meth:`fastvideo.eval.Evaluator.evaluate` consumes ``**kwargs``.
+
+To add a new benchmark:
+
+1. Subclass :class:`PromptDataset`, populate ``self._rows`` with dicts in
+   ``__init__``.
+2. Decorate with ``@register_dataset("my_bench")``.
+
+Convention for ``auxiliary_info``: a *flat* dict of metric-keyed values
+(e.g. ``{"color": "red"}``). Benchmarks with nested aux schemas (VBench's
+``{dim: {key: val}}``) flatten at load time so every consumer sees the
+same shape.
+"""
+from __future__ import annotations
+
+from typing import TypedDict
+from collections.abc import Iterator
+
+
+class Sample(TypedDict, total=False):
+    """Documented schema for a row yielded by :class:`PromptDataset`.
+
+    Only ``prompt`` is required. Extra keys beyond these are forwarded to
+    the runner's eval-kwargs builder verbatim, so action-conditioned or
+    audio-bearing benchmarks can add their own fields without changing
+    the base class.
+    """
+    prompt: str
+    n_samples: int
+    dimensions: list[str]
+    auxiliary_info: dict
+    image_path: str
+    reference_video: str
+
+
+class PromptDataset:
+    """Iterable corpus of sample dicts. Subclasses populate ``self._rows``."""
+
+    name: str = ""
+    description: str = ""
+    supports_dimensions: bool = False
+    requires_reference_image: bool = False
+    requires_reference_video: bool = False
+
+    def __init__(self) -> None:
+        self._rows: list[dict] = []
+
+    def __iter__(self) -> Iterator[dict]:
+        return iter(self._rows)
+
+    def __len__(self) -> int:
+        return len(self._rows)
+
+    def __getitem__(self, i: int) -> dict:
+        return self._rows[i]
+
+    def by_dimension(self) -> dict[str, list[dict]]:
+        """Group samples by dimension. A multi-dim sample appears under each."""
+        out: dict[str, list[dict]] = {}
+        for s in self._rows:
+            for d in s.get("dimensions", ()):
+                out.setdefault(d, []).append(s)
+        return out
+
+
+# Back-compat alias for callers still importing the old class name.
+BasePromptDataset = PromptDataset

fastvideo/eval/datasets/physics_iq.py

@@ -0,0 +1,444 @@
+"""Physics-IQ benchmark prompt corpus.
+
+Yields one sample dict per take-1 scenario, paired with its take-2
+reference and both takes' real motion masks. Each row drops straight
+into :meth:`fastvideo.eval.Evaluator.evaluate` for the ``physics_iq``
+metric:
+
+    {
+        "prompt": <description>,
+        "reference":            "<take-1 mp4>",
+        "reference_take2":      "<take-2 mp4>",
+        "reference_mask":       "<take-1 mask mp4>",
+        "reference_take2_mask": "<take-2 mask mp4>",
+        "scenario": <scenario_id>,
+        "view": <camera view>,
+        "auxiliary_info": { ... metadata ... },
+    }
+
+Self-contained dataset: the manifest CSV is vendored under
+``fastvideo/eval/metrics/physics_iq/_vendored/descriptions.csv``;
+per-scenario videos/masks/switch-frames auto-fetch on first use from the public
+DeepMind bucket into ``${FASTVIDEO_EVAL_CACHE}/datasets/physics_iq/``.
+Pass ``auto_download=False`` (or ``dataset_root=`` pointing at a
+pre-downloaded copy) to opt out of network fetches.
+"""
+
+from __future__ import annotations
+
+import csv
+import os
+from dataclasses import dataclass
+from pathlib import Path
+from urllib.request import urlretrieve
+
+import cv2
+import numpy as np
+
+from fastvideo.eval.datasets.base import PromptDataset
+from fastvideo.eval.datasets.registry import register_dataset
+from fastvideo.eval.models import get_cache_dir
+
+VIEWS = ("perspective-left", "perspective-center", "perspective-right")
+TAKE1_TOKEN = "take-1"
+TAKE2_TOKEN = "take-2"
+
+# FPS the dataset rows should resolve to. Source release is recorded at
+# 30 FPS; if a different value is requested the loader transcodes once
+# into a per-repo-root cache directory.
+_DEFAULT_FPS = 30
+_DEFAULT_DURATION_SECONDS = 5
+
+# Vendored manifest under ``fastvideo/eval/metrics/physics_iq/_vendored/``
+# is the same file shipped by upstream's git repo. The ``_vendored/``
+# subdir is the project-wide convention for upstream-provenance files
+# (matches the ``_``-prefixed auto-discovery skip and a single
+# codespell skip glob).
+_VENDORED_DESCRIPTIONS_CSV = (Path(__file__).resolve().parent.parent / "metrics" / "physics_iq" / "_vendored" /
+                              "descriptions.csv")
+
+# Public DeepMind bucket; HTTPS-readable, no auth. Override via
+# ``FASTVIDEO_PHYSICS_IQ_BUCKET_URL`` (e.g. for an internal mirror).
+_DEFAULT_BUCKET_URL = "https://storage.googleapis.com/physics-iq-benchmark"
+
+
+def _bucket_url() -> str:
+    return os.environ.get("FASTVIDEO_PHYSICS_IQ_BUCKET_URL", _DEFAULT_BUCKET_URL)
+
+
+def _default_dataset_root() -> Path:
+    """Sibling to ``models/torch/clip/`` under the eval cache root."""
+    return get_cache_dir() / "datasets" / "physics_iq"
+
+
+@dataclass(frozen=True)
+class PhysicsIQScenario:
+    """One row of the Physics-IQ manifest, fully resolved on disk."""
+    scenario_id: str
+    view: str
+    scenario_name: str
+    take1_video_path: str
+    take2_video_path: str
+    switch_frame_path: str
+    caption: str
+    expected_gen_filename: str
+    generated_video_path: str | None = None
+    take1_mask_path: str | None = None
+    take2_mask_path: str | None = None
+
+
+@register_dataset("physics_iq")
+class PhysicsIQPromptDataset(PromptDataset):
+    """Physics-IQ benchmark prompt corpus.
+
+    Self-contained: ``get_dataset("physics_iq")`` works with no kwargs.
+    The manifest CSV is vendored next to the metric, and per-scenario
+    assets auto-fetch on first miss from the public bucket into
+    ``${FASTVIDEO_EVAL_CACHE}/datasets/physics_iq/``.
+
+    Args:
+        dataset_root: path to a pre-downloaded copy of the Physics-IQ
+            release. Defaults to ``${FASTVIDEO_EVAL_CACHE}/datasets/physics_iq``;
+            override only if you already have a local mirror.
+        fps: target frame rate. The release ships at 30 FPS; other rates
+            transcode once on first access into ``<root>/.physics_iq_cache/``.
+        limit: optional truncation for quick smoke runs. Apply this kwarg
+            (not a post-construction slice) so we only fetch the assets
+            for the scenarios actually requested.
+        generated_dir: optional directory of pre-generated videos —
+            attaches each manifest row's expected output path to the
+            sample dict under ``auxiliary_info["generated_video_path"]``.
+        auto_download: when True (the default), missing testing videos,
+            masks, and switch frames are fetched from the public bucket
+            into ``dataset_root``. Set False for air-gapped runs; the
+            loader will then raise ``FileNotFoundError`` on miss.
+    """
+
+    description = ("Physics-IQ benchmark, 396 take-1 scenarios across 66 unique physics "
+                   "setups × 3 perspective views, each paired with a take-2 reference.")
+    requires_reference_video = True
+
+    def __init__(
+        self,
+        dataset_root: str | Path | None = None,
+        *,
+        fps: int = _DEFAULT_FPS,
+        limit: int | None = None,
+        generated_dir: str | Path | None = None,
+        auto_download: bool = True,
+    ) -> None:
+        super().__init__()
+
+        repo_root = Path(dataset_root or _default_dataset_root()).expanduser().resolve()
+        self.repo_root = repo_root
+        self.dataset_dir = _resolve_dataset_dir(repo_root)
+        self.descriptions_path = _resolve_descriptions_path(repo_root, self.dataset_dir)
+        self.cache_dir = repo_root / ".physics_iq_cache"
+        self.fps = fps
+        self.auto_download = auto_download
+        self.bucket_url = _bucket_url()
+
+        scenarios = self._iter_scenarios(
+            fps=fps,
+            generated_dir=generated_dir,
+            limit=limit,
+        )
+        self._rows = [_scenario_to_row(s) for s in scenarios]
+
+    def _iter_scenarios(
+        self,
+        *,
+        fps: int,
+        generated_dir: str | Path | None,
+        limit: int | None,
+    ) -> list[PhysicsIQScenario]:
+        with self.descriptions_path.open("r", newline="") as handle:
+            rows = list(csv.DictReader(handle))
+
+        take2_by_suffix = {_scenario_suffix(row["scenario"]): row for row in rows if TAKE2_TOKEN in row["scenario"]}
+        take1_rows = [row for row in rows if TAKE1_TOKEN in row["scenario"]]
+        if limit is not None:
+            take1_rows = take1_rows[:limit]
+
+        generated_dir_path = (Path(generated_dir).expanduser().resolve() if generated_dir else None)
+        scenarios: list[PhysicsIQScenario] = []
+
+        for row in take1_rows:
+            scenario_filename = row["scenario"]
+            scenario_id, view, _, scenario_name = _parse_scenario_filename(scenario_filename)
+            take2_row = take2_by_suffix.get(_scenario_suffix(scenario_filename))
+            if take2_row is None:
+                raise FileNotFoundError(f"Could not find take-2 row matching {scenario_filename}")
+            take2_id, _, _, _ = _parse_scenario_filename(take2_row["scenario"])
+
+            take1_video_path = self._resolve_testing_video_path(
+                scenario_id=scenario_id,
+                view=view,
+                take=TAKE1_TOKEN,
+                scenario_name=scenario_name,
+                fps=fps,
+            )
+            take2_video_path = self._resolve_testing_video_path(
+                scenario_id=take2_id,
+                view=view,
+                take=TAKE2_TOKEN,
+                scenario_name=scenario_name,
+                fps=fps,
+            )
+            switch_frame_path = self._resolve_switch_frame_path(
+                scenario_id=scenario_id,
+                view=view,
+                scenario_name=scenario_name,
+            )
+            take1_mask_path = self._resolve_real_mask_path(
+                scenario_id=scenario_id,
+                view=view,
+                take=TAKE1_TOKEN,
+                scenario_name=scenario_name,
+                fps=fps,
+            )
+            take2_mask_path = self._resolve_real_mask_path(
+                scenario_id=take2_id,
+                view=view,
+                take=TAKE2_TOKEN,
+                scenario_name=scenario_name,
+                fps=fps,
+            )
+            generated_video_path = (str(generated_dir_path /
+                                        row["generated_video_name"]) if generated_dir_path is not None else None)
+
+            scenarios.append(
+                PhysicsIQScenario(
+                    scenario_id=scenario_id,
+                    view=view,
+                    scenario_name=scenario_name,
+                    take1_video_path=str(take1_video_path),
+                    take2_video_path=str(take2_video_path),
+                    switch_frame_path=str(switch_frame_path),
+                    caption=row["description"],
+                    expected_gen_filename=row["generated_video_name"],
+                    generated_video_path=generated_video_path,
+                    take1_mask_path=str(take1_mask_path),
+                    take2_mask_path=str(take2_mask_path),
+                ))
+        return scenarios
+
+    def _resolve_testing_video_path(
+        self,
+        *,
+        scenario_id: str,
+        view: str,
+        take: str,
+        scenario_name: str,
+        fps: int,
+    ) -> Path:
+        target_dir = self.dataset_dir / "split-videos" / "testing" / f"{fps}FPS"
+        target_name = (f"{scenario_id}_testing-videos_{fps}FPS_{view}_{take}_{scenario_name}.mp4")
+        target_path = target_dir / target_name
+        if target_path.exists():
+            return target_path
+
+        # 30-FPS source: either present locally or auto-fetchable.
+        source_name = (f"{scenario_id}_testing-videos_30FPS_{view}_{take}_{scenario_name}.mp4")
+        source_rel = f"split-videos/testing/30FPS/{source_name}"
+        source_path = self.dataset_dir / source_rel
+        self._ensure_remote_asset(source_rel, source_path)
+        if fps == _DEFAULT_FPS:
+            return source_path
+
+        # FPS-convert and cache so repeat runs are free.
+        cache_dir = self.cache_dir / "split-videos" / "testing" / f"{fps}FPS"
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        cached_path = cache_dir / target_name
+        if not cached_path.exists():
+            _convert_video_fps(source_path, cached_path, fps_new=fps)
+        return cached_path
+
+    def _resolve_switch_frame_path(
+        self,
+        *,
+        scenario_id: str,
+        view: str,
+        scenario_name: str,
+    ) -> Path:
+        rel = (f"switch-frames/{scenario_id}_switch-frames_anyFPS_{view}_{scenario_name}.jpg")
+        target_path = self.dataset_dir / rel
+        self._ensure_remote_asset(rel, target_path)
+        return target_path
+
+    def _resolve_real_mask_path(
+        self,
+        *,
+        scenario_id: str,
+        view: str,
+        take: str,
+        scenario_name: str,
+        fps: int,
+    ) -> Path:
+        # Source release ships masks at 30 FPS only; non-30 rates are
+        # regenerated downstream from the (downsampled) real videos by
+        # the metric — see upstream ``run_physics_iq.py::ensure_binary_mask_structure``.
+        # We only auto-fetch 30 FPS here.
+        rel = (f"video-masks/real/30FPS/"
+               f"{scenario_id}_video-masks_30FPS_{view}_{take}_{scenario_name}.mp4")
+        target_path = self.dataset_dir / rel
+        self._ensure_remote_asset(rel, target_path)
+        if fps == _DEFAULT_FPS:
+            return target_path
+        # Caller asked for a non-30 rate; metric layer handles the
+        # regeneration. Return the canonical 30 FPS path so the metric
+        # always sees a valid mp4 it can transcode.
+        return target_path
+
+    def _ensure_remote_asset(self, rel_path: str, target_path: Path) -> Path:
+        """Download ``<bucket_url>/<rel_path>`` into *target_path* on miss.
+
+        Atomic via a sibling ``.part`` file; safe under concurrent runs
+        because the final ``rename`` is atomic on POSIX. Raises
+        ``FileNotFoundError`` if the file is missing and ``auto_download``
+        is False.
+        """
+        if target_path.exists():
+            return target_path
+        if not self.auto_download:
+            raise FileNotFoundError(f"Physics-IQ asset missing: {target_path}. "
+                                    "Set auto_download=True or pass dataset_root= a pre-downloaded copy.")
+        target_path.parent.mkdir(parents=True, exist_ok=True)
+        url = f"{self.bucket_url}/{rel_path.lstrip('/')}"
+        tmp_path = target_path.with_suffix(target_path.suffix + ".part")
+        try:
+            urlretrieve(url, tmp_path)
+        except Exception as exc:
+            if tmp_path.exists():
+                tmp_path.unlink()
+            raise FileNotFoundError(f"Failed to fetch Physics-IQ asset {url} -> {target_path}: "
+                                    f"{type(exc).__name__}: {exc}") from exc
+        tmp_path.rename(target_path)
+        return target_path
+
+
+# ---------------------------------------------------------------------------
+# Internal helpers
+# ---------------------------------------------------------------------------
+
+
+def _resolve_dataset_dir(repo_root: Path) -> Path:
+    nested = repo_root / "physics-IQ-benchmark"
+    if nested.exists():
+        return nested
+    return repo_root
+
+
+def _resolve_descriptions_path(repo_root: Path, dataset_dir: Path) -> Path:
+    """Prefer a co-located CSV under the user's dataset_root; fall back
+    to the copy vendored in this repo so ``get_dataset("physics_iq")``
+    works without external setup.
+    """
+    candidates = (
+        repo_root / "descriptions" / "descriptions.csv",
+        dataset_dir / "descriptions" / "descriptions.csv",
+    )
+    for path in candidates:
+        if path.exists():
+            return path
+    if _VENDORED_DESCRIPTIONS_CSV.is_file():
+        return _VENDORED_DESCRIPTIONS_CSV
+    raise FileNotFoundError("Could not locate Physics-IQ descriptions/descriptions.csv "
+                            f"(checked {[str(c) for c in candidates]} and vendored "
+                            f"{_VENDORED_DESCRIPTIONS_CSV})")
+
+
+def _parse_scenario_filename(filename: str) -> tuple[str, str, str, str]:
+    stem = Path(filename).name
+    if stem.endswith(".mp4"):
+        stem = stem[:-4]
+    parts = stem.split("_")
+    if len(parts) < 4:
+        raise ValueError(f"Unexpected Physics-IQ filename format: {filename}")
+    return parts[0], parts[1], parts[2], "_".join(parts[3:])
+
+
+def _scenario_suffix(filename: str) -> str:
+    _, view, _, scenario_name = _parse_scenario_filename(filename)
+    return f"{view}_{scenario_name}"
+
+
+def _scenario_to_row(scenario: PhysicsIQScenario) -> dict:
+    """Flatten a :class:`PhysicsIQScenario` into the public sample-dict shape."""
+    aux: dict = {
+        "scenario_id": scenario.scenario_id,
+        "scenario_name": scenario.scenario_name,
+        "switch_frame_path": scenario.switch_frame_path,
+        "expected_gen_filename": scenario.expected_gen_filename,
+    }
+    if scenario.generated_video_path is not None:
+        aux["generated_video_path"] = scenario.generated_video_path
+
+    row: dict = {
+        "prompt": scenario.caption,
+        "reference": scenario.take1_video_path,
+        "reference_take2": scenario.take2_video_path,
+        "scenario": scenario.scenario_id,
+        "view": scenario.view,
+        "auxiliary_info": aux,
+    }
+    if scenario.take1_mask_path is not None:
+        row["reference_mask"] = scenario.take1_mask_path
+    if scenario.take2_mask_path is not None:
+        row["reference_take2_mask"] = scenario.take2_mask_path
+    return row
+
+
+def _convert_video_fps(input_path: str | Path, output_path: str | Path, *, fps_new: int) -> None:
+    """Trim *input_path* to ``_DEFAULT_DURATION_SECONDS`` and re-encode at
+    *fps_new*, writing the result to *output_path*. Used to materialize
+    Physics-IQ's 30-FPS source release at user-requested rates.
+    """
+    input_path = Path(input_path)
+    output_path = Path(output_path)
+
+    cap = cv2.VideoCapture(str(input_path))
+    if not cap.isOpened():
+        raise FileNotFoundError(f"Could not open video for FPS conversion: {input_path}")
+
+    fps_original = cap.get(cv2.CAP_PROP_FPS)
+    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    duration = frame_count / fps_original if fps_original else 0.0
+    width, height = width - width % 2, height - height % 2
+    subclip_duration = min(_DEFAULT_DURATION_SECONDS, duration)
+
+    frames: list[np.ndarray] = []
+    for _ in range(int(subclip_duration * fps_original)):
+        ret, frame = cap.read()
+        if not ret:
+            break
+        frames.append(frame)
+    cap.release()
+
+    if not frames:
+        raise ValueError(f"No frames decoded from {input_path}")
+
+    frame_count_new = int(subclip_duration * fps_new)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    writer = cv2.VideoWriter(
+        str(output_path),
+        cv2.VideoWriter_fourcc(*"avc1"),
+        fps_new,
+        (width, height),
+    )
+    if frame_count_new <= 1:
+        writer.write(frames[0])
+        writer.release()
+        return
+
+    frame_count_original = len(frames)
+    for j in range(frame_count_new):
+        alpha = j * (frame_count_original - 1) / (frame_count_new - 1)
+        idx = int(alpha)
+        alpha -= idx
+        f1 = frames[idx].astype(np.float32)
+        f2 = frames[min(idx + 1, frame_count_original - 1)].astype(np.float32)
+        writer.write(((1.0 - alpha) * f1 + alpha * f2).astype(np.uint8))
+    writer.release()

fastvideo/eval/datasets/registry.py

@@ -0,0 +1,41 @@
+"""Registry for prompt-corpus datasets, mirroring :mod:`fastvideo.eval.registry`."""
+from __future__ import annotations
+
+from typing import Any, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from fastvideo.eval.datasets.base import BasePromptDataset
+
+_REGISTRY: dict[str, type[BasePromptDataset]] = {}
+
+
+def register_dataset(name: str):
+    """Decorator to register a prompt-dataset class.
+
+    Usage::
+
+        @register_dataset("vbench")
+        class VBenchPromptDataset(BasePromptDataset):
+            ...
+    """
+
+    def wrapper(cls):
+        cls.name = name
+        _REGISTRY[name] = cls
+        return cls
+
+    return wrapper
+
+
+def get_dataset(name: str, **kwargs: Any) -> BasePromptDataset:
+    """Instantiate a registered dataset by name."""
+    cls = _REGISTRY.get(name)
+    if cls is None:
+        available = ", ".join(sorted(_REGISTRY.keys()))
+        raise KeyError(f"Unknown dataset '{name}'. Available: {available}")
+    return cls(**kwargs)
+
+
+def list_datasets() -> list[str]:
+    """Return sorted list of all registered dataset names."""
+    return sorted(_REGISTRY.keys())

fastvideo/eval/datasets/vbench.py

@@ -0,0 +1,126 @@
+"""VBench prompt corpus.
+
+Single source of truth: upstream's ``VBench_full_info.json`` (946 entries,
+each with ``prompt_en``, a ``dimension`` list, optional ``auxiliary_info``
+keyed by dimension).
+"""
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+
+from fastvideo.eval.datasets.base import PromptDataset
+from fastvideo.eval.datasets.registry import register_dataset
+
+# VBench's official sampling protocol: 5 generations per prompt, except
+# temporal_flickering which requires 25 (averaging over 5 is too noisy
+# for a high-frequency-noise metric). See upstream prompts/README.md.
+TEMPORAL_FLICKERING_SAMPLES = 25
+DEFAULT_SAMPLES = 5
+
+_FULL_INFO_REL = "fastvideo/third_party/eval/vbench/vbench/VBench_full_info.json"
+
+
+def _locate_full_info() -> Path:
+    env = os.environ.get("VBENCH_FULL_INFO_JSON")
+    if env:
+        p = Path(env)
+        if p.is_file():
+            return p
+        raise FileNotFoundError(f"VBENCH_FULL_INFO_JSON={env} does not point at a file")
+    here = Path(__file__).resolve()
+    for ancestor in here.parents:
+        candidate = ancestor / _FULL_INFO_REL
+        if candidate.is_file():
+            return candidate
+        if (ancestor / ".git").exists():
+            break
+    raise FileNotFoundError("Could not locate VBench_full_info.json. Initialize the upstream "
+                            "submodule (`git submodule update --init "
+                            "fastvideo/third_party/eval/vbench`) or set VBENCH_FULL_INFO_JSON.")
+
+
+@register_dataset("vbench")
+class VBenchPromptDataset(PromptDataset):
+    """VBench prompts filtered by evaluation dimension.
+
+    Args:
+        dimensions: List of dimension names, or ``"all"``. Unknown
+            dimensions raise ``ValueError``.
+        full_info_path: Optional override for ``VBench_full_info.json``;
+            defaults to autodetection.
+
+    A prompt that belongs to several requested dimensions is yielded once;
+    its ``dimensions`` list carries all matches so the scorer can route.
+    """
+
+    description = ("VBench (Vchitect) prompt corpus, 946 prompts across 16 "
+                   "evaluation dimensions.")
+    supports_dimensions = True
+
+    def __init__(
+        self,
+        dimensions: list[str] | str = "all",
+        full_info_path: str | Path | None = None,
+    ) -> None:
+        super().__init__()
+        path = Path(full_info_path) if full_info_path else _locate_full_info()
+        with path.open() as f:
+            entries = json.load(f)
+
+        all_dims = sorted({d for e in entries for d in e["dimension"]})
+        if dimensions == "all":
+            self.dimensions: list[str] = all_dims
+        else:
+            unknown = set(dimensions) - set(all_dims)
+            if unknown:
+                raise ValueError(f"Unknown VBench dimensions: {sorted(unknown)}. "
+                                 f"Available: {all_dims}")
+            self.dimensions = list(dimensions)
+
+        wanted = set(self.dimensions)
+        for entry in entries:
+            relevant = [d for d in entry["dimension"] if d in wanted]
+            if not relevant:
+                continue
+            n = (TEMPORAL_FLICKERING_SAMPLES if "temporal_flickering" in relevant else DEFAULT_SAMPLES)
+
+            # Strip the outer {dim_name: ...} wrapper from upstream's aux
+            # schema so every metric reads its inputs from a flat dict.
+            #
+            # This unwraps exactly one level — the dimension key. Whatever
+            # shape lives inside is the metric's contract:
+            #
+            #   color:                {"color": {"color": "red"}}
+            #     → flat: {"color": "red"}                  (scalar)
+            #
+            #   object_class:         {"object_class": {"object": "person"}}
+            #     → flat: {"object": "person"}              (scalar)
+            #
+            #   multiple_objects:     {"multiple_objects": {"object": "a and b"}}
+            #     → flat: {"object": "a and b"}             (scalar)
+            #
+            #   spatial_relationship: {"spatial_relationship":
+            #                            {"spatial_relationship":
+            #                                {"object_a": ..., "object_b": ...,
+            #                                 "relationship": ...}}}
+            #     → flat: {"spatial_relationship": {object_a,object_b,relationship}}
+            #
+            # Note the spatial_relationship case keeps a nested inner dict
+            # by design — upstream double-wraps it, the SpatialRelationship
+            # metric reads ``aux["spatial_relationship"]`` expecting that
+            # inner dict. Don't "simplify" the wrapping away.
+            raw_aux = entry.get("auxiliary_info") or {}
+            flat_aux: dict = {}
+            for v in raw_aux.values():
+                if isinstance(v, dict):
+                    flat_aux.update(v)
+
+            self._rows.append({
+                "prompt": entry["prompt_en"],
+                "n_samples": n,
+                "dimensions": relevant,
+                "auxiliary_info": flat_aux,
+            })
+        self.full_info_path = path

fastvideo/eval/evaluator.py

@@ -0,0 +1,255 @@
+"""User-facing scorer.
+
+Layering (mirrors FastVideo's VideoGenerator → Worker pattern, but
+in-process)::
+
+    Evaluator               ← user-facing
+      └── EvalWorker × N    ← single-GPU; owns metric replicas
+      └── VideoPool         ← async path-→-tensor prefetch (per evaluate call)
+
+The constructor builds one :class:`EvalWorker` per GPU and loads every
+metric on every worker eagerly. :meth:`evaluate` is the single entry
+point: pass kwargs for one sample, or pass a list of sample dicts to
+fan-out across GPU replicas with pipelined decoding — same method,
+return type follows the input shape.
+"""
+from __future__ import annotations
+
+import threading
+from collections.abc import Iterable
+from typing import Any
+
+from fastvideo.eval.registry import (list_metrics, missing_dependencies, resolve_group)
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.worker import EvalWorker, add_pool_decode_ms
+from fastvideo.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class Evaluator:
+    """Pre-initialized scorer for repeated evaluation.
+
+    Parameters
+    ----------
+    metrics : list[str] | str
+        Metric names, group prefixes (``"vbench"``), or ``"all"``.
+    device : str
+        Single-GPU device (e.g. ``"cuda:0"``). Ignored when *num_gpus* > 1.
+    num_gpus : int
+        Number of GPU replicas. Each gets its own :class:`EvalWorker`.
+    compile : bool
+        Apply :func:`torch.compile` to each metric's ``_model``.
+    loader_threads : int
+        Background decode threads in the :class:`VideoPool`. Default 1
+        (hide decode behind compute). Bump for I/O-heavy benchmark sets
+        where one loader can't keep up with the workers.
+    prefetch_factor : int
+        ``pool max_size = prefetch_factor * num_workers``. Default 2 —
+        one sample being consumed, one prefetched per worker.
+    pre_upload : bool
+        If ``True`` (default), the worker uploads ``video`` /
+        ``reference`` tensors to its device once per sample so every
+        metric in the loop consumes the same GPU-resident tensor (no
+        per-metric ``.to(self.device)`` traffic). Set ``False`` for
+        training-time eval where the shared GPU-resident tensor would
+        compete with the training step for VRAM — each metric then
+        uploads its own copy as before.
+    """
+
+    def __init__(
+        self,
+        metrics: list[str] | str = "all",
+        device: str = "cuda:0",
+        num_gpus: int = 1,
+        compile: bool = False,
+        *,
+        loader_threads: int = 1,
+        prefetch_factor: int = 2,
+        pre_upload: bool = True,
+    ) -> None:
+        names = _resolve_metric_names(metrics)
+        if num_gpus > 1:
+            self._workers = [
+                EvalWorker(names, f"cuda:{i}", compile=compile, pre_upload=pre_upload) for i in range(num_gpus)
+            ]
+        else:
+            self._workers = [EvalWorker(names, device, compile=compile, pre_upload=pre_upload)]
+        self._loader_threads = max(1, loader_threads)
+        self._prefetch_factor = max(1, prefetch_factor)
+
+    @property
+    def num_gpus(self) -> int:
+        return len(self._workers)
+
+    @property
+    def metric_names(self) -> list[str]:
+        return self._workers[0].metric_names
+
+    def evaluate(
+        self,
+        samples: Iterable[dict] | None = None,
+        **kwargs,
+    ) -> dict[str, MetricResult] | list[dict[str, MetricResult]]:
+        """Score one sample (kwargs form) or many samples (list form).
+
+        ``video`` and ``reference`` may be either a pre-loaded
+        ``(T, C, H, W)`` tensor or a path-like (``str`` / ``Path``).
+        Paths in the list form are decoded asynchronously by a
+        :class:`VideoPool` that runs alongside metric compute, hiding
+        decode latency behind GPU work.
+
+        One sample::
+
+            ev.evaluate(video=tensor, text_prompt="...", fps=24.0)
+            ev.evaluate(video="path/to/clip.mp4", fps=24.0)
+
+        Many samples — pipelined decode + work-stealing across replicas::
+
+            ev.evaluate(samples=[
+                {"video": "a.mp4", "reference": "ref_a.mp4"},
+                {"video": "b.mp4", "reference": "ref_b.mp4"},
+                ...
+            ])
+
+        Multi-GPU dispatch fires automatically when ``num_gpus > 1`` and
+        the list form is used: every worker runs a consumer thread,
+        pulling decoded samples from the shared pool as it frees up.
+        The kwargs form always runs on worker 0 with no pool overhead.
+        """
+        if samples is None:
+            return self._workers[0].evaluate(**kwargs)
+
+        samples = list(samples)
+        if not samples:
+            return []
+        return self._evaluate_with_pool(samples)
+
+    def _evaluate_with_pool(self, samples: list[dict]) -> list[dict[str, MetricResult]]:
+        """Pipelined dispatch: ``VideoPool`` prefetches decoded samples;
+        consumers (one per worker) pop them and run metrics.
+
+        Decode order in the pool is non-deterministic — each pool item
+        carries its original input index so results are written back in
+        input order.
+        """
+        from fastvideo.eval.pool import VideoPool
+
+        n_workers = len(self._workers)
+        max_size = self._prefetch_factor * n_workers
+        results: list[Any] = [None] * len(samples)
+
+        with VideoPool(samples, loader_threads=self._loader_threads, max_size=max_size) as pool:
+            if n_workers == 1:
+                # Single-GPU: this thread is the consumer.
+                while True:
+                    item = pool.get()
+                    if item is None:
+                        break
+                    idx, decoded = item
+                    results[idx] = self._workers[0].evaluate(**decoded)
+            else:
+                # Multi-GPU: each worker runs its own consumer thread,
+                # pulling from the shared pool (work-stealing).
+                threads: list[threading.Thread] = []
+                for w in self._workers:
+                    t = threading.Thread(
+                        target=self._consumer_loop,
+                        args=(w, pool, results),
+                        daemon=True,
+                    )
+                    t.start()
+                    threads.append(t)
+                for t in threads:
+                    t.join()
+
+            # Attribute pool-side decode ms to the same global counter
+            # the worker uses so ``pop_timings()`` returns total decode
+            # time regardless of where the decode happened.
+            add_pool_decode_ms(pool.decode_ms_total)
+
+        return results
+
+    @staticmethod
+    def _consumer_loop(worker: EvalWorker, pool: Any, results: list) -> None:
+        while True:
+            item = pool.get()
+            if item is None:
+                return
+            idx, decoded = item
+            results[idx] = worker.evaluate(**decoded)
+
+    def release_cuda_memory(self) -> None:
+        """Free CUDA caches on every replica without dropping models."""
+        for w in self._workers:
+            w.release_cuda_memory()
+
+    def unload(self) -> None:
+        """Drop metric refs on every replica. Reverse with :meth:`reload`."""
+        for w in self._workers:
+            w.unload()
+
+    def reload(self) -> None:
+        """Rebuild metrics dropped by :meth:`unload`."""
+        for w in self._workers:
+            w.reload()
+
+    def shutdown(self) -> None:
+        """No-op kept for API compatibility.
+
+        Earlier versions of this class held a long-lived
+        ``ThreadPoolExecutor`` for multi-GPU round-robin dispatch and
+        needed an explicit shutdown to drain it. The current design
+        builds and tears down a :class:`VideoPool` per ``evaluate``
+        call, so there's no long-lived state to release here.
+        """
+
+
+def create_evaluator(
+    metrics: list[str] | str = "all",
+    device: str = "cuda:0",
+    num_gpus: int = 1,
+    compile: bool = False,
+) -> Evaluator:
+    return Evaluator(metrics=metrics, device=device, num_gpus=num_gpus, compile=compile)
+
+
+def _resolve_metric_names(metrics: list[str] | str) -> list[str]:
+    """Resolve metric names, supporting groups (``"vbench"``) and ``"all"``.
+
+    Group / ``"all"`` selectors silently skip metrics whose declared
+    dependencies aren't importable in this environment, with a single
+    warning per skipped metric. Explicit names (e.g. ``"vbench.color"``)
+    always pass through unchanged — the missing dep then surfaces as
+    :class:`ImportError` at construction time, which is what the user
+    asked for.
+    """
+    if metrics == "all":
+        return _filter_satisfied(list_metrics(), context="all")
+    if isinstance(metrics, str):
+        metrics = [metrics]
+
+    seen: set[str] = set()
+    names: list[str] = []
+    for m in metrics:
+        group = resolve_group(m)
+        candidates = _filter_satisfied(group, context=m) if group is not None else [m]
+        for n in candidates:
+            if n not in seen:
+                seen.add(n)
+                names.append(n)
+    return names
+
+
+def _filter_satisfied(names: list[str], *, context: str) -> list[str]:
+    """Drop metrics with missing deps from a group expansion."""
+    keep: list[str] = []
+    for n in names:
+        missing = missing_dependencies(n)
+        if missing:
+            logger.warning(
+                "eval: skipping %s in group '%s'; missing dependency: %s. "
+                "Install instructions: pass the metric name explicitly to see them.", n, context, ", ".join(missing))
+            continue
+        keep.append(n)
+    return keep

fastvideo/eval/io/__init__.py

@@ -0,0 +1,11 @@
+from fastvideo.eval.io.paths import (build_eval_kwargs, default_filename, glob_videos, sanitize_prompt)
+from fastvideo.eval.io.video import extract_frames, load_video
+
+__all__ = [
+    "load_video",
+    "extract_frames",
+    "sanitize_prompt",
+    "default_filename",
+    "glob_videos",
+    "build_eval_kwargs",
+]

fastvideo/eval/io/paths.py

@@ -0,0 +1,63 @@
+"""Filesystem helpers shared by eval scripts.
+
+Provides the prompt-sanitization, default filename convention, and
+``(row, video_path) → eval-kwargs`` builder. Free functions, not a
+class — :class:`fastvideo.eval.Evaluator` is the only stateful object
+in the eval surface; loops live in user scripts.
+"""
+from __future__ import annotations
+
+import re
+from pathlib import Path
+from typing import Any
+
+# Filesystem-unsafe characters mirrored from VideoGenerator's output-path
+# sanitizer so on-disk filenames match what the generator writes.
+_INVALID_CHARS = re.compile(r'[\\/:*?"<>|]')
+
+
+def sanitize_prompt(prompt: str, max_len: int = 100) -> str:
+    """Prompt → safe filename stem."""
+    s = _INVALID_CHARS.sub("", prompt[:max_len]).strip().strip(".")
+    return re.sub(r"\s+", " ", s) or "output"
+
+
+def default_filename(row: dict, idx: int, ext: str = ".mp4") -> str:
+    """``<sanitized-prompt>-<idx>.mp4`` — VBench-style."""
+    return f"{sanitize_prompt(row['prompt'])}-{idx}{ext}"
+
+
+def glob_videos(videos_dir: Path, row: dict, ext: str = ".mp4") -> list[Path]:
+    """Find every generated video for *row*, sorted by trailing ``-<idx>``."""
+    pattern = f"{sanitize_prompt(row['prompt'])}-*{ext}"
+    files = list(videos_dir.glob(pattern))
+
+    def _idx(p: Path) -> int:
+        try:
+            return int(p.stem.rsplit("-", 1)[1])
+        except (IndexError, ValueError):
+            return -1
+
+    return sorted(files, key=_idx)
+
+
+def build_eval_kwargs(row: dict, video_path: Path, *, fps: float = 24.0) -> dict[str, Any]:
+    """Build evaluator kwargs from a sample row + a video on disk.
+
+    Loads the video as ``(T,C,H,W)`` and adds the leading batch dim.
+    Forwards ``prompt`` (as ``text_prompt=[prompt]``) and
+    ``auxiliary_info`` (as ``[aux]``) when present on the row.
+    """
+    from fastvideo.eval.io.video import load_video
+
+    video = load_video(str(video_path))  # (T, C, H, W) in [0, 1]
+    kwargs: dict[str, Any] = {
+        "video": video.unsqueeze(0),  # (1, T, C, H, W)
+        "fps": fps,
+    }
+    if "prompt" in row:
+        kwargs["text_prompt"] = [row["prompt"]]
+    aux = row.get("auxiliary_info")
+    if aux:
+        kwargs["auxiliary_info"] = [aux]
+    return kwargs

fastvideo/eval/io/video.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+import torch
+from PIL import Image
+
+
+def load_video(source: str | torch.Tensor | list, **kwargs) -> torch.Tensor:
+    """Load a video as a ``(T, C, H, W)`` float32 tensor in ``[0, 1]``.
+
+    Supported *source* types:
+
+    * **str / Path** – path to ``.mp4`` / ``.avi`` / ``.gif`` file, or a
+      directory of frame images (sorted alphabetically).
+    * **torch.Tensor** – returned as-is after shape validation.
+    * **list[PIL.Image]** – stacked into a tensor.
+    """
+    if isinstance(source, torch.Tensor):
+        if source.ndim != 4:
+            raise ValueError(f"Expected video tensor with 4 dims (T,C,H,W), got {source.ndim}")
+        return source.float()
+
+    if isinstance(source, list):
+        frames = [_pil_to_tensor(img) for img in source]
+        return torch.stack(frames)
+
+    path = Path(source)
+    if path.is_dir():
+        return _load_frame_dir(path)
+    return _load_video_file(str(path))
+
+
+def extract_frames(video: torch.Tensor, n_frames: int | None = None) -> torch.Tensor:
+    """Uniformly sample *n_frames* from a ``(T, C, H, W)`` video tensor."""
+    if n_frames is None or n_frames >= video.shape[0]:
+        return video
+    indices = torch.linspace(0, video.shape[0] - 1, n_frames).long()
+    return video[indices]
+
+
+# --- Internal helpers ---
+
+
+def _pil_to_tensor(img: Image.Image) -> torch.Tensor:
+    arr = np.array(img.convert("RGB"))  # (H, W, 3) uint8
+    return torch.from_numpy(arr).permute(2, 0, 1).float() / 255.0
+
+
+def _load_frame_dir(path: Path) -> torch.Tensor:
+    exts = {".png", ".jpg", ".jpeg", ".bmp", ".webp"}
+    files = sorted(f for f in path.iterdir() if f.suffix.lower() in exts)
+    if not files:
+        raise FileNotFoundError(f"No image files found in {path}")
+    frames = [_pil_to_tensor(Image.open(f)) for f in files]
+    return torch.stack(frames)
+
+
+def _load_video_file(path: str) -> torch.Tensor:
+    # Try decord first (faster), fall back to torchvision
+    try:
+        return _load_with_decord(path)
+    except ImportError:
+        pass
+    return _load_with_torchvision(path)
+
+
+def _load_with_decord(path: str) -> torch.Tensor:
+    from decord import VideoReader, cpu
+
+    vr = VideoReader(path, ctx=cpu(0))
+    # (T, H, W, C) uint8
+    frames = vr.get_batch(list(range(len(vr)))).asnumpy()
+    # → (T, C, H, W) float32 [0, 1]
+    tensor = torch.from_numpy(frames).permute(0, 3, 1, 2).float() / 255.0
+    return tensor
+
+
+def _load_with_torchvision(path: str) -> torch.Tensor:
+    import torchvision.io
+
+    video, _, _ = torchvision.io.read_video(path, pts_unit="sec")
+    # torchvision returns (T, H, W, C) uint8
+    return video.permute(0, 3, 1, 2).float() / 255.0

fastvideo/eval/memory.py

@@ -0,0 +1,14 @@
+"""Small GPU-memory helper used by :class:`EvalWorker`."""
+
+from __future__ import annotations
+
+import gc
+
+import torch
+
+
+def clear_cache() -> None:
+    """Free GPU cache + run garbage collection."""
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()

fastvideo/eval/metrics/__init__.py

@@ -0,0 +1,29 @@
+"""Auto-discover and register all built-in metrics (recursive).
+
+Walks the metrics package tree and imports any leaf-package's
+``metric`` module so ``@register`` decorators fire. Path components
+starting with ``_`` are skipped (used for shared helpers like
+``optical_flow/_shared.py`` and `vbench/_grit_helper.py`).
+"""
+
+import importlib
+import os
+import contextlib
+
+
+def _walk(path: str, prefix: str):
+    """Recursively yield (module_name, is_pkg) for non-underscore packages."""
+    for entry in os.listdir(path):
+        if entry.startswith("_") or entry.startswith("."):
+            continue
+        full = os.path.join(path, entry)
+        if os.path.isdir(full) and os.path.exists(os.path.join(full, "__init__.py")):
+            sub_prefix = f"{prefix}.{entry}"
+            yield (sub_prefix, True)
+            yield from _walk(full, sub_prefix)
+
+
+for _pkg_path in __path__:
+    for _modname, _ispkg in _walk(_pkg_path, __name__):
+        with contextlib.suppress(ModuleNotFoundError):
+            importlib.import_module(f"{_modname}.metric")

fastvideo/eval/metrics/base.py

@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+
+import torch
+
+from fastvideo.eval.types import MetricResult
+
+
+class BaseMetric(ABC):
+    """Abstract base class for all eval metrics.
+
+    Subclasses must implement :meth:`compute`. Optionally override
+    :meth:`setup` to eagerly load models.
+
+    Metrics that need to chunk along the time dimension (frames or frame
+    pairs) for memory reasons should hardcode their own chunk size in
+    ``__init__`` (see ``optical_flow`` for the canonical example). Eval
+    always processes one video per :meth:`Evaluator.evaluate` call;
+    ``compute`` therefore receives a single sample, not a batch.
+    """
+
+    name: str = ""
+    requires_reference: bool = True
+    higher_is_better: bool = True
+    dependencies: list[str] = []
+    needs_gpu: bool = False
+    backbone: str | None = None
+
+    # Default time-dim chunk size for metrics that batch internally over
+    # frames or frame-pairs. Override in subclass __init__ if needed
+    # (see ``optical_flow``, ``motion_smoothness``, ``dynamic_degree``).
+    _chunk_size: int | None = None
+
+    def __init__(self) -> None:
+        self._device: torch.device = torch.device("cpu")
+
+    @property
+    def device(self) -> torch.device:
+        return self._device
+
+    def to(self, device: str | torch.device) -> BaseMetric:
+        """Move metric (and its internal models) to *device*."""
+        self._device = torch.device(device)
+        return self
+
+    def setup(self) -> None:  # noqa: B027 - intentionally optional override
+        """Eagerly load models. Called once by :class:`EvalWorker`.
+
+        Default is a no-op; metrics with no eager state (pixel math,
+        closed-form ops) inherit this. Override only if your metric
+        needs to load weights.
+        """
+
+    def _skip(self, sample: dict, reason: str) -> MetricResult:
+        """Return a skipped result (``score=None`` + reason in details)."""
+        return MetricResult(name=self.name, score=None, details={"skipped": reason})
+
+    @abstractmethod
+    def compute(self, sample: dict) -> MetricResult:
+        """Compute the metric on a single sample.
+
+        ``sample["video"]`` is ``(T, C, H, W)`` float in ``[0, 1]``.
+        ``sample["reference"]`` (if used) has the same shape.
+
+        If required inputs are missing, return ``self._skip(sample, reason)``
+        instead of raising.
+        """
+        ...

fastvideo/eval/metrics/common/lpips/metric.py

@@ -0,0 +1,70 @@
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("common.lpips")
+class LPIPSMetric(BaseMetric):
+    name = "common.lpips"
+    requires_reference = True
+    higher_is_better = False
+    needs_gpu = True
+    dependencies = ["lpips"]
+
+    def __init__(self, net: str = "alex", chunk_size: int = 8) -> None:
+        super().__init__()
+        self.net = net
+        # Per-frame AlexNet feature maps at 1080p run ~500 MB each. A
+        # full 121-frame chunk peaks around 60 GB; chunking to 8 frames
+        # drops that to ~5 GB with identical numerical output.
+        self._chunk_size = chunk_size
+        self._model: Any = None
+
+    def to(self, device: str | torch.device) -> LPIPSMetric:
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        import lpips as lpips_lib
+        self._model = lpips_lib.LPIPS(net=self.net).to(self.device)
+        self._model.eval()
+
+    def compute(self, sample: dict) -> MetricResult:
+        if self._model is None:
+            self.setup()
+
+        # When the worker pre-uploaded inputs (default), these are
+        # already on ``self.device`` and the ``.to(...)`` below is a
+        # no-op. With ``pre_upload=False`` the worker keeps them on CPU
+        # and this metric pays the transfer just like before.
+        gen = sample["video"].float().to(self.device, non_blocking=True)
+        ref = sample["reference"].float().to(self.device, non_blocking=True)
+
+        n = min(gen.shape[0], ref.shape[0])
+        gen, ref = gen[:n] * 2.0 - 1.0, ref[:n] * 2.0 - 1.0
+
+        chunk = self._chunk_size or n
+        all_scores = []
+        with torch.no_grad():
+            for i in range(0, n, chunk):
+                s = self._model(gen[i:i + chunk], ref[i:i + chunk]).squeeze()
+                if s.dim() == 0:
+                    s = s.unsqueeze(0)
+                all_scores.append(s)
+        scores = torch.cat(all_scores)  # (n,)
+
+        return MetricResult(
+            name=self.name,
+            score=float(scores.mean()),
+            details={"per_frame": scores.tolist()},
+        )

fastvideo/eval/metrics/common/psnr/metric.py

@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("common.psnr")
+class PSNRMetric(BaseMetric):
+    name = "common.psnr"
+    requires_reference = True
+    higher_is_better = True
+    # PSNR is `((gen - ref)**2).mean(...)` plus log — memory-bandwidth-
+    # bound on host (~6 GB read + 3 GB write per video pair at 1080p ×
+    # 121 fr). Trivial on GPU and frees the host bus for the loader.
+    needs_gpu = True
+
+    def __init__(self, max_val: float = 1.0, chunk_size: int = 32) -> None:
+        super().__init__()
+        self.max_val = max_val
+        # (gen - ref)**2 at 1080p × 121 fr allocates a full ~3 GB
+        # intermediate. chunk=32 caps that at ~800 MB with identical
+        # numerical output.
+        self._chunk_size = chunk_size
+
+    def compute(self, sample: dict) -> MetricResult:
+        gen = sample["video"].float().to(self.device)  # (T, C, H, W)
+        ref = sample["reference"].float().to(self.device)
+        n = min(gen.shape[0], ref.shape[0])
+        gen, ref = gen[:n], ref[:n]
+
+        # Per-frame MSE → PSNR, chunked so the squared-diff intermediate
+        # never holds the whole clip at once.
+        chunk = self._chunk_size or n
+        mse_parts = []
+        for i in range(0, n, chunk):
+            g = gen[i:i + chunk]
+            r = ref[i:i + chunk]
+            mse_parts.append(((g - r)**2).mean(dim=(1, 2, 3)))
+        mse = torch.cat(mse_parts)  # (T,)
+        psnr = 10.0 * torch.log10(self.max_val**2 / mse.clamp(min=1e-10))
+
+        return MetricResult(
+            name=self.name,
+            score=psnr.mean().item(),
+            details={"per_frame": psnr.tolist()},
+        )

fastvideo/eval/metrics/common/ssim/metric.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+
+import torch
+import torch.nn.functional as F
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+def _ssim_per_frame(
+    x: torch.Tensor,
+    y: torch.Tensor,
+    window_size: int = 11,
+    C1: float = 0.01**2,
+    C2: float = 0.03**2,
+) -> torch.Tensor:
+    """Compute SSIM for each frame.  Returns ``(N,)`` tensor where N = number of frames."""
+    channels = x.shape[1]
+    kernel = _gaussian_kernel(window_size, 1.5, channels, x.device, x.dtype)
+
+    mu_x = F.conv2d(x, kernel, groups=channels, padding=window_size // 2)
+    mu_y = F.conv2d(y, kernel, groups=channels, padding=window_size // 2)
+
+    mu_x2 = mu_x * mu_x
+    mu_y2 = mu_y * mu_y
+    mu_xy = mu_x * mu_y
+
+    sigma_x2 = F.conv2d(x * x, kernel, groups=channels, padding=window_size // 2) - mu_x2
+    sigma_y2 = F.conv2d(y * y, kernel, groups=channels, padding=window_size // 2) - mu_y2
+    sigma_xy = F.conv2d(x * y, kernel, groups=channels, padding=window_size // 2) - mu_xy
+
+    num = (2 * mu_xy + C1) * (2 * sigma_xy + C2)
+    den = (mu_x2 + mu_y2 + C1) * (sigma_x2 + sigma_y2 + C2)
+    ssim_map = num / den
+
+    return ssim_map.mean(dim=(1, 2, 3))
+
+
+def _gaussian_kernel(size: int, sigma: float, channels: int, device, dtype):
+    coords = torch.arange(size, device=device, dtype=dtype) - size // 2
+    g = torch.exp(-coords**2 / (2 * sigma**2))
+    g = g / g.sum()
+    kernel_2d = g.unsqueeze(1) * g.unsqueeze(0)
+    kernel = kernel_2d.unsqueeze(0).unsqueeze(0).repeat(channels, 1, 1, 1)
+    return kernel
+
+
+@register("common.ssim")
+class SSIMMetric(BaseMetric):
+    name = "common.ssim"
+    requires_reference = True
+    higher_is_better = True
+    # SSIM is 5 depthwise conv2d's per chunk — pure GPU territory at any
+    # interesting resolution. At 1080p × 121 frames the CPU path takes
+    # ~5–10 s per pair vs <100 ms on GPU. Keeping it on CPU also made
+    # the metric fight the pipelined loader thread for DDR bandwidth.
+    needs_gpu = True
+
+    def __init__(self, window_size: int = 11, chunk_size: int = 16) -> None:
+        super().__init__()
+        self.window_size = window_size
+        # Each conv2d output is (chunk, 3, H, W); SSIM allocates ~5–7
+        # such intermediates plus inputs. At 1080p with chunk=121 this
+        # peaks ~25 GB; chunk=16 brings peak to ~4 GB with no change
+        # in output.
+        self._chunk_size = chunk_size
+
+    def compute(self, sample: dict) -> MetricResult:
+        gen = sample["video"].float().to(self.device)  # (T, C, H, W)
+        ref = sample["reference"].float().to(self.device)
+        n = min(gen.shape[0], ref.shape[0])
+        gen, ref = gen[:n], ref[:n]
+
+        chunk = self._chunk_size or n
+        parts = []
+        for i in range(0, n, chunk):
+            parts.append(_ssim_per_frame(gen[i:i + chunk], ref[i:i + chunk], self.window_size))
+        per_frame = torch.cat(parts)  # (n,)
+
+        return MetricResult(
+            name=self.name,
+            score=per_frame.mean().item(),
+            details={"per_frame": per_frame.tolist()},
+        )

fastvideo/eval/metrics/optical_flow/_shared.py

@@ -0,0 +1,297 @@
+"""Shared helpers for optical-flow metrics.
+
+Both ``optical_flow.gt_optical_flow`` and
+``optical_flow.synthetic_optical_flow`` extract per-frame flow with
+``ptlflow`` and reduce it through the same per-pixel / per-frame /
+temporal aggregation pipeline. The pipeline lives here so the two
+metrics stay byte-identical on the comparison side and only differ in
+how they construct the *reference* flow field.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import torch
+
+_PER_FRAME_AGG_KEYS: tuple[str, ...] = (
+    "mf_epe",
+    "mf_angle_err",
+    "mf_cosine",
+    "mf_mag_ratio",
+    "pixel_epe_mean",
+    "pixel_epe_max",
+    "px_angle_rmse",
+    "grid_epe_mean",
+    "grid_epe_max",
+    "fl_all",
+    "foe_dist",
+    "flow_kl_2d",
+)
+
+
+def _trapezoid(vals: np.ndarray) -> float:
+    fn = getattr(np, "trapezoid", None) or np.trapz
+    return float(fn(vals))
+
+
+def _estimate_foe(
+    flow: np.ndarray,
+    step: int = 8,
+    min_mag: float = 0.5,
+) -> tuple[float, float]:
+    """Least-squares Focus of Expansion. Returns (fx, fy)."""
+    H, W = flow.shape[:2]
+    ys = np.arange(step // 2, H, step)
+    xs = np.arange(step // 2, W, step)
+    yy, xx = np.meshgrid(ys, xs, indexing="ij")
+    yy = yy.ravel()
+    xx = xx.ravel()
+    uu = flow[yy, xx, 0]
+    vv = flow[yy, xx, 1]
+
+    mag = np.sqrt(uu**2 + vv**2)
+    valid = mag > min_mag
+    if valid.sum() < 10:
+        return W / 2.0, H / 2.0
+
+    xx = xx[valid].astype(np.float64)
+    yy = yy[valid].astype(np.float64)
+    uu = uu[valid].astype(np.float64)
+    vv = vv[valid].astype(np.float64)
+
+    # v * fx - u * fy = v * x - u * y
+    A = np.column_stack([vv, -uu])
+    b = vv * xx - uu * yy
+    result, _, _, _ = np.linalg.lstsq(A, b, rcond=None)
+    return float(result[0]), float(result[1])
+
+
+def _flow_kl_2d(
+    flow_a: np.ndarray,
+    flow_b: np.ndarray,
+    n_angle_bins: int = 36,
+    n_mag_bins: int = 20,
+    min_mag: float = 0.5,
+) -> float:
+    """KL(P_a || P_b) over a joint (angle, log-magnitude) histogram."""
+
+    def _hist(flow: np.ndarray) -> np.ndarray | None:
+        u, v = flow[:, :, 0].ravel(), flow[:, :, 1].ravel()
+        mag = np.sqrt(u**2 + v**2)
+        angle = np.degrees(np.arctan2(v, u)) % 360
+        valid = mag >= min_mag
+        if valid.sum() < 10:
+            return None
+        mag = mag[valid]
+        angle = angle[valid]
+        mag_max = max(mag.max(), min_mag + 1.0)
+        mag_edges = np.logspace(np.log10(min_mag), np.log10(mag_max), n_mag_bins + 1)
+        angle_edges = np.linspace(0, 360, n_angle_bins + 1)
+        h, _, _ = np.histogram2d(angle, mag, bins=[angle_edges, mag_edges])
+        return h
+
+    ha, hb = _hist(flow_a), _hist(flow_b)
+    if ha is None or hb is None:
+        return 0.0
+    eps = 1.0
+    p = (ha + eps) / (ha + eps).sum()
+    q = (hb + eps) / (hb + eps).sum()
+    return float((p * np.log(p / q)).sum())
+
+
+def compute_frame_metrics(
+    flow_gt: np.ndarray,
+    flow_gen: np.ndarray,
+    grid_size: int = 8,
+    min_mag: float = 0.5,
+    max_mag_pct: float = 80.0,
+) -> dict[str, float]:
+    """Per-frame comparison metrics between two HxWx2 flow fields.
+
+    Port of mhuo's compute_frame_metrics — see ptlflow_validation.py.
+    """
+    metrics: dict[str, float] = {}
+
+    gt_mag_map = np.linalg.norm(flow_gt, axis=2)
+    gen_mag_map = np.linalg.norm(flow_gen, axis=2)
+    max_mag_map = np.maximum(gt_mag_map, gen_mag_map)
+    mag_hi = np.percentile(max_mag_map, max_mag_pct)
+    mag_mask = (max_mag_map >= min_mag) & (max_mag_map <= mag_hi)
+    n_valid = int(mag_mask.sum())
+
+    if n_valid > 0:
+        mean_gt = flow_gt[mag_mask].mean(axis=0)
+        mean_gen = flow_gen[mag_mask].mean(axis=0)
+    else:
+        mean_gt = flow_gt.reshape(-1, 2).mean(axis=0)
+        mean_gen = flow_gen.reshape(-1, 2).mean(axis=0)
+
+    metrics["mf_epe"] = float(np.linalg.norm(mean_gt - mean_gen))
+
+    mf_min_mag = 0.1
+    mag_gt = float(np.linalg.norm(mean_gt))
+    mag_gen = float(np.linalg.norm(mean_gen))
+    if mag_gt < mf_min_mag and mag_gen < mf_min_mag:
+        metrics["mf_angle_err"] = 0.0
+        metrics["mf_cosine"] = 1.0
+    elif mag_gt < mf_min_mag or mag_gen < mf_min_mag:
+        metrics["mf_angle_err"] = 90.0
+        metrics["mf_cosine"] = 0.0
+    elif mag_gt > 1e-6 and mag_gen > 1e-6:
+        cos_sim = float(np.dot(mean_gt, mean_gen) / (mag_gt * mag_gen))
+        cos_sim = float(np.clip(cos_sim, -1.0, 1.0))
+        metrics["mf_angle_err"] = float(np.degrees(np.arccos(cos_sim)))
+        metrics["mf_cosine"] = cos_sim
+    else:
+        metrics["mf_angle_err"] = 0.0
+        metrics["mf_cosine"] = 1.0
+
+    metrics["mf_mag_ratio"] = float(mag_gen / mag_gt) if mag_gt > 1e-6 else 1.0
+
+    epe_map = np.linalg.norm(flow_gt - flow_gen, axis=2)
+    if n_valid > 0:
+        metrics["pixel_epe_mean"] = float(epe_map[mag_mask].mean())
+        metrics["pixel_epe_max"] = float(epe_map[mag_mask].max())
+    else:
+        metrics["pixel_epe_mean"] = float(epe_map.mean())
+        metrics["pixel_epe_max"] = float(epe_map.max())
+
+    valid = mag_mask & (gt_mag_map > 0.5) & (gen_mag_map > 0.5)
+    if valid.sum() > 0:
+        dot = (flow_gt[:, :, 0] * flow_gen[:, :, 0] + flow_gt[:, :, 1] * flow_gen[:, :, 1])
+        cos_map = np.clip(dot / (gt_mag_map * gen_mag_map + 1e-8), -1.0, 1.0)
+        angle_map = np.degrees(np.arccos(cos_map))
+        metrics["px_angle_rmse"] = float(np.sqrt((angle_map[valid]**2).mean()))
+    else:
+        metrics["px_angle_rmse"] = 0.0
+
+    H, W = epe_map.shape
+    gh, gw = H // grid_size, W // grid_size
+    grid_vals = []
+    for gi in range(grid_size):
+        for gj in range(grid_size):
+            cell_mask = mag_mask[gi * gh:(gi + 1) * gh, gj * gw:(gj + 1) * gw]
+            cell_epe = epe_map[gi * gh:(gi + 1) * gh, gj * gw:(gj + 1) * gw]
+            if cell_mask.sum() > 0:
+                grid_vals.append(float(cell_epe[cell_mask].mean()))
+            else:
+                grid_vals.append(float(cell_epe.mean()))
+    metrics["grid_epe_mean"] = float(np.mean(grid_vals))
+    metrics["grid_epe_max"] = float(np.max(grid_vals))
+
+    if n_valid > 0:
+        outlier = (epe_map > 3.0) & (epe_map > 0.05 * gt_mag_map) & mag_mask
+        metrics["fl_all"] = float(outlier.sum() / n_valid)
+    else:
+        outlier = (epe_map > 3.0) & (epe_map > 0.05 * gt_mag_map)
+        metrics["fl_all"] = float(outlier.mean())
+
+    foe_gt_x, foe_gt_y = _estimate_foe(flow_gt)
+    foe_gen_x, foe_gen_y = _estimate_foe(flow_gen)
+    metrics["foe_dist"] = float(np.sqrt((foe_gt_x - foe_gen_x)**2 + (foe_gt_y - foe_gen_y)**2))
+
+    metrics["flow_kl_2d"] = _flow_kl_2d(flow_gt, flow_gen)
+    return metrics
+
+
+def aggregate_temporal(per_frame: list[dict[str, float]], ) -> dict[str, float | int | None]:
+    """Aggregate per-frame metric dicts into mean/std/max/auc/onset summaries.
+
+    Port of mhuo's compute_temporal_metrics.
+    """
+    n = len(per_frame)
+    if n == 0:
+        return {"n_frames": 0}
+
+    summary: dict[str, float | int | None] = {"n_frames": n}
+    series: dict[str, np.ndarray] = {k: np.array([m[k] for m in per_frame]) for k in _PER_FRAME_AGG_KEYS}
+    for name, vals in series.items():
+        summary[f"{name}_mean"] = float(vals.mean())
+        summary[f"{name}_std"] = float(vals.std())
+        summary[f"{name}_max"] = float(vals.max())
+        summary[f"{name}_auc"] = _trapezoid(vals) / max(n - 1, 1)
+
+    epe_series = series["pixel_epe_mean"]
+    window = min(5, n)
+    if n >= window:
+        baseline = float(np.median(epe_series[:window]))
+        threshold = max(baseline * 2.0, 1.0)
+        kernel = np.ones(window) / window
+        smoothed = np.convolve(epe_series, kernel, mode="valid")
+        divergence_frame: int | None = None
+        for i, val in enumerate(smoothed):
+            if val > threshold:
+                divergence_frame = int(i)
+                break
+        summary["divergence_onset_frame"] = divergence_frame
+        summary["divergence_threshold"] = float(threshold)
+    else:
+        summary["divergence_onset_frame"] = None
+        summary["divergence_threshold"] = None
+    return summary
+
+
+def tensor_to_bgr_list(video: torch.Tensor) -> list[np.ndarray]:
+    """Convert ``(T, C, H, W)`` float [0,1] to a list of HWC BGR uint8 frames.
+
+    Performs the cast + permute + BGR swap on the input's device, then
+    transfers once. This avoids 121 per-frame ``.cpu().numpy()`` calls
+    moving 3 GB of float32 across PCIe when the input lives on GPU
+    (which is what ``Evaluator(pre_upload=True)`` produces). One uint8
+    transfer is ~4× less bytes than the per-frame float32 round-trips.
+    """
+    # (T, C, H, W) float [0,1] → (T, H, W, C) uint8 BGR, all on-device.
+    bgr_u8 = (video.float() * 255.0).clamp(0, 255).to(torch.uint8).permute(0, 2, 3, 1).flip(-1).contiguous()
+    arr = bgr_u8.cpu().numpy()  # single transfer
+    return [arr[t] for t in range(arr.shape[0])]
+
+
+def load_ptlflow_model(model_name: str, ckpt: str, device: torch.device):
+    """Load a ``ptlflow`` model on *device* in eval mode."""
+    import ptlflow
+    model = ptlflow.get_model(model_name, ckpt_path=ckpt)
+    model.eval()
+    return model.to(device)
+
+
+def extract_video_flows(
+    model,
+    video: torch.Tensor,  # (T, C, H, W) float [0, 1]
+    *,
+    chunk: int,
+    device: torch.device,
+) -> list[np.ndarray]:
+    """Run *model* on every consecutive frame pair in *video*.
+
+    Returns a list of HxWx2 flow arrays of length ``T - 1``.
+    """
+    from ptlflow.utils.io_adapter import IOAdapter
+
+    h, w = video.shape[2], video.shape[3]
+    io_adapter = IOAdapter(
+        output_stride=model.output_stride,
+        input_size=(h, w),
+        cuda=(device.type == "cuda"),
+    )
+    bgr_frames = tensor_to_bgr_list(video)
+    pairs = [(bgr_frames[i], bgr_frames[i + 1]) for i in range(len(bgr_frames) - 1)]
+
+    flows: list[np.ndarray] = []
+    for start in range(0, len(pairs), chunk):
+        end = min(start + chunk, len(pairs))
+        pair_tensors = []
+        for f1, f2 in pairs[start:end]:
+            inputs = io_adapter.prepare_inputs([f1, f2])
+            pair_tensors.append(inputs["images"])
+        batched_images = torch.cat(pair_tensors, dim=0)
+        with torch.no_grad():
+            preds = model({"images": batched_images})
+        preds["images"] = batched_images
+        preds = io_adapter.unscale(preds)
+        flows_tensor = preds["flows"]
+        if flows_tensor.dim() == 5:
+            flows_tensor = flows_tensor.squeeze(1)
+        for i in range(flows_tensor.shape[0]):
+            flow = flows_tensor[i].detach().cpu().permute(1, 2, 0).numpy()
+            flows.append(flow)
+    return flows

fastvideo/eval/metrics/optical_flow/gt_optical_flow/metric.py

@@ -0,0 +1,118 @@
+"""Compare optical flow extracted from a generated video against optical
+flow extracted from a ground-truth reference video.
+
+Both flows are produced by the same ``ptlflow`` model (default
+``dpflow``/``things``). The resulting per-pixel / per-frame / temporal
+metric set is identical to ``synthetic_optical_flow`` — only the way the
+*reference* flow is constructed differs.
+"""
+
+from __future__ import annotations
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.metrics.optical_flow._shared import (
+    aggregate_temporal,
+    compute_frame_metrics,
+    extract_video_flows,
+    load_ptlflow_model,
+)
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("optical_flow.gt_optical_flow")
+class GtOpticalFlowMetric(BaseMetric):
+    """Per-pixel / per-frame / temporal flow comparison vs. a reference video.
+
+    The headline ``score`` is ``pixel_epe_mean_mean`` (lower is better);
+    every other scalar lives in ``details`` so downstream consumers can
+    pick whichever one they care about.
+    """
+
+    name = "optical_flow.gt_optical_flow"
+    requires_reference = True
+    higher_is_better = False
+    needs_gpu = True
+    backbone = "optical_flow"
+    dependencies = ["ptlflow"]
+
+    def __init__(
+        self,
+        model_name: str = "dpflow",
+        ckpt: str = "things",
+        min_mag: float = 0.5,
+        max_mag_pct: float = 80.0,
+        grid_size: int = 8,
+    ) -> None:
+        super().__init__()
+        self.model_name = model_name
+        self.ckpt = ckpt
+        self.min_mag = min_mag
+        self.max_mag_pct = max_mag_pct
+        self.grid_size = grid_size
+        self._model = None
+        # Frame-pair batch size per DPFlow forward. The correlation
+        # volume scales ~quadratically with input resolution and the
+        # cost-volume tensor is roughly 4 GB per pair at 1080p — so
+        # batching 16 pairs requested ~63 GB and OOMed on H200 (matches
+        # mhuo's reference impl, which runs one pair per forward, see
+        # ``mhuo/ptlflow/eval_flow_divergence.py:99-105``). Default 1
+        # is memory-safe at any resolution; bump for low-res to amortize
+        # Python loop overhead.
+        self._chunk_size = 1
+
+    def to(self, device: str | torch.device) -> GtOpticalFlowMetric:
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        self._model = load_ptlflow_model(self.model_name, self.ckpt, self.device)
+
+    def compute(self, sample: dict) -> MetricResult:
+        if self._model is None:
+            self.setup()
+
+        gen_video = sample["video"].float()  # (T, C, H, W)
+        ref_video = sample["reference"].float()
+        n = min(gen_video.shape[0], ref_video.shape[0])
+        gen_video, ref_video = gen_video[:n], ref_video[:n]
+        if n < 2:
+            raise ValueError("Need at least 2 frames to compute optical flow")
+
+        chunk = self._chunk_size or 16
+        gen_flows = extract_video_flows(
+            self._model,
+            gen_video,
+            chunk=chunk,
+            device=self.device,
+        )
+        ref_flows = extract_video_flows(
+            self._model,
+            ref_video,
+            chunk=chunk,
+            device=self.device,
+        )
+        per_frame = [
+            compute_frame_metrics(
+                rf,
+                gf,
+                grid_size=self.grid_size,
+                min_mag=self.min_mag,
+                max_mag_pct=self.max_mag_pct,
+            ) for rf, gf in zip(ref_flows, gen_flows, strict=False)
+        ]
+        summary = aggregate_temporal(per_frame)
+        score = summary.get("pixel_epe_mean_mean")
+        details = dict(summary)
+        details["per_frame_metrics"] = per_frame
+        return MetricResult(
+            name=self.name,
+            score=float(score) if score is not None else None,
+            details=details,
+        )

fastvideo/eval/metrics/optical_flow/synthetic_optical_flow/_thirdperson.py

@@ -0,0 +1,332 @@
+"""Third-person synthetic optical-flow generator (Option B, no depth).
+
+Camera frame convention (OpenCV): x=right, y=down, z=forward.
+
+Per-frame inputs from the action stream:
+    keyboard : (6,) — [W, S, A, D, turn_left, turn_right]
+    mouse    : (2,) — [pitch, yaw]   (pitch may be sign-flipped per sample,
+                                      see ``mouse_pitch_sign``)
+
+Mapping to camera kinematics:
+    omega_x = alpha_pitch * mouse_pitch                       (camera pitch)
+    omega_y = alpha_yaw   * mouse_yaw
+            + alpha_turn  * (turn_right - turn_left)          (camera yaw)
+    omega_z = 0                                               (no roll)
+
+    T_avatar_x = beta_strafe * (D - A)
+    T_avatar_y = 0
+    T_avatar_z = beta_fwd    * (W - S)
+
+Off-pivot correction. The orbit camera rotates about the avatar pivot at
+``r = (0, r_y, r_z)`` in camera-local coords, not about the optical
+center. A rotation by omega about that pivot is kinematically equivalent
+to a rotation about the optical center plus a translation
+``T_orbit = -(omega x r)``. So:
+
+    T_total = T_avatar + T_orbit
+
+Flow (no depth, Z = 1):
+    u_R = (xy/f)*ωx - (f + x²/f)*ωy + y*ωz
+    v_R = (f + y²/f)*ωx - (xy/f)*ωy - x*ωz
+    u_T = -f*Tx + x*Tz
+    v_T = -f*Ty + y*Tz
+
+The Z=1 collapse means strafe (Tx-only) produces a uniform horizontal
+field. Forward motion (Tz) still has the right radial direction
+structure since u_T scales with x, just no depth-modulated magnitude.
+Angle-family metrics survive this; per-pixel magnitude metrics will be
+biased on parallax-rich backgrounds. That's the explicit cost of
+declining to use generated-video depth.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import asdict, dataclass, field
+from pathlib import Path
+
+import numpy as np
+
+
+@dataclass
+class ThirdPersonCalibration:
+    """Fitted parameters for a third-person rig.
+
+    ``r_y`` defaults to 0 (avatar at camera height). ``focal_length`` is
+    in pixels. ``init_pitch`` is the camera's rest pitch (radians, negative
+    means tilted down — typical over-the-shoulder framing). User mouse-pitch
+    input is integrated on top of this baseline.
+    """
+    alpha_yaw: float
+    alpha_pitch: float
+    alpha_turn: float
+    beta_fwd: float
+    beta_strafe: float
+    focal_length: float
+    r_z: float
+    r_y: float = 0.0
+    init_pitch: float = 0.0
+    notes: str = ""
+    fit_metadata: dict = field(default_factory=dict)
+
+    def to_json(self, path: str | Path) -> None:
+        Path(path).write_text(json.dumps(asdict(self), indent=2))
+
+    @classmethod
+    def from_dict(cls, d: dict) -> ThirdPersonCalibration:
+        known = {f.name for f in cls.__dataclass_fields__.values()}
+        return cls(**{k: v for k, v in d.items() if k in known})
+
+
+def load_calibration(path: str | Path) -> ThirdPersonCalibration:
+    return ThirdPersonCalibration.from_dict(json.loads(Path(path).read_text()))
+
+
+class ThirdPersonFlowGenerator:
+    """Vectorized 3P synthetic-flow generator. No depth.
+
+    Parameters
+    ----------
+    calibration : ThirdPersonCalibration
+    frame_shape : (H, W)
+    mouse_pitch_sign : +1 or -1 — sample-level flag from metadata
+        (``mouse_pitch_flipped: true`` in mhuo's data ⇒ -1).
+    """
+
+    def __init__(
+        self,
+        calibration: ThirdPersonCalibration,
+        frame_shape: tuple[int, int],
+        mouse_pitch_sign: int = +1,
+    ) -> None:
+        self.cal = calibration
+        self.H, self.W = frame_shape
+        self.mouse_pitch_sign = int(mouse_pitch_sign)
+
+        f = self.cal.focal_length
+        cx, cy = self.W / 2.0, self.H / 2.0
+        xs = np.arange(self.W, dtype=np.float64) - cx
+        ys = np.arange(self.H, dtype=np.float64) - cy
+        self.x_grid, self.y_grid = np.meshgrid(xs, ys)  # H,W
+
+        # Pre-compute LH rotation kernels (depend only on pixel coords + f).
+        self.xy_over_f = self.x_grid * self.y_grid / f
+        self.f_plus_x2_over_f = f + self.x_grid**2 / f
+        self.f_plus_y2_over_f = f + self.y_grid**2 / f
+
+    @staticmethod
+    def _action_to_kinematics(
+        keyboard: np.ndarray,
+        mouse: np.ndarray,
+        cal: ThirdPersonCalibration,
+        mouse_pitch_sign: int,
+    ) -> tuple[np.ndarray, np.ndarray]:
+        """Return (omega (3,), T_total (3,)) for one frame's action."""
+        kb = np.asarray(keyboard, dtype=np.float64).reshape(-1)
+        mo = np.asarray(mouse, dtype=np.float64).reshape(-1)
+
+        pitch = mo[0] * mouse_pitch_sign
+        yaw = mo[1]
+
+        omega_x = cal.alpha_pitch * pitch
+        omega_y = cal.alpha_yaw * yaw
+        if kb.shape[0] >= 6:
+            omega_y += cal.alpha_turn * (kb[5] - kb[4])
+        omega = np.array([omega_x, omega_y, 0.0])
+
+        T_avatar = np.array([
+            cal.beta_strafe * (kb[3] - kb[2]),
+            0.0,
+            cal.beta_fwd * (kb[0] - kb[1]),
+        ])
+
+        # T_orbit = -(omega x r) with r = (0, r_y, r_z).
+        # cross([wx,wy,0], [0,ry,rz]) = (wy*rz, -wx*rz, wx*ry)
+        T_orbit = -np.array([
+            omega[1] * cal.r_z,
+            -omega[0] * cal.r_z,
+            omega[0] * cal.r_y,
+        ])
+        return omega, T_avatar + T_orbit
+
+    def _flow_from_kinematics(self, omega: np.ndarray, T: np.ndarray) -> np.ndarray:
+        """Compose rotation + translation flow at every pixel. Z=1."""
+        wx, wy, wz = omega
+        Tx, Ty, Tz = T
+        f = self.cal.focal_length
+
+        u_R = self.xy_over_f * wx - self.f_plus_x2_over_f * wy + self.y_grid * wz
+        v_R = self.f_plus_y2_over_f * wx - self.xy_over_f * wy - self.x_grid * wz
+        u_T = -f * Tx + self.x_grid * Tz
+        v_T = -f * Ty + self.y_grid * Tz
+        return np.stack([u_R + u_T, v_R + v_T], axis=-1).astype(np.float32)
+
+    def generate_flow(
+        self,
+        keyboard: np.ndarray,
+        mouse: np.ndarray,
+    ) -> np.ndarray:
+        """Synthesize HxWx2 flow for one frame's action."""
+        omega, T = self._action_to_kinematics(
+            keyboard,
+            mouse,
+            self.cal,
+            self.mouse_pitch_sign,
+        )
+        return self._flow_from_kinematics(omega, T)
+
+    def generate_flow_sequence(
+        self,
+        actions: dict,
+        n_pairs: int | None = None,
+    ) -> list[np.ndarray]:
+        """Generate flow for each consecutive frame pair.
+
+        Returns ``n`` flows where ``n = n_pairs`` if supplied, else
+        ``len(keyboard) - 1``.
+        """
+        kb = actions["keyboard"]
+        mo = actions["mouse"]
+        T = len(kb)
+        n = (T - 1) if n_pairs is None else min(n_pairs, T - 1)
+        return [self.generate_flow(kb[i], mo[i]) for i in range(n)]
+
+
+# ---------------------------------------------------------------------------
+# Linear-features form (used by the calibration fitter).
+# ---------------------------------------------------------------------------
+
+# Number of free params we fit. Order is fixed and shared with the fitter.
+PARAM_NAMES = (
+    "alpha_yaw",
+    "alpha_pitch",
+    "alpha_turn",
+    "beta_fwd",
+    "beta_strafe",
+    "focal_length",
+    "r_z",
+    "r_y",
+)
+
+
+def predict_flow_at_pixels(
+    keyboard: np.ndarray,  # (6,)
+    mouse: np.ndarray,  # (2,)
+    xs_centered: np.ndarray,  # (N,) pixel x relative to principal point
+    ys_centered: np.ndarray,  # (N,)
+    cal: ThirdPersonCalibration,
+    mouse_pitch_sign: int,
+) -> np.ndarray:
+    """Vectorized flow prediction at an arbitrary pixel set. Returns (N, 2).
+
+    Stateless: assumes camera is level (theta_pitch=0). For 3P games where
+    the camera tilts independently of the avatar's facing, use
+    :func:`predict_flow_at_pixels_stateful` and pass the integrated pitch.
+    """
+    omega, T = ThirdPersonFlowGenerator._action_to_kinematics(
+        keyboard,
+        mouse,
+        cal,
+        mouse_pitch_sign,
+    )
+    wx, wy, wz = omega
+    Tx, Ty, Tz = T
+    f = cal.focal_length
+
+    xy_over_f = xs_centered * ys_centered / f
+    f_plus_x2_over_f = f + xs_centered**2 / f
+    f_plus_y2_over_f = f + ys_centered**2 / f
+
+    u_R = xy_over_f * wx - f_plus_x2_over_f * wy + ys_centered * wz
+    v_R = f_plus_y2_over_f * wx - xy_over_f * wy - xs_centered * wz
+    u_T = -f * Tx + xs_centered * Tz
+    v_T = -f * Ty + ys_centered * Tz
+    return np.stack([u_R + u_T, v_R + v_T], axis=-1)
+
+
+def predict_flow_at_pixels_stateful(
+    keyboard: np.ndarray,
+    mouse: np.ndarray,
+    xs_centered: np.ndarray,
+    ys_centered: np.ndarray,
+    cal: ThirdPersonCalibration,
+    theta_pitch: float,
+) -> np.ndarray:
+    """Stateful flow prediction that accounts for accumulated camera pitch.
+
+    In a 3P game the avatar moves in the world's horizontal plane in the
+    direction the camera is yawed. When the camera is also pitched
+    (looking down at the avatar / up at the sky), this world-horizontal
+    motion has a non-zero y component in the camera frame. Concretely:
+
+        T_cam = β_fwd · (W − S) · (0, sin θ_pitch, cos θ_pitch)
+              + β_strafe · (D − A) · (1, 0, 0)              # strafe is pitch-invariant
+
+    Yaw doesn't appear because the camera frame is yaw-aligned by construction
+    (avatar and camera yaw together). Mouse rotation contributions to ω are
+    unchanged.
+
+    Parameters
+    ----------
+    theta_pitch : float (radians)
+        Accumulated camera pitch state at this frame, integrated from
+        prior mouse-pitch input. Positive = camera looking up.
+    """
+    f = cal.focal_length
+    cos_p = float(np.cos(theta_pitch))
+    sin_p = float(np.sin(theta_pitch))
+
+    # Rotational velocities (unchanged from stateless)
+    pitch_in = float(mouse[0])
+    yaw_in = float(mouse[1])
+    omega_x = cal.alpha_pitch * pitch_in
+    omega_y = cal.alpha_yaw * yaw_in
+    if keyboard.shape[0] >= 6:
+        omega_y += cal.alpha_turn * (keyboard[5] - keyboard[4])
+
+    # Avatar-frame translations (in world-horizontal plane, avatar-yaw-aligned)
+    avatar_strafe = cal.beta_strafe * (keyboard[3] - keyboard[2])
+    avatar_fwd = cal.beta_fwd * (keyboard[0] - keyboard[1])
+
+    # Map to camera frame using current pitch
+    Tx = avatar_strafe
+    Ty = avatar_fwd * sin_p
+    Tz = avatar_fwd * cos_p
+
+    xy_over_f = xs_centered * ys_centered / f
+    f_plus_x2_over_f = f + xs_centered**2 / f
+    f_plus_y2_over_f = f + ys_centered**2 / f
+
+    u_R = xy_over_f * omega_x - f_plus_x2_over_f * omega_y
+    v_R = f_plus_y2_over_f * omega_x - xy_over_f * omega_y
+    u_T = -f * Tx + xs_centered * Tz
+    v_T = -f * Ty + ys_centered * Tz
+    return np.stack([u_R + u_T, v_R + v_T], axis=-1)
+
+
+def integrate_pitch_state(
+    mouse: np.ndarray,  # (T, 2) raw or cached actions
+    cal: ThirdPersonCalibration,
+    *,
+    init_pitch: float = 0.0,
+    frames_per_step: int = 1,
+) -> np.ndarray:
+    """Integrate per-frame mouse-pitch input into accumulated camera pitch.
+
+    Returns a (T,) array where ``out[t]`` is the camera's accumulated pitch
+    angle (radians) AT THE START of frame ``t`` — i.e. the pose under which
+    frame ``t``'s action is interpreted.
+
+    For raw-frame action sequences pass ``frames_per_step=1``. For cached
+    actions where each sample represents N raw frames of integration
+    (cache stride = N), pass ``frames_per_step=N``.
+
+    NOTE: this only models the explicit user-input pitch. Cinematic
+    auto-pitch (camera tilting to track the avatar over uneven terrain)
+    isn't in the action stream and isn't captured here. For that you need
+    visual odometry (Option B / WorldCam-style ViPE pipeline).
+    """
+    per_step = cal.alpha_pitch * np.asarray(mouse[:, 0], dtype=np.float64) * frames_per_step
+    cum = np.cumsum(per_step)
+    # out[t] = pose BEFORE frame t's input is applied → shift by one
+    return init_pitch + np.concatenate([[0.0], cum[:-1]]).astype(np.float64)

fastvideo/eval/metrics/optical_flow/synthetic_optical_flow/metric.py

@@ -0,0 +1,169 @@
+"""Compare optical flow extracted from a generated video against optical
+flow synthesized analytically from per-frame actions.
+
+The reference flow is *not* observed from a ground-truth video — it's
+predicted from the action stream via a third-person camera-kinematics
+model (Longuet-Higgins linearization + off-pivot translation correction;
+no depth). Observed flow comes from the same ``ptlflow`` model used by
+``gt_optical_flow``, and the two are compared with the identical metric
+set, so scores are directly comparable across the two metrics.
+
+Required sample keys
+--------------------
+``video``
+    ``(B, T, C, H, W)`` float in ``[0, 1]``.
+``actions``
+    ``dict`` (or list-of-dicts of length B) with two ``np.ndarray`` keys:
+
+    * ``keyboard`` of shape ``(T, 6)`` — ``[W, S, A, D, turn_left, turn_right]``
+    * ``mouse`` of shape ``(T, 2)`` — ``[pitch, yaw]``
+``calibration``
+    Either a path to a ``ThirdPersonCalibration`` JSON file, or a dict
+    of fitted parameters. May also be set once at construction time via
+    ``calibration_path=`` and reused across samples.
+
+Optional sample keys
+--------------------
+``mouse_pitch_sign``
+    ``+1`` (default) or ``-1`` if the dataset's mouse-pitch sign is
+    flipped (mhuo's data carries this in metadata).
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.metrics.optical_flow._shared import (
+    aggregate_temporal,
+    compute_frame_metrics,
+    extract_video_flows,
+    load_ptlflow_model,
+)
+from fastvideo.eval.metrics.optical_flow.synthetic_optical_flow._thirdperson import (
+    ThirdPersonCalibration,
+    ThirdPersonFlowGenerator,
+    load_calibration,
+)
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+def _resolve_calibration(obj: str | Path | dict | ThirdPersonCalibration, ) -> ThirdPersonCalibration:
+    if isinstance(obj, ThirdPersonCalibration):
+        return obj
+    if isinstance(obj, dict):
+        return ThirdPersonCalibration.from_dict(obj)
+    return load_calibration(obj)
+
+
+@register("optical_flow.synthetic_optical_flow")
+class SyntheticOpticalFlowMetric(BaseMetric):
+    """Action-driven synthetic flow vs. video-extracted observed flow.
+
+    Pass ``calibration_path`` at construction to bind the calibration
+    once across all samples; otherwise supply ``sample["calibration"]``
+    per call. Missing actions or calibration produce a skipped result
+    (``score=None``) rather than raising.
+    """
+
+    name = "optical_flow.synthetic_optical_flow"
+    requires_reference = False
+    higher_is_better = False
+    needs_gpu = True
+    backbone = "optical_flow"
+    dependencies = ["ptlflow"]
+
+    def __init__(
+        self,
+        model_name: str = "dpflow",
+        ckpt: str = "things",
+        calibration_path: str | Path | None = None,
+        min_mag: float = 0.5,
+        max_mag_pct: float = 80.0,
+        grid_size: int = 8,
+    ) -> None:
+        super().__init__()
+        self.model_name = model_name
+        self.ckpt = ckpt
+        self.min_mag = min_mag
+        self.max_mag_pct = max_mag_pct
+        self.grid_size = grid_size
+        self._calibration: ThirdPersonCalibration | None = (_resolve_calibration(calibration_path)
+                                                            if calibration_path else None)
+        self._model = None
+        # See gt_optical_flow note: 1 frame pair per DPFlow forward.
+        # Batching at 1080p OOMs because the cost volume is ~4 GB/pair.
+        self._chunk_size = 1
+
+    def to(self, device: str | torch.device) -> SyntheticOpticalFlowMetric:
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        self._model = load_ptlflow_model(self.model_name, self.ckpt, self.device)
+
+    def compute(self, sample: dict) -> MetricResult:
+        if self._model is None:
+            self.setup()
+
+        actions = sample.get("actions")
+        if actions is None:
+            return self._skip(sample, "missing 'actions' (keyboard + mouse)")
+
+        cal_obj = sample.get("calibration")
+        cal = self._calibration if cal_obj is None else _resolve_calibration(cal_obj)
+        if cal is None:
+            return self._skip(
+                sample,
+                "missing 'calibration' (pass calibration_path= at construction "
+                "or sample['calibration'] per call)",
+            )
+
+        video = sample["video"].float()  # (T, C, H, W)
+        T, _, H, W = video.shape
+        if T < 2:
+            raise ValueError("Need at least 2 frames to compute optical flow")
+        n_pairs = T - 1
+
+        mouse_pitch_sign = int(sample.get("mouse_pitch_sign", 1))
+        chunk = self._chunk_size or 16
+
+        observed = extract_video_flows(
+            self._model,
+            video,
+            chunk=chunk,
+            device=self.device,
+        )
+        predictor = ThirdPersonFlowGenerator(
+            calibration=cal,
+            frame_shape=(H, W),
+            mouse_pitch_sign=mouse_pitch_sign,
+        )
+        predicted = predictor.generate_flow_sequence(actions, n_pairs=n_pairs)
+
+        n = min(len(observed), len(predicted))
+        per_frame = [
+            compute_frame_metrics(
+                predicted[i],
+                observed[i],
+                grid_size=self.grid_size,
+                min_mag=self.min_mag,
+                max_mag_pct=self.max_mag_pct,
+            ) for i in range(n)
+        ]
+        summary = aggregate_temporal(per_frame)
+        score = summary.get("pixel_epe_mean_mean")
+        details = dict(summary)
+        details["per_frame_metrics"] = per_frame
+        return MetricResult(
+            name=self.name,
+            score=float(score) if score is not None else None,
+            details=details,
+        )

fastvideo/eval/metrics/physics_iq/_vendored/LICENSE

@@ -0,0 +1,19 @@
+descriptions.csv is vendored from the upstream Physics-IQ benchmark
+(https://github.com/google-deepmind/physics-IQ-benchmark, file
+descriptions/descriptions.csv) without modification. The associated video,
+mask, and switch-frame assets are NOT vendored — they auto-fetch on first
+use from gs://physics-iq-benchmark (public bucket; HTTPS-readable at
+https://storage.googleapis.com/physics-iq-benchmark/).
+
+Licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0).
+See https://creativecommons.org/licenses/by/4.0/legalcode for the full
+license text.
+
+Citation:
+    @article{motamed2025physics,
+      title={Do generative video models understand physical principles?},
+      author={Saman Motamed and Laura Culp and Kevin Swersky and
+              Priyank Jaini and Robert Geirhos},
+      journal={arXiv preprint arXiv:2501.09038},
+      year={2025}
+    }

fastvideo/eval/metrics/physics_iq/_vendored/descriptions.csv

@@ -0,0 +1,397 @@
+scenario,description,category,generated_video_name
+0001_perspective-left_take-1_trimmed-ball-and-block-fall.mp4,Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The grabber tools let go of the ball and block. Static shot with no camera movement.,Solid Mechanics,0001_perspective-left_trimmed-ball-and-block-fall.mp4
+0002_perspective-center_take-1_trimmed-ball-and-block-fall.mp4,Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The grabber tools let go of the ball and block. Static shot with no camera movement.,Solid Mechanics,0002_perspective-center_trimmed-ball-and-block-fall.mp4
+0003_perspective-right_take-1_trimmed-ball-and-block-fall.mp4,Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The grabber tools let go of the ball and block. Static shot with no camera movement.,Solid Mechanics,0003_perspective-right_trimmed-ball-and-block-fall.mp4
+0004_perspective-left_take-1_trimmed-ball-behind-rotating-paper.mp4,A grabber arm is holding a tennis ball above a piece of cardstock propped up on a rotating platform sitting on a table that rotates clockwise. The grabber lowers the ball and places is on the table as the cardstock rotates. Static shot with no camera movement.,Solid Mechanics,0004_perspective-left_trimmed-ball-behind-rotating-paper.mp4
+0005_perspective-center_take-1_trimmed-ball-behind-rotating-paper.mp4,A grabber arm is holding a tennis ball above a piece of cardstock propped up on a rotating platform sitting on a table that rotates clockwise. The grabber lowers the ball and places is on the table as the cardstock rotates. Static shot with no camera movement.,Solid Mechanics,0005_perspective-center_trimmed-ball-behind-rotating-paper.mp4
+0006_perspective-right_take-1_trimmed-ball-behind-rotating-paper.mp4,A grabber arm is holding a tennis ball above a piece of cardstock propped up on a rotating platform sitting on a table that rotates clockwise. The grabber lowers the ball and places is on the table as the cardstock rotates. Static shot with no camera movement.,Solid Mechanics,0006_perspective-right_trimmed-ball-behind-rotating-paper.mp4
+0007_perspective-left_take-1_trimmed-ball-hits-duck.mp4,A light beige coffee table with a small yellow rubber ducky on it. A mustard yellow couch is in the background. There is a black pipe on one end of the table and a brown tennis ball rolls out of it towards the rubber ducky. Static shot with no camera movement.,Solid Mechanics,0007_perspective-left_trimmed-ball-hits-duck.mp4
+0008_perspective-center_take-1_trimmed-ball-hits-duck.mp4,A light beige coffee table with a small yellow rubber ducky on it. A mustard yellow couch is in the background. There is a black pipe on one end of the table and a brown tennis ball rolls out of it towards the rubber ducky. Static shot with no camera movement.,Solid Mechanics,0008_perspective-center_trimmed-ball-hits-duck.mp4
+0009_perspective-right_take-1_trimmed-ball-hits-duck.mp4,A light beige coffee table with a small yellow rubber ducky on it. A mustard yellow couch is in the background. There is a black pipe on one end of the table and a brown tennis ball rolls out of it towards the rubber ducky. Static shot with no camera movement.,Solid Mechanics,0009_perspective-right_trimmed-ball-hits-duck.mp4
+0010_perspective-left_take-1_trimmed-ball-hits-nothing.mp4,A light-colored wooden coffee table with a few small objects on it including a tennis ball and a smaller red ball. An orange ball rolls out of a black pipe that is sitting on the table towards the right side. Static shot with no camera movement.,Solid Mechanics,0010_perspective-left_trimmed-ball-hits-nothing.mp4
+0011_perspective-center_take-1_trimmed-ball-hits-nothing.mp4,A light-colored wooden coffee table with a few small objects on it including a tennis ball and a smaller red ball. An orange ball rolls out of a black pipe that is sitting on the table towards the right side. Static shot with no camera movement.,Solid Mechanics,0011_perspective-center_trimmed-ball-hits-nothing.mp4
+0012_perspective-right_take-1_trimmed-ball-hits-nothing.mp4,A light-colored wooden coffee table with a few small objects on it including a tennis ball and a smaller red ball. An orange ball rolls out of a black pipe that is sitting on the table towards the right side. Static shot with no camera movement.,Solid Mechanics,0012_perspective-right_trimmed-ball-hits-nothing.mp4
+0013_perspective-left_take-1_trimmed-ball-in-basket.mp4,An orange inflatable basketball is suspended above a black plastic crate placed on a wooden table. The ball is then released. Static shot with no camera movement.,Solid Mechanics,0013_perspective-left_trimmed-ball-in-basket.mp4
+0014_perspective-center_take-1_trimmed-ball-in-basket.mp4,An orange inflatable basketball is suspended above a black plastic crate placed on a wooden table. The ball is then released. Static shot with no camera movement.,Solid Mechanics,0014_perspective-center_trimmed-ball-in-basket.mp4
+0015_perspective-right_take-1_trimmed-ball-in-basket.mp4,An orange inflatable basketball is suspended above a black plastic crate placed on a wooden table. The ball is then released. Static shot with no camera movement.,Solid Mechanics,0015_perspective-right_trimmed-ball-in-basket.mp4
+0016_perspective-left_take-1_trimmed-ball-in-sand.mp4,A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball. Static shot with no camera movement.,Solid Mechanics,0016_perspective-left_trimmed-ball-in-sand.mp4
+0017_perspective-center_take-1_trimmed-ball-in-sand.mp4,A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball. Static shot with no camera movement.,Solid Mechanics,0017_perspective-center_trimmed-ball-in-sand.mp4
+0018_perspective-right_take-1_trimmed-ball-in-sand.mp4,A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball. Static shot with no camera movement.,Solid Mechanics,0018_perspective-right_trimmed-ball-in-sand.mp4
+0019_perspective-left_take-1_trimmed-ball-ramp.mp4,A simple ramp made of cardboard propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame and a yellow tennis ball rolls out of the pipe towards the ramp. Static shot with no camera movement.,Solid Mechanics,0019_perspective-left_trimmed-ball-ramp.mp4
+0020_perspective-center_take-1_trimmed-ball-ramp.mp4,A simple ramp made of cardboard propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame and a yellow tennis ball rolls out of the pipe towards the ramp. Static shot with no camera movement.,Solid Mechanics,0020_perspective-center_trimmed-ball-ramp.mp4
+0021_perspective-right_take-1_trimmed-ball-ramp.mp4,A simple ramp made of cardboard propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame and a yellow tennis ball rolls out of the pipe towards the ramp. Static shot with no camera movement.,Solid Mechanics,0021_perspective-right_trimmed-ball-ramp.mp4
+0022_perspective-left_take-1_trimmed-ball-rolls-off.mp4,A light wood coffee table in the foreground with a black pipe on the end of the table. A grey tennis ball rolls out of the pipe towards the right and onto the table. Static shot with no camera movement.,Solid Mechanics,0022_perspective-left_trimmed-ball-rolls-off.mp4
+0023_perspective-center_take-1_trimmed-ball-rolls-off.mp4,A light wood coffee table in the foreground with a black pipe on the end of the table. A grey tennis ball rolls out of the pipe towards the right and onto the table. Static shot with no camera movement.,Solid Mechanics,0023_perspective-center_trimmed-ball-rolls-off.mp4
+0024_perspective-right_take-1_trimmed-ball-rolls-off.mp4,A light wood coffee table in the foreground with a black pipe on the end of the table. A grey tennis ball rolls out of the pipe towards the right and onto the table. Static shot with no camera movement.,Solid Mechanics,0024_perspective-right_trimmed-ball-rolls-off.mp4
+0025_perspective-left_take-1_trimmed-ball-rolls-on-glass.mp4,A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. A blue tennis ball rolls on the wooden table and towards the glass. Static shot with no camera movement.,Solid Mechanics,0025_perspective-left_trimmed-ball-rolls-on-glass.mp4
+0026_perspective-center_take-1_trimmed-ball-rolls-on-glass.mp4,A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. A blue tennis ball rolls on the wooden table and towards the glass. Static shot with no camera movement.,Solid Mechanics,0026_perspective-center_trimmed-ball-rolls-on-glass.mp4
+0027_perspective-right_take-1_trimmed-ball-rolls-on-glass.mp4,A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. A blue tennis ball rolls on the wooden table and towards the glass. Static shot with no camera movement.,Solid Mechanics,0027_perspective-right_trimmed-ball-rolls-on-glass.mp4
+0028_perspective-left_take-1_trimmed-ball-train.mp4,"A light-colored coffee table with two tennis balls, one orange and one brown, placed near the center back to back. A grey tennis ball rolls out of a black pipe sitting on the table and towards the other two balls. Static shot with no camera movement.",Solid Mechanics,0028_perspective-left_trimmed-ball-train.mp4
+0029_perspective-center_take-1_trimmed-ball-train.mp4,"A light-colored coffee table with two tennis balls, one orange and one brown, placed near the center back to back. A grey tennis ball rolls out of a black pipe sitting on the table and towards the other two balls. Static shot with no camera movement.",Solid Mechanics,0029_perspective-center_trimmed-ball-train.mp4
+0030_perspective-right_take-1_trimmed-ball-train.mp4,"A light-colored coffee table with two tennis balls, one orange and one brown, placed near the center back to back. A grey tennis ball rolls out of a black pipe sitting on the table and towards the other two balls. Static shot with no camera movement.",Solid Mechanics,0030_perspective-right_trimmed-ball-train.mp4
+0031_perspective-left_take-1_trimmed-balls-collide.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0031_perspective-left_trimmed-balls-collide.mp4
+0032_perspective-center_take-1_trimmed-balls-collide.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0032_perspective-center_trimmed-balls-collide.mp4
+0033_perspective-right_take-1_trimmed-balls-collide.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0033_perspective-right_trimmed-balls-collide.mp4
+0034_perspective-left_take-1_trimmed-block-domino.mp4,A row of colorful wooden blocks lined up on a wooden table with a wooden stick attached to a black rotating platform. The platform rotates clockwise and the wooden stick hits the first block as it rotates. Static shot with no camera movement.,Solid Mechanics,0034_perspective-left_trimmed-block-domino.mp4
+0035_perspective-center_take-1_trimmed-block-domino.mp4,A row of colorful wooden blocks lined up on a wooden table with a wooden stick attached to a black rotating platform. The platform rotates clockwise and the wooden stick hits the first block as it rotates. Static shot with no camera movement.,Solid Mechanics,0035_perspective-center_trimmed-block-domino.mp4
+0036_perspective-right_take-1_trimmed-block-domino.mp4,A row of colorful wooden blocks lined up on a wooden table with a wooden stick attached to a black rotating platform. The platform rotates clockwise and the wooden stick hits the first block as it rotates. Static shot with no camera movement.,Solid Mechanics,0036_perspective-right_trimmed-block-domino.mp4
+0037_perspective-left_take-1_trimmed-blow-balloon.mp4,A black balloon is attached to a fixed electric air pump hose on a table with a plain wall in the background. Air is being pumped in the balloon. Static shot with no camera movement.,Fluid Dynamics,0037_perspective-left_trimmed-blow-balloon.mp4
+0038_perspective-center_take-1_trimmed-blow-balloon.mp4,A black balloon is attached to a fixed electric air pump hose on a table with a plain wall in the background. Air is being pumped in the balloon. Static shot with no camera movement.,Fluid Dynamics,0038_perspective-center_trimmed-blow-balloon.mp4
+0039_perspective-right_take-1_trimmed-blow-balloon.mp4,A black balloon is attached to a fixed electric air pump hose on a table with a plain wall in the background. Air is being pumped in the balloon. Static shot with no camera movement.,Fluid Dynamics,0039_perspective-right_trimmed-blow-balloon.mp4
+0040_perspective-left_take-1_trimmed-cut-orange.mp4,A tangerine that has been cut in half is placed on a glass cutting board. A knife is slicing through the tangerine. Static shot with no camera movement.,Solid Mechanics,0040_perspective-left_trimmed-cut-orange.mp4
+0041_perspective-center_take-1_trimmed-cut-orange.mp4,A tangerine that has been cut in half is placed on a glass cutting board. A knife is slicing through the tangerine. Static shot with no camera movement.,Solid Mechanics,0041_perspective-center_trimmed-cut-orange.mp4
+0042_perspective-right_take-1_trimmed-cut-orange.mp4,A tangerine that has been cut in half is placed on a glass cutting board. A knife is slicing through the tangerine. Static shot with no camera movement.,Solid Mechanics,0042_perspective-right_trimmed-cut-orange.mp4
+0043_perspective-left_take-1_trimmed-cut-paper.mp4,"Two black and blue gripping tools are pulling a piece of green paper from its two corners, causing it to tear. Static shot with no camera movement.",Solid Mechanics,0043_perspective-left_trimmed-cut-paper.mp4
+0044_perspective-center_take-1_trimmed-cut-paper.mp4,"Two black and blue gripping tools are pulling a piece of green paper from its two corners, causing it to tear. Static shot with no camera movement.",Solid Mechanics,0044_perspective-center_trimmed-cut-paper.mp4
+0045_perspective-right_take-1_trimmed-cut-paper.mp4,"Two black and blue gripping tools are pulling a piece of green paper from its two corners, causing it to tear. Static shot with no camera movement.",Solid Mechanics,0045_perspective-right_trimmed-cut-paper.mp4
+0046_perspective-left_take-1_trimmed-domino-in-juice.mp4,A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.,Fluid Dynamics,0046_perspective-left_trimmed-domino-in-juice.mp4
+0047_perspective-center_take-1_trimmed-domino-in-juice.mp4,A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.,Fluid Dynamics,0047_perspective-center_trimmed-domino-in-juice.mp4
+0048_perspective-right_take-1_trimmed-domino-in-juice.mp4,A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.,Fluid Dynamics,0048_perspective-right_trimmed-domino-in-juice.mp4
+0049_perspective-left_take-1_trimmed-dominos-with-space.mp4,Two rows of alternating black and white dominoes are set up on a wooden table with a gap between the two rows. A wooden stick attached to a rotating platform rotates clockwise and knocks the first domino in the first row. Static shot with no camera movement.,Solid Mechanics,0049_perspective-left_trimmed-dominos-with-space.mp4
+0050_perspective-center_take-1_trimmed-dominos-with-space.mp4,Two rows of alternating black and white dominoes are set up on a wooden table with a gap between the two rows. A wooden stick attached to a rotating platform rotates clockwise and knocks the first domino in the first row. Static shot with no camera movement.,Solid Mechanics,0050_perspective-center_trimmed-dominos-with-space.mp4
+0051_perspective-right_take-1_trimmed-dominos-with-space.mp4,Two rows of alternating black and white dominoes are set up on a wooden table with a gap between the two rows. A wooden stick attached to a rotating platform rotates clockwise and knocks the first domino in the first row. Static shot with no camera movement.,Solid Mechanics,0051_perspective-right_trimmed-dominos-with-space.mp4
+0052_perspective-left_take-1_trimmed-double-cradle.mp4,A Newton's cradle device on the table and two of the metal balls are held up by a blue handled grabber tool. The claw releases the two balls. Static shot with no camera movement.,Solid Mechanics,0052_perspective-left_trimmed-double-cradle.mp4
+0053_perspective-center_take-1_trimmed-double-cradle.mp4,A Newton's cradle device on the table and two of the metal balls are held up by a blue handled grabber tool. The claw releases the two balls. Static shot with no camera movement.,Solid Mechanics,0053_perspective-center_trimmed-double-cradle.mp4
+0054_perspective-right_take-1_trimmed-double-cradle.mp4,A Newton's cradle device on the table and two of the metal balls are held up by a blue handled grabber tool. The claw releases the two balls. Static shot with no camera movement.,Solid Mechanics,0054_perspective-right_trimmed-double-cradle.mp4
+0055_perspective-left_take-1_trimmed-duck-and-dominos.mp4,A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.,Solid Mechanics,0055_perspective-left_trimmed-duck-and-dominos.mp4
+0056_perspective-center_take-1_trimmed-duck-and-dominos.mp4,A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.,Solid Mechanics,0056_perspective-center_trimmed-duck-and-dominos.mp4
+0057_perspective-right_take-1_trimmed-duck-and-dominos.mp4,A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.,Solid Mechanics,0057_perspective-right_trimmed-duck-and-dominos.mp4
+0058_perspective-left_take-1_trimmed-duck-falls-in-box.mp4,A yellow rubber ducky is suspended above an open dark green fabric box on a wooden table. The duck is then released. Static shot with no camera movement.,Solid Mechanics,0058_perspective-left_trimmed-duck-falls-in-box.mp4
+0059_perspective-center_take-1_trimmed-duck-falls-in-box.mp4,A yellow rubber ducky is suspended above an open dark green fabric box on a wooden table. The duck is then released. Static shot with no camera movement.,Solid Mechanics,0059_perspective-center_trimmed-duck-falls-in-box.mp4
+0060_perspective-right_take-1_trimmed-duck-falls-in-box.mp4,A yellow rubber ducky is suspended above an open dark green fabric box on a wooden table. The duck is then released. Static shot with no camera movement.,Solid Mechanics,0060_perspective-right_trimmed-duck-falls-in-box.mp4
+0061_perspective-left_take-1_trimmed-duck-static.mp4,A stationary yellow rubber duck on a light brown wooden table against a plain white background. Static shot with no camera movement.,Solid Mechanics,0061_perspective-left_trimmed-duck-static.mp4
+0062_perspective-center_take-1_trimmed-duck-static.mp4,A stationary yellow rubber duck on a light brown wooden table against a plain white background. Static shot with no camera movement.,Solid Mechanics,0062_perspective-center_trimmed-duck-static.mp4
+0063_perspective-right_take-1_trimmed-duck-static.mp4,A stationary yellow rubber duck on a light brown wooden table against a plain white background. Static shot with no camera movement.,Solid Mechanics,0063_perspective-right_trimmed-duck-static.mp4
+0064_perspective-left_take-1_trimmed-fill-glass-red-drink.mp4,A glass beverage dispenser filled with a bright red liquid is set up on a woven basket and is pouring the liquid into a clear glass on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0064_perspective-left_trimmed-fill-glass-red-drink.mp4
+0065_perspective-center_take-1_trimmed-fill-glass-red-drink.mp4,A glass beverage dispenser filled with a bright red liquid is set up on a woven basket and is pouring the liquid into a clear glass on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0065_perspective-center_trimmed-fill-glass-red-drink.mp4
+0066_perspective-right_take-1_trimmed-fill-glass-red-drink.mp4,A glass beverage dispenser filled with a bright red liquid is set up on a woven basket and is pouring the liquid into a clear glass on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0066_perspective-right_trimmed-fill-glass-red-drink.mp4
+0067_perspective-left_take-1_trimmed-glass-stays-same.mp4,A glass beverage dispenser filled with a bright red liquid is set on a wicker base. Under the dispenser there is a glass half filled with red liquid. Static shot with no camera movement.,Fluid Dynamics,0067_perspective-left_trimmed-glass-stays-same.mp4
+0068_perspective-center_take-1_trimmed-glass-stays-same.mp4,A glass beverage dispenser filled with a bright red liquid is set on a wicker base. Under the dispenser there is a glass half filled with red liquid. Static shot with no camera movement.,Fluid Dynamics,0068_perspective-center_trimmed-glass-stays-same.mp4
+0069_perspective-right_take-1_trimmed-glass-stays-same.mp4,A glass beverage dispenser filled with a bright red liquid is set on a wicker base. Under the dispenser there is a glass half filled with red liquid. Static shot with no camera movement.,Fluid Dynamics,0069_perspective-right_trimmed-glass-stays-same.mp4
+0070_perspective-left_take-1_trimmed-juice-in-water.mp4,A glass beverage dispenser pouring grapefruit juice into a glass that has some water inside. Static shot with no camera movement.,Fluid Dynamics,0070_perspective-left_trimmed-juice-in-water.mp4
+0071_perspective-center_take-1_trimmed-juice-in-water.mp4,A glass beverage dispenser pouring grapefruit juice into a glass that has some water inside. Static shot with no camera movement.,Fluid Dynamics,0071_perspective-center_trimmed-juice-in-water.mp4
+0072_perspective-right_take-1_trimmed-juice-in-water.mp4,A glass beverage dispenser pouring grapefruit juice into a glass that has some water inside. Static shot with no camera movement.,Fluid Dynamics,0072_perspective-right_trimmed-juice-in-water.mp4
+0073_perspective-left_take-1_trimmed-light-on-block.mp4,A blue rectangular wooden block is placed on a black rotating turntable that rotates clockwise illuminated by a spotlight casting a long shadow on the wall behind it. Static shot with no camera movement.,Optics,0073_perspective-left_trimmed-light-on-block.mp4
+0074_perspective-center_take-1_trimmed-light-on-block.mp4,A blue rectangular wooden block is placed on a black rotating turntable that rotates clockwise illuminated by a spotlight casting a long shadow on the wall behind it. Static shot with no camera movement.,Optics,0074_perspective-center_trimmed-light-on-block.mp4
+0075_perspective-right_take-1_trimmed-light-on-block.mp4,A blue rectangular wooden block is placed on a black rotating turntable that rotates clockwise illuminated by a spotlight casting a long shadow on the wall behind it. Static shot with no camera movement.,Optics,0075_perspective-right_trimmed-light-on-block.mp4
+0076_perspective-left_take-1_trimmed-light-on-mug.mp4,A yellow mug is placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting a shadow on the wall behind it. Static shot with no camera movement.,Optics,0076_perspective-left_trimmed-light-on-mug.mp4
+0077_perspective-center_take-1_trimmed-light-on-mug.mp4,A yellow mug is placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting a shadow on the wall behind it. Static shot with no camera movement.,Optics,0077_perspective-center_trimmed-light-on-mug.mp4
+0078_perspective-right_take-1_trimmed-light-on-mug.mp4,A yellow mug is placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting a shadow on the wall behind it. Static shot with no camera movement.,Optics,0078_perspective-right_trimmed-light-on-mug.mp4
+0079_perspective-left_take-1_trimmed-light-on-mug-block.mp4,A yellow mug and a blue wooden block are placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting their shadow on the wall behind it. Static shot with no camera movement.,Optics,0079_perspective-left_trimmed-light-on-mug-block.mp4
+0080_perspective-center_take-1_trimmed-light-on-mug-block.mp4,A yellow mug and a blue wooden block are placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting their shadow on the wall behind it. Static shot with no camera movement.,Optics,0080_perspective-center_trimmed-light-on-mug-block.mp4
+0081_perspective-right_take-1_trimmed-light-on-mug-block.mp4,A yellow mug and a blue wooden block are placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting their shadow on the wall behind it. Static shot with no camera movement.,Optics,0081_perspective-right_trimmed-light-on-mug-block.mp4
+0082_perspective-left_take-1_trimmed-light-on-statue.mp4,A small statue made of porcelain illuminated by a spotlight on a rotating base that rotates clockwise. The spotlight casts a large shadow of the statue onto the wall behind it. Static shot with no camera movement.,Optics,0082_perspective-left_trimmed-light-on-statue.mp4
+0083_perspective-center_take-1_trimmed-light-on-statue.mp4,A small statue made of porcelain illuminated by a spotlight on a rotating base that rotates clockwise. The spotlight casts a large shadow of the statue onto the wall behind it. Static shot with no camera movement.,Optics,0083_perspective-center_trimmed-light-on-statue.mp4
+0084_perspective-right_take-1_trimmed-light-on-statue.mp4,A small statue made of porcelain illuminated by a spotlight on a rotating base that rotates clockwise. The spotlight casts a large shadow of the statue onto the wall behind it. Static shot with no camera movement.,Optics,0084_perspective-right_trimmed-light-on-statue.mp4
+0085_perspective-left_take-1_trimmed-liquid-on-duck.mp4,A yellow rubber ducky is placed in an empty black baking pan on a wooden table.  A beverage dispenser with red liquid inside sits on a woven basket behind it. The liquid pours on the duck from the dispenser. Static shot with no camera movement.,Fluid Dynamics,0085_perspective-left_trimmed-liquid-on-duck.mp4
+0086_perspective-center_take-1_trimmed-liquid-on-duck.mp4,A yellow rubber ducky is placed in an empty black baking pan on a wooden table.  A beverage dispenser with red liquid inside sits on a woven basket behind it. The liquid pours on the duck from the dispenser. Static shot with no camera movement.,Fluid Dynamics,0086_perspective-center_trimmed-liquid-on-duck.mp4
+0087_perspective-right_take-1_trimmed-liquid-on-duck.mp4,A yellow rubber ducky is placed in an empty black baking pan on a wooden table.  A beverage dispenser with red liquid inside sits on a woven basket behind it. The liquid pours on the duck from the dispenser. Static shot with no camera movement.,Fluid Dynamics,0087_perspective-right_trimmed-liquid-on-duck.mp4
+0088_perspective-left_take-1_trimmed-liquid-overfill.mp4,A bright red liquid being poured from a dispenser into a glass which is placed on a dark baking tray on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0088_perspective-left_trimmed-liquid-overfill.mp4
+0089_perspective-center_take-1_trimmed-liquid-overfill.mp4,A bright red liquid being poured from a dispenser into a glass which is placed on a dark baking tray on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0089_perspective-center_trimmed-liquid-overfill.mp4
+0090_perspective-right_take-1_trimmed-liquid-overfill.mp4,A bright red liquid being poured from a dispenser into a glass which is placed on a dark baking tray on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0090_perspective-right_trimmed-liquid-overfill.mp4
+0091_perspective-left_take-1_trimmed-lit-candle.mp4,Two candle holders that have tall red candles in them are placed on a wooden table. One of the candles is burning. Static shot with no camera movement.,Thermodynamics,0091_perspective-left_trimmed-lit-candle.mp4
+0092_perspective-center_take-1_trimmed-lit-candle.mp4,Two candle holders that have tall red candles in them are placed on a wooden table. One of the candles is burning. Static shot with no camera movement.,Thermodynamics,0092_perspective-center_trimmed-lit-candle.mp4
+0093_perspective-right_take-1_trimmed-lit-candle.mp4,Two candle holders that have tall red candles in them are placed on a wooden table. One of the candles is burning. Static shot with no camera movement.,Thermodynamics,0093_perspective-right_trimmed-lit-candle.mp4
+0094_perspective-left_take-1_trimmed-magnet-domino.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small white plastic domino block is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0094_perspective-left_trimmed-magnet-domino.mp4
+0095_perspective-center_take-1_trimmed-magnet-domino.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small white plastic domino block is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0095_perspective-center_trimmed-magnet-domino.mp4
+0096_perspective-right_take-1_trimmed-magnet-domino.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small white plastic domino block is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0096_perspective-right_trimmed-magnet-domino.mp4
+0097_perspective-left_take-1_trimmed-magnet-transparent-peakaboo.mp4,A clear acrylic box suspended from a cord hangs above a tennis ball with a smiley face drawn on it positioned on a wooden table. The box is lowered to cover the ball. Static shot with no camera movement.,Solid Mechanics,0097_perspective-left_trimmed-magnet-transparent-peakaboo.mp4
+0098_perspective-center_take-1_trimmed-magnet-transparent-peakaboo.mp4,A clear acrylic box suspended from a cord hangs above a tennis ball with a smiley face drawn on it positioned on a wooden table. The box is lowered to cover the ball. Static shot with no camera movement.,Solid Mechanics,0098_perspective-center_trimmed-magnet-transparent-peakaboo.mp4
+0099_perspective-right_take-1_trimmed-magnet-transparent-peakaboo.mp4,A clear acrylic box suspended from a cord hangs above a tennis ball with a smiley face drawn on it positioned on a wooden table. The box is lowered to cover the ball. Static shot with no camera movement.,Solid Mechanics,0099_perspective-right_trimmed-magnet-transparent-peakaboo.mp4
+0100_perspective-left_take-1_trimmed-magnet-wrench.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small metal wrench is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0100_perspective-left_trimmed-magnet-wrench.mp4
+0101_perspective-center_take-1_trimmed-magnet-wrench.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small metal wrench is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0101_perspective-center_trimmed-magnet-wrench.mp4
+0102_perspective-right_take-1_trimmed-magnet-wrench.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small metal wrench is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0102_perspective-right_trimmed-magnet-wrench.mp4
+0103_perspective-left_take-1_trimmed-marble-run-x.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0103_perspective-left_trimmed-marble-run-x.mp4
+0104_perspective-center_take-1_trimmed-marble-run-x.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0104_perspective-center_trimmed-marble-run-x.mp4
+0105_perspective-right_take-1_trimmed-marble-run-x.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0105_perspective-right_trimmed-marble-run-x.mp4
+0106_perspective-left_take-1_trimmed-marble-run-y.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0106_perspective-left_trimmed-marble-run-y.mp4
+0107_perspective-center_take-1_trimmed-marble-run-y.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0107_perspective-center_trimmed-marble-run-y.mp4
+0108_perspective-right_take-1_trimmed-marble-run-y.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0108_perspective-right_trimmed-marble-run-y.mp4
+0109_perspective-left_take-1_trimmed-match.mp4,A lit match is being lowered into a glass of water. Static shot with no camera movement.,Fluid Dynamics,0109_perspective-left_trimmed-match.mp4
+0110_perspective-center_take-1_trimmed-match.mp4,A lit match is being lowered into a glass of water. Static shot with no camera movement.,Fluid Dynamics,0110_perspective-center_trimmed-match.mp4
+0111_perspective-right_take-1_trimmed-match.mp4,A lit match is being lowered into a glass of water. Static shot with no camera movement.,Fluid Dynamics,0111_perspective-right_trimmed-match.mp4
+0112_perspective-left_take-1_trimmed-match-blows-balloon.mp4,A black balloon is sitting on a wooden table next to a small rotating platform with a lit matchstick taped to it. The match rotates clockwise and touches the balloon. Static shot with no camera movement.,Thermodynamics,0112_perspective-left_trimmed-match-blows-balloon.mp4
+0113_perspective-center_take-1_trimmed-match-blows-balloon.mp4,A black balloon is sitting on a wooden table next to a small rotating platform with a lit matchstick taped to it. The match rotates clockwise and touches the balloon. Static shot with no camera movement.,Thermodynamics,0113_perspective-center_trimmed-match-blows-balloon.mp4
+0114_perspective-right_take-1_trimmed-match-blows-balloon.mp4,A black balloon is sitting on a wooden table next to a small rotating platform with a lit matchstick taped to it. The match rotates clockwise and touches the balloon. Static shot with no camera movement.,Thermodynamics,0114_perspective-right_trimmed-match-blows-balloon.mp4
+0115_perspective-left_take-1_trimmed-mirror-ball-fall.mp4,A tennis ball attached to a magnet and string is hanging in front of a mirror and creating an illusion of two tennis balls. The string lowers the ball slowly. Static shot with no camera movement.,Optics,0115_perspective-left_trimmed-mirror-ball-fall.mp4
+0116_perspective-center_take-1_trimmed-mirror-ball-fall.mp4,A tennis ball attached to a magnet and string is hanging in front of a mirror and creating an illusion of two tennis balls. The string lowers the ball slowly. Static shot with no camera movement.,Optics,0116_perspective-center_trimmed-mirror-ball-fall.mp4
+0117_perspective-right_take-1_trimmed-mirror-ball-fall.mp4,A tennis ball attached to a magnet and string is hanging in front of a mirror and creating an illusion of two tennis balls. The string lowers the ball slowly. Static shot with no camera movement.,Optics,0117_perspective-right_trimmed-mirror-ball-fall.mp4
+0118_perspective-left_take-1_trimmed-mirror-ball-rotate.mp4,A tennis ball with a smiley face drawn on it is slowly rotating on a black rotating platform that rotates clockwise in front of a mirror and reflecting the side of the ball that has no smiley face. Static shot with no camera movement.,Optics,0118_perspective-left_trimmed-mirror-ball-rotate.mp4
+0119_perspective-center_take-1_trimmed-mirror-ball-rotate.mp4,A tennis ball with a smiley face drawn on it is slowly rotating on a black rotating platform that rotates clockwise in front of a mirror and reflecting the side of the ball that has no smiley face. Static shot with no camera movement.,Optics,0119_perspective-center_trimmed-mirror-ball-rotate.mp4
+0120_perspective-right_take-1_trimmed-mirror-ball-rotate.mp4,A tennis ball with a smiley face drawn on it is slowly rotating on a black rotating platform that rotates clockwise in front of a mirror and reflecting the side of the ball that has no smiley face. Static shot with no camera movement.,Optics,0120_perspective-right_trimmed-mirror-ball-rotate.mp4
+0121_perspective-left_take-1_trimmed-mirror-teapot-rotate.mp4,A teapot on a rotating display base that rotates clockwise in front of a mirror reflecting the teapot's image. Static shot with no camera movement.,Optics,0121_perspective-left_trimmed-mirror-teapot-rotate.mp4
+0122_perspective-center_take-1_trimmed-mirror-teapot-rotate.mp4,A teapot on a rotating display base that rotates clockwise in front of a mirror reflecting the teapot's image. Static shot with no camera movement.,Optics,0122_perspective-center_trimmed-mirror-teapot-rotate.mp4
+0123_perspective-right_take-1_trimmed-mirror-teapot-rotate.mp4,A teapot on a rotating display base that rotates clockwise in front of a mirror reflecting the teapot's image. Static shot with no camera movement.,Optics,0123_perspective-right_trimmed-mirror-teapot-rotate.mp4
+0124_perspective-left_take-1_trimmed-mug-breaks.mp4,A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug. Static shot with no camera movement.,Solid Mechanics,0124_perspective-left_trimmed-mug-breaks.mp4
+0125_perspective-center_take-1_trimmed-mug-breaks.mp4,A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug. Static shot with no camera movement.,Solid Mechanics,0125_perspective-center_trimmed-mug-breaks.mp4
+0126_perspective-right_take-1_trimmed-mug-breaks.mp4,A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug. Static shot with no camera movement.,Solid Mechanics,0126_perspective-right_trimmed-mug-breaks.mp4
+0127_perspective-left_take-1_trimmed-napkin-soak.mp4,A grabber tool holds a piece of paper towel over a shallow dish of light blue liquid on a wooden table. The grabber releases the paper towel on the dish. Static shot with no camera movement.,Fluid Dynamics,0127_perspective-left_trimmed-napkin-soak.mp4
+0128_perspective-center_take-1_trimmed-napkin-soak.mp4,A grabber tool holds a piece of paper towel over a shallow dish of light blue liquid on a wooden table. The grabber releases the paper towel on the dish. Static shot with no camera movement.,Fluid Dynamics,0128_perspective-center_trimmed-napkin-soak.mp4
+0129_perspective-right_take-1_trimmed-napkin-soak.mp4,A grabber tool holds a piece of paper towel over a shallow dish of light blue liquid on a wooden table. The grabber releases the paper towel on the dish. Static shot with no camera movement.,Fluid Dynamics,0129_perspective-right_trimmed-napkin-soak.mp4
+0130_perspective-left_take-1_trimmed-paint-on-glass.mp4,A clear acrylic sheet placed on a wooden table with a small dollop of red paint. A rotating paintbrush attached to a rotating platform rotates clockwise and goes through the paint. Static shot with no camera movement.,Fluid Dynamics,0130_perspective-left_trimmed-paint-on-glass.mp4
+0131_perspective-center_take-1_trimmed-paint-on-glass.mp4,A clear acrylic sheet placed on a wooden table with a small dollop of red paint. A rotating paintbrush attached to a rotating platform rotates clockwise and goes through the paint. Static shot with no camera movement.,Fluid Dynamics,0131_perspective-center_trimmed-paint-on-glass.mp4
+0132_perspective-right_take-1_trimmed-paint-on-glass.mp4,A clear acrylic sheet placed on a wooden table with a small dollop of red paint. A rotating paintbrush attached to a rotating platform rotates clockwise and goes through the paint. Static shot with no camera movement.,Fluid Dynamics,0132_perspective-right_trimmed-paint-on-glass.mp4
+0133_perspective-left_take-1_trimmed-paper-fall-water.mp4,A grabber tool is holding a crumpled piece of paper over a bowl of water on a wooden table. The grabber then releases the crumpled paper onto the bowl. Static shot with no camera movement.,Fluid Dynamics,0133_perspective-left_trimmed-paper-fall-water.mp4
+0134_perspective-center_take-1_trimmed-paper-fall-water.mp4,A grabber tool is holding a crumpled piece of paper over a bowl of water on a wooden table. The grabber then releases the crumpled paper onto the bowl. Static shot with no camera movement.,Fluid Dynamics,0134_perspective-center_trimmed-paper-fall-water.mp4
+0135_perspective-right_take-1_trimmed-paper-fall-water.mp4,A grabber tool is holding a crumpled piece of paper over a bowl of water on a wooden table. The grabber then releases the crumpled paper onto the bowl. Static shot with no camera movement.,Fluid Dynamics,0135_perspective-right_trimmed-paper-fall-water.mp4
+0136_perspective-left_take-1_trimmed-paper-in-water.mp4,A small piece of crumpled white paper is being lowered into a tall glass containing blue liquid with a green band showing the water level. The crumpled paper is released into the glass. Static shot with no camera movement.,Fluid Dynamics,0136_perspective-left_trimmed-paper-in-water.mp4
+0137_perspective-center_take-1_trimmed-paper-in-water.mp4,A small piece of crumpled white paper is being lowered into a tall glass containing blue liquid with a green band showing the water level. The crumpled paper is released into the glass. Static shot with no camera movement.,Fluid Dynamics,0137_perspective-center_trimmed-paper-in-water.mp4
+0138_perspective-right_take-1_trimmed-paper-in-water.mp4,A small piece of crumpled white paper is being lowered into a tall glass containing blue liquid with a green band showing the water level. The crumpled paper is released into the glass. Static shot with no camera movement.,Fluid Dynamics,0138_perspective-right_trimmed-paper-in-water.mp4
+0139_perspective-left_take-1_trimmed-paper-smoke.mp4,A piece of folded paper is placed on a glass cutting board. The paper is being burnt and white smoke is emitting from it. Static shot with no camera movement.,Thermodynamics,0139_perspective-left_trimmed-paper-smoke.mp4
+0140_perspective-center_take-1_trimmed-paper-smoke.mp4,A piece of folded paper is placed on a glass cutting board. The paper is being burnt and white smoke is emitting from it. Static shot with no camera movement.,Thermodynamics,0140_perspective-center_trimmed-paper-smoke.mp4
+0141_perspective-right_take-1_trimmed-paper-smoke.mp4,A piece of folded paper is placed on a glass cutting board. The paper is being burnt and white smoke is emitting from it. Static shot with no camera movement.,Thermodynamics,0141_perspective-right_trimmed-paper-smoke.mp4
+0142_perspective-left_take-1_trimmed-potato-in-water.mp4,A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass. Static shot with no camera movement.,Fluid Dynamics,0142_perspective-left_trimmed-potato-in-water.mp4
+0143_perspective-center_take-1_trimmed-potato-in-water.mp4,A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass. Static shot with no camera movement.,Fluid Dynamics,0143_perspective-center_trimmed-potato-in-water.mp4
+0144_perspective-right_take-1_trimmed-potato-in-water.mp4,A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass. Static shot with no camera movement.,Fluid Dynamics,0144_perspective-right_trimmed-potato-in-water.mp4
+0145_perspective-left_take-1_trimmed-roll-behind-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0145_perspective-left_trimmed-roll-behind-box.mp4
+0146_perspective-center_take-1_trimmed-roll-behind-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0146_perspective-center_trimmed-roll-behind-box.mp4
+0147_perspective-right_take-1_trimmed-roll-behind-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0147_perspective-right_trimmed-roll-behind-box.mp4
+0148_perspective-left_take-1_trimmed-roll-front-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0148_perspective-left_trimmed-roll-front-box.mp4
+0149_perspective-center_take-1_trimmed-roll-front-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0149_perspective-center_trimmed-roll-front-box.mp4
+0150_perspective-right_take-1_trimmed-roll-front-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0150_perspective-right_trimmed-roll-front-box.mp4
+0151_perspective-left_take-1_trimmed-roll-in-box.mp4,An olive green fabric box is on a light wood surface. A brown tennis ball rolls out of the black tube sitting on the table and rolls towards the box. Static shot with no camera movement.,Solid Mechanics,0151_perspective-left_trimmed-roll-in-box.mp4
+0152_perspective-center_take-1_trimmed-roll-in-box.mp4,An olive green fabric box is on a light wood surface. A brown tennis ball rolls out of the black tube sitting on the table and rolls towards the box. Static shot with no camera movement.,Solid Mechanics,0152_perspective-center_trimmed-roll-in-box.mp4
+0153_perspective-right_take-1_trimmed-roll-in-box.mp4,An olive green fabric box is on a light wood surface. A brown tennis ball rolls out of the black tube sitting on the table and rolls towards the box. Static shot with no camera movement.,Solid Mechanics,0153_perspective-right_trimmed-roll-in-box.mp4
+0154_perspective-left_take-1_trimmed-rolling-reflection.mp4,A 30lb kettlebell resting on a wooden table next to a mirror. A tennis ball rolls towards the kettlebell. Static shot with no camera movement.,Optics,0154_perspective-left_trimmed-rolling-reflection.mp4
+0155_perspective-center_take-1_trimmed-rolling-reflection.mp4,A 30lb kettlebell resting on a wooden table next to a mirror. A tennis ball rolls towards the kettlebell. Static shot with no camera movement.,Optics,0155_perspective-center_trimmed-rolling-reflection.mp4
+0156_perspective-right_take-1_trimmed-rolling-reflection.mp4,A 30lb kettlebell resting on a wooden table next to a mirror. A tennis ball rolls towards the kettlebell. Static shot with no camera movement.,Optics,0156_perspective-right_trimmed-rolling-reflection.mp4
+0157_perspective-left_take-1_trimmed-silk-cover.mp4,A teapot is placed on a wooden table. a piece of silk fabric is lowered on the teapot to cover it. Static shot with no camera movement.,Solid Mechanics,0157_perspective-left_trimmed-silk-cover.mp4
+0158_perspective-center_take-1_trimmed-silk-cover.mp4,A teapot is placed on a wooden table. a piece of silk fabric is lowered on the teapot to cover it. Static shot with no camera movement.,Solid Mechanics,0158_perspective-center_trimmed-silk-cover.mp4
+0159_perspective-right_take-1_trimmed-silk-cover.mp4,A teapot is placed on a wooden table. a piece of silk fabric is lowered on the teapot to cover it. Static shot with no camera movement.,Solid Mechanics,0159_perspective-right_trimmed-silk-cover.mp4
+0160_perspective-left_take-1_trimmed-single-cradle.mp4,A Newton's cradle device on the table and one of the metal balls is held up by a blue handled grabber tool. The claw releases the ball. Static shot with no camera movement.,Solid Mechanics,0160_perspective-left_trimmed-single-cradle.mp4
+0161_perspective-center_take-1_trimmed-single-cradle.mp4,A Newton's cradle device on the table and one of the metal balls is held up by a blue handled grabber tool. The claw releases the ball. Static shot with no camera movement.,Solid Mechanics,0161_perspective-center_trimmed-single-cradle.mp4
+0162_perspective-right_take-1_trimmed-single-cradle.mp4,A Newton's cradle device on the table and one of the metal balls is held up by a blue handled grabber tool. The claw releases the ball. Static shot with no camera movement.,Solid Mechanics,0162_perspective-right_trimmed-single-cradle.mp4
+0163_perspective-left_take-1_trimmed-siphon.mp4,A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks. Static shot with no camera movement.,Fluid Dynamics,0163_perspective-left_trimmed-siphon.mp4
+0164_perspective-center_take-1_trimmed-siphon.mp4,A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks. Static shot with no camera movement.,Fluid Dynamics,0164_perspective-center_trimmed-siphon.mp4
+0165_perspective-right_take-1_trimmed-siphon.mp4,A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks. Static shot with no camera movement.,Fluid Dynamics,0165_perspective-right_trimmed-siphon.mp4
+0166_perspective-left_take-1_trimmed-smiley-ball-rotates.mp4,A tennis ball with a smiley face drawn on it is placed on a rotating black platform that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0166_perspective-left_trimmed-smiley-ball-rotates.mp4
+0167_perspective-center_take-1_trimmed-smiley-ball-rotates.mp4,A tennis ball with a smiley face drawn on it is placed on a rotating black platform that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0167_perspective-center_trimmed-smiley-ball-rotates.mp4
+0168_perspective-right_take-1_trimmed-smiley-ball-rotates.mp4,A tennis ball with a smiley face drawn on it is placed on a rotating black platform that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0168_perspective-right_trimmed-smiley-ball-rotates.mp4
+0169_perspective-left_take-1_trimmed-solid-ball-peakaboo.mp4,A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again. Static shot with no camera movement.,Solid Mechanics,0169_perspective-left_trimmed-solid-ball-peakaboo.mp4
+0170_perspective-center_take-1_trimmed-solid-ball-peakaboo.mp4,A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again. Static shot with no camera movement.,Solid Mechanics,0170_perspective-center_trimmed-solid-ball-peakaboo.mp4
+0171_perspective-right_take-1_trimmed-solid-ball-peakaboo.mp4,A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again. Static shot with no camera movement.,Solid Mechanics,0171_perspective-right_trimmed-solid-ball-peakaboo.mp4
+0172_perspective-left_take-1_trimmed-stable-blocks.mp4,A pink block is being lowered towards a simple structure made of colorful blocks resembling a gate. Static shot with no camera movement.,Solid Mechanics,0172_perspective-left_trimmed-stable-blocks.mp4
+0173_perspective-center_take-1_trimmed-stable-blocks.mp4,A pink block is being lowered towards a simple structure made of colorful blocks resembling a gate. Static shot with no camera movement.,Solid Mechanics,0173_perspective-center_trimmed-stable-blocks.mp4
+0174_perspective-right_take-1_trimmed-stable-blocks.mp4,A pink block is being lowered towards a simple structure made of colorful blocks resembling a gate. Static shot with no camera movement.,Solid Mechanics,0174_perspective-right_trimmed-stable-blocks.mp4
+0175_perspective-left_take-1_trimmed-teapot-rotates.mp4,A teapot is placed on a rotating display that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0175_perspective-left_trimmed-teapot-rotates.mp4
+0176_perspective-center_take-1_trimmed-teapot-rotates.mp4,A teapot is placed on a rotating display that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0176_perspective-center_trimmed-teapot-rotates.mp4
+0177_perspective-right_take-1_trimmed-teapot-rotates.mp4,A teapot is placed on a rotating display that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0177_perspective-right_trimmed-teapot-rotates.mp4
+0178_perspective-left_take-1_trimmed-two-balls-pass.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards eachother. Static shot with no camera movement.,Solid Mechanics,0178_perspective-left_trimmed-two-balls-pass.mp4
+0179_perspective-center_take-1_trimmed-two-balls-pass.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards eachother. Static shot with no camera movement.,Solid Mechanics,0179_perspective-center_trimmed-two-balls-pass.mp4
+0180_perspective-right_take-1_trimmed-two-balls-pass.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards eachother. Static shot with no camera movement.,Solid Mechanics,0180_perspective-right_trimmed-two-balls-pass.mp4
+0181_perspective-left_take-1_trimmed-unstable-block-stack.mp4,A grabber tool carefully placing a blue wooden block on top of a yellow block which is balanced on a red block forming an L shape. Static shot with no camera movement.,Solid Mechanics,0181_perspective-left_trimmed-unstable-block-stack.mp4
+0182_perspective-center_take-1_trimmed-unstable-block-stack.mp4,A grabber tool carefully placing a blue wooden block on top of a yellow block which is balanced on a red block forming an L shape. Static shot with no camera movement.,Solid Mechanics,0182_perspective-center_trimmed-unstable-block-stack.mp4
+0183_perspective-right_take-1_trimmed-unstable-block-stack.mp4,A grabber tool carefully placing a blue wooden block on top of a yellow block which is balanced on a red block forming an L shape. Static shot with no camera movement.,Solid Mechanics,0183_perspective-right_trimmed-unstable-block-stack.mp4
+0184_perspective-left_take-1_trimmed-water-in-juice.mp4,A glass beverage dispenser is dispensing water into a glass which has some grapefruit juice in it. Static shot with no camera movement.,Fluid Dynamics,0184_perspective-left_trimmed-water-in-juice.mp4
+0185_perspective-center_take-1_trimmed-water-in-juice.mp4,A glass beverage dispenser is dispensing water into a glass which has some grapefruit juice in it. Static shot with no camera movement.,Fluid Dynamics,0185_perspective-center_trimmed-water-in-juice.mp4
+0186_perspective-right_take-1_trimmed-water-in-juice.mp4,A glass beverage dispenser is dispensing water into a glass which has some grapefruit juice in it. Static shot with no camera movement.,Fluid Dynamics,0186_perspective-right_trimmed-water-in-juice.mp4
+0187_perspective-left_take-1_trimmed-weight-on-ceramic.mp4,A 30lb kettlebell is slowly lowered on top of a yellow ceramic coffee mug placed on a wooden table. Static shot with no camera movement.,Solid Mechanics,0187_perspective-left_trimmed-weight-on-ceramic.mp4
+0188_perspective-center_take-1_trimmed-weight-on-ceramic.mp4,A 30lb kettlebell is slowly lowered on top of a yellow ceramic coffee mug placed on a wooden table. Static shot with no camera movement.,Solid Mechanics,0188_perspective-center_trimmed-weight-on-ceramic.mp4
+0189_perspective-right_take-1_trimmed-weight-on-ceramic.mp4,A 30lb kettlebell is slowly lowered on top of a yellow ceramic coffee mug placed on a wooden table. Static shot with no camera movement.,Solid Mechanics,0189_perspective-right_trimmed-weight-on-ceramic.mp4
+0190_perspective-left_take-1_trimmed-weight-on-paper.mp4,A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.,Solid Mechanics,0190_perspective-left_trimmed-weight-on-paper.mp4
+0191_perspective-center_take-1_trimmed-weight-on-paper.mp4,A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.,Solid Mechanics,0191_perspective-center_trimmed-weight-on-paper.mp4
+0192_perspective-right_take-1_trimmed-weight-on-paper.mp4,A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.,Solid Mechanics,0192_perspective-right_trimmed-weight-on-paper.mp4
+0193_perspective-left_take-1_trimmed-weight-on-pillow.mp4,A 30lb kettlebell and a green piece of paper are lowered onto two pillows. Static shot with no camera movement.,Solid Mechanics,0193_perspective-left_trimmed-weight-on-pillow.mp4
+0194_perspective-center_take-1_trimmed-weight-on-pillow.mp4,A 30lb kettlebell and a green piece of paper are lowered onto two pillows. Static shot with no camera movement.,Solid Mechanics,0194_perspective-center_trimmed-weight-on-pillow.mp4
+0195_perspective-right_take-1_trimmed-weight-on-pillow.mp4,A 30lb kettlebell and a green piece of paper are lowered onto two pillows. Static shot with no camera movement.,Solid Mechanics,0195_perspective-right_trimmed-weight-on-pillow.mp4
+0196_perspective-left_take-1_trimmed-weight-protects-duck.mp4,A light beige coffee table with a black kettlebell and a yellow rubber duck on it. A grey tennis ball rolls out of the black tube sitting on the table and towards the duck and kettlebell. Static shot with no camera movement.,Solid Mechanics,0196_perspective-left_trimmed-weight-protects-duck.mp4
+0197_perspective-center_take-1_trimmed-weight-protects-duck.mp4,A light beige coffee table with a black kettlebell and a yellow rubber duck on it. A grey tennis ball rolls out of the black tube sitting on the table and towards the duck and kettlebell. Static shot with no camera movement.,Solid Mechanics,0197_perspective-center_trimmed-weight-protects-duck.mp4
+0198_perspective-right_take-1_trimmed-weight-protects-duck.mp4,A light beige coffee table with a black kettlebell and a yellow rubber duck on it. A grey tennis ball rolls out of the black tube sitting on the table and towards the duck and kettlebell. Static shot with no camera movement.,Solid Mechanics,0198_perspective-right_trimmed-weight-protects-duck.mp4
+0199_perspective-left_take-2_trimmed-ball-and-block-fall.mp4,Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The camera is static and the grabber tools let go of the ball and block. Static shot with no camera movement.,Solid Mechanics,0199_perspective-left_trimmed-ball-and-block-fall.mp4
+0200_perspective-center_take-2_trimmed-ball-and-block-fall.mp4,Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The camera is static and the grabber tools let go of the ball and block. Static shot with no camera movement.,Solid Mechanics,0200_perspective-center_trimmed-ball-and-block-fall.mp4
+0201_perspective-right_take-2_trimmed-ball-and-block-fall.mp4,Two pillows on a table and two grabber tools hanging above them from which a brown tennis ball and an orange block are suspended. The camera is static and the grabber tools let go of the ball and block. Static shot with no camera movement.,Solid Mechanics,0201_perspective-right_trimmed-ball-and-block-fall.mp4
+0202_perspective-left_take-2_trimmed-ball-behind-rotating-paper.mp4,A grabber arm is holding a tennis ball above a piece of cardstock propped up on a rotating platform sitting on a table that rotates clockwise. The grabber lowers the ball and places is on the table as the cardstock rotates. Static shot with no camera movement.,Solid Mechanics,0202_perspective-left_trimmed-ball-behind-rotating-paper.mp4
+0203_perspective-center_take-2_trimmed-ball-behind-rotating-paper.mp4,A grabber arm is holding a tennis ball above a piece of cardstock propped up on a rotating platform sitting on a table that rotates clockwise. The grabber lowers the ball and places is on the table as the cardstock rotates. Static shot with no camera movement.,Solid Mechanics,0203_perspective-center_trimmed-ball-behind-rotating-paper.mp4
+0204_perspective-right_take-2_trimmed-ball-behind-rotating-paper.mp4,A grabber arm is holding a tennis ball above a piece of cardstock propped up on a rotating platform sitting on a table that rotates clockwise. The grabber lowers the ball and places is on the table as the cardstock rotates. Static shot with no camera movement.,Solid Mechanics,0204_perspective-right_trimmed-ball-behind-rotating-paper.mp4
+0205_perspective-left_take-2_trimmed-ball-hits-duck.mp4,A light beige coffee table with a small yellow rubber ducky on it. A mustard yellow couch is in the background. There is a black pipe on one end of the table and a brown tennis ball rolls out of it towards the rubber ducky. Static shot with no camera movement.,Solid Mechanics,0205_perspective-left_trimmed-ball-hits-duck.mp4
+0206_perspective-center_take-2_trimmed-ball-hits-duck.mp4,A light beige coffee table with a small yellow rubber ducky on it. A mustard yellow couch is in the background. There is a black pipe on one end of the table and a brown tennis ball rolls out of it towards the rubber ducky. Static shot with no camera movement.,Solid Mechanics,0206_perspective-center_trimmed-ball-hits-duck.mp4
+0207_perspective-right_take-2_trimmed-ball-hits-duck.mp4,A light beige coffee table with a small yellow rubber ducky on it. A mustard yellow couch is in the background. There is a black pipe on one end of the table and a brown tennis ball rolls out of it towards the rubber ducky. Static shot with no camera movement.,Solid Mechanics,0207_perspective-right_trimmed-ball-hits-duck.mp4
+0208_perspective-left_take-2_trimmed-ball-hits-nothing.mp4,A light-colored wooden coffee table with a few small objects on it including a tennis ball and a smaller red ball. An orange ball rolls out of a black pipe that is sitting on the table towards the right side. Static shot with no camera movement.,Solid Mechanics,0208_perspective-left_trimmed-ball-hits-nothing.mp4
+0209_perspective-center_take-2_trimmed-ball-hits-nothing.mp4,A light-colored wooden coffee table with a few small objects on it including a tennis ball and a smaller red ball. An orange ball rolls out of a black pipe that is sitting on the table towards the right side. Static shot with no camera movement.,Solid Mechanics,0209_perspective-center_trimmed-ball-hits-nothing.mp4
+0210_perspective-right_take-2_trimmed-ball-hits-nothing.mp4,A light-colored wooden coffee table with a few small objects on it including a tennis ball and a smaller red ball. An orange ball rolls out of a black pipe that is sitting on the table towards the right side. Static shot with no camera movement.,Solid Mechanics,0210_perspective-right_trimmed-ball-hits-nothing.mp4
+0211_perspective-left_take-2_trimmed-ball-in-basket.mp4,An orange inflatable basketball is suspended above a black plastic crate placed on a wooden table. The ball is then released. Static shot with no camera movement.,Solid Mechanics,0211_perspective-left_trimmed-ball-in-basket.mp4
+0212_perspective-center_take-2_trimmed-ball-in-basket.mp4,An orange inflatable basketball is suspended above a black plastic crate placed on a wooden table. The ball is then released. Static shot with no camera movement.,Solid Mechanics,0212_perspective-center_trimmed-ball-in-basket.mp4
+0213_perspective-right_take-2_trimmed-ball-in-basket.mp4,An orange inflatable basketball is suspended above a black plastic crate placed on a wooden table. The ball is then released. Static shot with no camera movement.,Solid Mechanics,0213_perspective-right_trimmed-ball-in-basket.mp4
+0214_perspective-left_take-2_trimmed-ball-in-sand.mp4,A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball. Static shot with no camera movement.,Solid Mechanics,0214_perspective-left_trimmed-ball-in-sand.mp4
+0215_perspective-center_take-2_trimmed-ball-in-sand.mp4,A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball. Static shot with no camera movement.,Solid Mechanics,0215_perspective-center_trimmed-ball-in-sand.mp4
+0216_perspective-right_take-2_trimmed-ball-in-sand.mp4,A blue grabber tool holds a tennis ball above a pile of green kinetic sand on a wooden table. The grabber then releases the ball. Static shot with no camera movement.,Solid Mechanics,0216_perspective-right_trimmed-ball-in-sand.mp4
+0217_perspective-left_take-2_trimmed-ball-ramp.mp4,A simple ramp made of cardboard propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame and a yellow tennis ball rolls out of the pipe towards the ramp. Static shot with no camera movement.,Solid Mechanics,0217_perspective-left_trimmed-ball-ramp.mp4
+0218_perspective-center_take-2_trimmed-ball-ramp.mp4,A simple ramp made of cardboard propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame and a yellow tennis ball rolls out of the pipe towards the ramp. Static shot with no camera movement.,Solid Mechanics,0218_perspective-center_trimmed-ball-ramp.mp4
+0219_perspective-right_take-2_trimmed-ball-ramp.mp4,A simple ramp made of cardboard propped up by a blue block on a light-colored wooden table. There's a black pipe to the left of the frame and a yellow tennis ball rolls out of the pipe towards the ramp. Static shot with no camera movement.,Solid Mechanics,0219_perspective-right_trimmed-ball-ramp.mp4
+0220_perspective-left_take-2_trimmed-ball-rolls-off.mp4,A light wood coffee table in the foreground with a black pipe on the end of the table. A grey tennis ball rolls out of the pipe towards the right and onto the table. Static shot with no camera movement.,Solid Mechanics,0220_perspective-left_trimmed-ball-rolls-off.mp4
+0221_perspective-center_take-2_trimmed-ball-rolls-off.mp4,A light wood coffee table in the foreground with a black pipe on the end of the table. A grey tennis ball rolls out of the pipe towards the right and onto the table. Static shot with no camera movement.,Solid Mechanics,0221_perspective-center_trimmed-ball-rolls-off.mp4
+0222_perspective-right_take-2_trimmed-ball-rolls-off.mp4,A light wood coffee table in the foreground with a black pipe on the end of the table. A grey tennis ball rolls out of the pipe towards the right and onto the table. Static shot with no camera movement.,Solid Mechanics,0222_perspective-right_trimmed-ball-rolls-off.mp4
+0223_perspective-left_take-2_trimmed-ball-rolls-on-glass.mp4,A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. A blue tennis ball rolls on the wooden table and towards the glass. Static shot with no camera movement.,Solid Mechanics,0223_perspective-left_trimmed-ball-rolls-on-glass.mp4
+0224_perspective-center_take-2_trimmed-ball-rolls-on-glass.mp4,A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. A blue tennis ball rolls on the wooden table and towards the glass. Static shot with no camera movement.,Solid Mechanics,0224_perspective-center_trimmed-ball-rolls-on-glass.mp4
+0225_perspective-right_take-2_trimmed-ball-rolls-on-glass.mp4,A piece of clear glass resting on the edge of a light-colored wooden table against a plain white wall. A blue tennis ball rolls on the wooden table and towards the glass. Static shot with no camera movement.,Solid Mechanics,0225_perspective-right_trimmed-ball-rolls-on-glass.mp4
+0226_perspective-left_take-2_trimmed-ball-train.mp4,"A light-colored coffee table with two tennis balls, one orange and one brown, placed near the center back to back. A grey tennis ball rolls out of a black pipe sitting on the table and towards the other two balls. Static shot with no camera movement.",Solid Mechanics,0226_perspective-left_trimmed-ball-train.mp4
+0227_perspective-center_take-2_trimmed-ball-train.mp4,"A light-colored coffee table with two tennis balls, one orange and one brown, placed near the center back to back. A grey tennis ball rolls out of a black pipe sitting on the table and towards the other two balls. Static shot with no camera movement.",Solid Mechanics,0227_perspective-center_trimmed-ball-train.mp4
+0228_perspective-right_take-2_trimmed-ball-train.mp4,"A light-colored coffee table with two tennis balls, one orange and one brown, placed near the center back to back. A grey tennis ball rolls out of a black pipe sitting on the table and towards the other two balls. Static shot with no camera movement.",Solid Mechanics,0228_perspective-right_trimmed-ball-train.mp4
+0229_perspective-left_take-2_trimmed-balls-collide.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0229_perspective-left_trimmed-balls-collide.mp4
+0230_perspective-center_take-2_trimmed-balls-collide.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0230_perspective-center_trimmed-balls-collide.mp4
+0231_perspective-right_take-2_trimmed-balls-collide.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0231_perspective-right_trimmed-balls-collide.mp4
+0232_perspective-left_take-2_trimmed-block-domino.mp4,A row of colorful wooden blocks lined up on a wooden table with a wooden stick attached to a black rotating platform. The platform rotates clockwise and the wooden stick hits the first block as it rotates. Static shot with no camera movement.,Solid Mechanics,0232_perspective-left_trimmed-block-domino.mp4
+0233_perspective-center_take-2_trimmed-block-domino.mp4,A row of colorful wooden blocks lined up on a wooden table with a wooden stick attached to a black rotating platform. The platform rotates clockwise and the wooden stick hits the first block as it rotates. Static shot with no camera movement.,Solid Mechanics,0233_perspective-center_trimmed-block-domino.mp4
+0234_perspective-right_take-2_trimmed-block-domino.mp4,A row of colorful wooden blocks lined up on a wooden table with a wooden stick attached to a black rotating platform. The platform rotates clockwise and the wooden stick hits the first block as it rotates. Static shot with no camera movement.,Solid Mechanics,0234_perspective-right_trimmed-block-domino.mp4
+0235_perspective-left_take-2_trimmed-blow-balloon.mp4,A black balloon is attached to a fixed electric air pump hose on a table with a plain wall in the background. Air is being pumped in the balloon. Static shot with no camera movement.,Fluid Dynamics,0235_perspective-left_trimmed-blow-balloon.mp4
+0236_perspective-center_take-2_trimmed-blow-balloon.mp4,A black balloon is attached to a fixed electric air pump hose on a table with a plain wall in the background. Air is being pumped in the balloon. Static shot with no camera movement.,Fluid Dynamics,0236_perspective-center_trimmed-blow-balloon.mp4
+0237_perspective-right_take-2_trimmed-blow-balloon.mp4,A black balloon is attached to a fixed electric air pump hose on a table with a plain wall in the background. Air is being pumped in the balloon. Static shot with no camera movement.,Fluid Dynamics,0237_perspective-right_trimmed-blow-balloon.mp4
+0238_perspective-left_take-2_trimmed-cut-orange.mp4,A tangerine that has been cut in half is placed on a glass cutting board. A knife is slicing through the tangerine. Static shot with no camera movement.,Solid Mechanics,0238_perspective-left_trimmed-cut-orange.mp4
+0239_perspective-center_take-2_trimmed-cut-orange.mp4,A tangerine that has been cut in half is placed on a glass cutting board. A knife is slicing through the tangerine. Static shot with no camera movement.,Solid Mechanics,0239_perspective-center_trimmed-cut-orange.mp4
+0240_perspective-right_take-2_trimmed-cut-orange.mp4,A tangerine that has been cut in half is placed on a glass cutting board. A knife is slicing through the tangerine. Static shot with no camera movement.,Solid Mechanics,0240_perspective-right_trimmed-cut-orange.mp4
+0241_perspective-left_take-2_trimmed-cut-paper.mp4,"Two black and blue gripping tools are pulling a piece of green paper from its two corners, causing it to tear. Static shot with no camera movement.",Solid Mechanics,0241_perspective-left_trimmed-cut-paper.mp4
+0242_perspective-center_take-2_trimmed-cut-paper.mp4,"Two black and blue gripping tools are pulling a piece of green paper from its two corners, causing it to tear. Static shot with no camera movement.",Solid Mechanics,0242_perspective-center_trimmed-cut-paper.mp4
+0243_perspective-right_take-2_trimmed-cut-paper.mp4,"Two black and blue gripping tools are pulling a piece of green paper from its two corners, causing it to tear. Static shot with no camera movement.",Solid Mechanics,0243_perspective-right_trimmed-cut-paper.mp4
+0244_perspective-left_take-2_trimmed-domino-in-juice.mp4,A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.,Fluid Dynamics,0244_perspective-left_trimmed-domino-in-juice.mp4
+0245_perspective-center_take-2_trimmed-domino-in-juice.mp4,A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.,Fluid Dynamics,0245_perspective-center_trimmed-domino-in-juice.mp4
+0246_perspective-right_take-2_trimmed-domino-in-juice.mp4,A grabber tool holding a white domino drops the domino into a dark-colored liquid in a blue mug that is on a wooden surface. Static shot with no camera movement.,Fluid Dynamics,0246_perspective-right_trimmed-domino-in-juice.mp4
+0247_perspective-left_take-2_trimmed-dominos-with-space.mp4,Two rows of alternating black and white dominoes are set up on a wooden table with a gap between the two rows. A wooden stick attached to a rotating platform rotates clockwise and knocks the first domino in the first row. Static shot with no camera movement.,Solid Mechanics,0247_perspective-left_trimmed-dominos-with-space.mp4
+0248_perspective-center_take-2_trimmed-dominos-with-space.mp4,Two rows of alternating black and white dominoes are set up on a wooden table with a gap between the two rows. A wooden stick attached to a rotating platform rotates clockwise and knocks the first domino in the first row. Static shot with no camera movement.,Solid Mechanics,0248_perspective-center_trimmed-dominos-with-space.mp4
+0249_perspective-right_take-2_trimmed-dominos-with-space.mp4,Two rows of alternating black and white dominoes are set up on a wooden table with a gap between the two rows. A wooden stick attached to a rotating platform rotates clockwise and knocks the first domino in the first row. Static shot with no camera movement.,Solid Mechanics,0249_perspective-right_trimmed-dominos-with-space.mp4
+0250_perspective-left_take-2_trimmed-double-cradle.mp4,A Newton's cradle device on the table and two of the metal balls are held up by a blue handled grabber tool. The claw releases the two balls. Static shot with no camera movement.,Solid Mechanics,0250_perspective-left_trimmed-double-cradle.mp4
+0251_perspective-center_take-2_trimmed-double-cradle.mp4,A Newton's cradle device on the table and two of the metal balls are held up by a blue handled grabber tool. The claw releases the two balls. Static shot with no camera movement.,Solid Mechanics,0251_perspective-center_trimmed-double-cradle.mp4
+0252_perspective-right_take-2_trimmed-double-cradle.mp4,A Newton's cradle device on the table and two of the metal balls are held up by a blue handled grabber tool. The claw releases the two balls. Static shot with no camera movement.,Solid Mechanics,0252_perspective-right_trimmed-double-cradle.mp4
+0253_perspective-left_take-2_trimmed-duck-and-dominos.mp4,A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.,Solid Mechanics,0253_perspective-left_trimmed-duck-and-dominos.mp4
+0254_perspective-center_take-2_trimmed-duck-and-dominos.mp4,A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.,Solid Mechanics,0254_perspective-center_trimmed-duck-and-dominos.mp4
+0255_perspective-right_take-2_trimmed-duck-and-dominos.mp4,A yellow rubber duck is positioned in the middle of a line of black and white dominoes on a wooden table. A stick attached to a black rotating platform rotates clockwise and knocks the first domino block. Static shot with no camera movement.,Solid Mechanics,0255_perspective-right_trimmed-duck-and-dominos.mp4
+0256_perspective-left_take-2_trimmed-duck-falls-in-box.mp4,A yellow rubber ducky is suspended above an open dark green fabric box on a wooden table. The duck is then released. Static shot with no camera movement.,Solid Mechanics,0256_perspective-left_trimmed-duck-falls-in-box.mp4
+0257_perspective-center_take-2_trimmed-duck-falls-in-box.mp4,A yellow rubber ducky is suspended above an open dark green fabric box on a wooden table. The duck is then released. Static shot with no camera movement.,Solid Mechanics,0257_perspective-center_trimmed-duck-falls-in-box.mp4
+0258_perspective-right_take-2_trimmed-duck-falls-in-box.mp4,A yellow rubber ducky is suspended above an open dark green fabric box on a wooden table. The duck is then released. Static shot with no camera movement.,Solid Mechanics,0258_perspective-right_trimmed-duck-falls-in-box.mp4
+0259_perspective-left_take-2_trimmed-duck-static.mp4,A stationary yellow rubber duck on a light brown wooden table against a plain white background. Static shot with no camera movement.,Solid Mechanics,0259_perspective-left_trimmed-duck-static.mp4
+0260_perspective-center_take-2_trimmed-duck-static.mp4,A stationary yellow rubber duck on a light brown wooden table against a plain white background. Static shot with no camera movement.,Solid Mechanics,0260_perspective-center_trimmed-duck-static.mp4
+0261_perspective-right_take-2_trimmed-duck-static.mp4,A stationary yellow rubber duck on a light brown wooden table against a plain white background. Static shot with no camera movement.,Solid Mechanics,0261_perspective-right_trimmed-duck-static.mp4
+0262_perspective-left_take-2_trimmed-fill-glass-red-drink.mp4,A glass beverage dispenser filled with a bright red liquid is set up on a woven basket and is pouring the liquid into a clear glass on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0262_perspective-left_trimmed-fill-glass-red-drink.mp4
+0263_perspective-center_take-2_trimmed-fill-glass-red-drink.mp4,A glass beverage dispenser filled with a bright red liquid is set up on a woven basket and is pouring the liquid into a clear glass on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0263_perspective-center_trimmed-fill-glass-red-drink.mp4
+0264_perspective-right_take-2_trimmed-fill-glass-red-drink.mp4,A glass beverage dispenser filled with a bright red liquid is set up on a woven basket and is pouring the liquid into a clear glass on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0264_perspective-right_trimmed-fill-glass-red-drink.mp4
+0265_perspective-left_take-2_trimmed-glass-stays-same.mp4,A glass beverage dispenser filled with a bright red liquid is set on a wicker base. Under the dispenser there is a glass half filled with red liquid. Static shot with no camera movement.,Fluid Dynamics,0265_perspective-left_trimmed-glass-stays-same.mp4
+0266_perspective-center_take-2_trimmed-glass-stays-same.mp4,A glass beverage dispenser filled with a bright red liquid is set on a wicker base. Under the dispenser there is a glass half filled with red liquid. Static shot with no camera movement.,Fluid Dynamics,0266_perspective-center_trimmed-glass-stays-same.mp4
+0267_perspective-right_take-2_trimmed-glass-stays-same.mp4,A glass beverage dispenser filled with a bright red liquid is set on a wicker base. Under the dispenser there is a glass half filled with red liquid. Static shot with no camera movement.,Fluid Dynamics,0267_perspective-right_trimmed-glass-stays-same.mp4
+0268_perspective-left_take-2_trimmed-juice-in-water.mp4,A glass beverage dispenser pouring grapefruit juice into a glass that has some water inside. Static shot with no camera movement.,Fluid Dynamics,0268_perspective-left_trimmed-juice-in-water.mp4
+0269_perspective-center_take-2_trimmed-juice-in-water.mp4,A glass beverage dispenser pouring grapefruit juice into a glass that has some water inside. Static shot with no camera movement.,Fluid Dynamics,0269_perspective-center_trimmed-juice-in-water.mp4
+0270_perspective-right_take-2_trimmed-juice-in-water.mp4,A glass beverage dispenser pouring grapefruit juice into a glass that has some water inside. Static shot with no camera movement.,Fluid Dynamics,0270_perspective-right_trimmed-juice-in-water.mp4
+0271_perspective-left_take-2_trimmed-light-on-block.mp4,A blue rectangular wooden block is placed on a black rotating turntable that rotates clockwise illuminated by a spotlight casting a long shadow on the wall behind it. Static shot with no camera movement.,Optics,0271_perspective-left_trimmed-light-on-block.mp4
+0272_perspective-center_take-2_trimmed-light-on-block.mp4,A blue rectangular wooden block is placed on a black rotating turntable that rotates clockwise illuminated by a spotlight casting a long shadow on the wall behind it. Static shot with no camera movement.,Optics,0272_perspective-center_trimmed-light-on-block.mp4
+0273_perspective-right_take-2_trimmed-light-on-block.mp4,A blue rectangular wooden block is placed on a black rotating turntable that rotates clockwise illuminated by a spotlight casting a long shadow on the wall behind it. Static shot with no camera movement.,Optics,0273_perspective-right_trimmed-light-on-block.mp4
+0274_perspective-left_take-2_trimmed-light-on-mug.mp4,A yellow mug is placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting a shadow on the wall behind it. Static shot with no camera movement.,Optics,0274_perspective-left_trimmed-light-on-mug.mp4
+0275_perspective-center_take-2_trimmed-light-on-mug.mp4,A yellow mug is placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting a shadow on the wall behind it. Static shot with no camera movement.,Optics,0275_perspective-center_trimmed-light-on-mug.mp4
+0276_perspective-right_take-2_trimmed-light-on-mug.mp4,A yellow mug is placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting a shadow on the wall behind it. Static shot with no camera movement.,Optics,0276_perspective-right_trimmed-light-on-mug.mp4
+0277_perspective-left_take-2_trimmed-light-on-mug-block.mp4,A yellow mug and a blue wooden block are placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting their shadow on the wall behind it. Static shot with no camera movement.,Optics,0277_perspective-left_trimmed-light-on-mug-block.mp4
+0278_perspective-center_take-2_trimmed-light-on-mug-block.mp4,A yellow mug and a blue wooden block are placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting their shadow on the wall behind it. Static shot with no camera movement.,Optics,0278_perspective-center_trimmed-light-on-mug-block.mp4
+0279_perspective-right_take-2_trimmed-light-on-mug-block.mp4,A yellow mug and a blue wooden block are placed on a rotating turntable that rotates clockwise illuminated by a spotlight casting their shadow on the wall behind it. Static shot with no camera movement.,Optics,0279_perspective-right_trimmed-light-on-mug-block.mp4
+0280_perspective-left_take-2_trimmed-light-on-statue.mp4,A small statue made of porcelain illuminated by a spotlight on a rotating base that rotates clockwise. The spotlight casts a large shadow of the statue onto the wall behind it. Static shot with no camera movement.,Optics,0280_perspective-left_trimmed-light-on-statue.mp4
+0281_perspective-center_take-2_trimmed-light-on-statue.mp4,A small statue made of porcelain illuminated by a spotlight on a rotating base that rotates clockwise. The spotlight casts a large shadow of the statue onto the wall behind it. Static shot with no camera movement.,Optics,0281_perspective-center_trimmed-light-on-statue.mp4
+0282_perspective-right_take-2_trimmed-light-on-statue.mp4,A small statue made of porcelain illuminated by a spotlight on a rotating base that rotates clockwise. The spotlight casts a large shadow of the statue onto the wall behind it. Static shot with no camera movement.,Optics,0282_perspective-right_trimmed-light-on-statue.mp4
+0283_perspective-left_take-2_trimmed-liquid-on-duck.mp4,A yellow rubber ducky is placed in an empty black baking pan on a wooden table.  A beverage dispenser with red liquid inside sits on a woven basket behind it. The liquid pours on the duck from the dispenser. Static shot with no camera movement.,Fluid Dynamics,0283_perspective-left_trimmed-liquid-on-duck.mp4
+0284_perspective-center_take-2_trimmed-liquid-on-duck.mp4,A yellow rubber ducky is placed in an empty black baking pan on a wooden table.  A beverage dispenser with red liquid inside sits on a woven basket behind it. The liquid pours on the duck from the dispenser. Static shot with no camera movement.,Fluid Dynamics,0284_perspective-center_trimmed-liquid-on-duck.mp4
+0285_perspective-right_take-2_trimmed-liquid-on-duck.mp4,A yellow rubber ducky is placed in an empty black baking pan on a wooden table.  A beverage dispenser with red liquid inside sits on a woven basket behind it. The liquid pours on the duck from the dispenser. Static shot with no camera movement.,Fluid Dynamics,0285_perspective-right_trimmed-liquid-on-duck.mp4
+0286_perspective-left_take-2_trimmed-liquid-overfill.mp4,A bright red liquid being poured from a dispenser into a glass which is placed on a dark baking tray on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0286_perspective-left_trimmed-liquid-overfill.mp4
+0287_perspective-center_take-2_trimmed-liquid-overfill.mp4,A bright red liquid being poured from a dispenser into a glass which is placed on a dark baking tray on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0287_perspective-center_trimmed-liquid-overfill.mp4
+0288_perspective-right_take-2_trimmed-liquid-overfill.mp4,A bright red liquid being poured from a dispenser into a glass which is placed on a dark baking tray on a wooden table. Static shot with no camera movement.,Fluid Dynamics,0288_perspective-right_trimmed-liquid-overfill.mp4
+0289_perspective-left_take-2_trimmed-lit-candle.mp4,Two candle holders that have tall red candles in them are placed on a wooden table. One of the candles is burning. Static shot with no camera movement.,Thermodynamics,0289_perspective-left_trimmed-lit-candle.mp4
+0290_perspective-center_take-2_trimmed-lit-candle.mp4,Two candle holders that have tall red candles in them are placed on a wooden table. One of the candles is burning. Static shot with no camera movement.,Thermodynamics,0290_perspective-center_trimmed-lit-candle.mp4
+0291_perspective-right_take-2_trimmed-lit-candle.mp4,Two candle holders that have tall red candles in them are placed on a wooden table. One of the candles is burning. Static shot with no camera movement.,Thermodynamics,0291_perspective-right_trimmed-lit-candle.mp4
+0292_perspective-left_take-2_trimmed-magnet-domino.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small white plastic domino block is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0292_perspective-left_trimmed-magnet-domino.mp4
+0293_perspective-center_take-2_trimmed-magnet-domino.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small white plastic domino block is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0293_perspective-center_trimmed-magnet-domino.mp4
+0294_perspective-right_take-2_trimmed-magnet-domino.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small white plastic domino block is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0294_perspective-right_trimmed-magnet-domino.mp4
+0295_perspective-left_take-2_trimmed-magnet-transparent-peakaboo.mp4,A clear acrylic box suspended from a cord hangs above a tennis ball with a smiley face drawn on it positioned on a wooden table. The box is lowered to cover the ball. Static shot with no camera movement.,Solid Mechanics,0295_perspective-left_trimmed-magnet-transparent-peakaboo.mp4
+0296_perspective-center_take-2_trimmed-magnet-transparent-peakaboo.mp4,A clear acrylic box suspended from a cord hangs above a tennis ball with a smiley face drawn on it positioned on a wooden table. The box is lowered to cover the ball. Static shot with no camera movement.,Solid Mechanics,0296_perspective-center_trimmed-magnet-transparent-peakaboo.mp4
+0297_perspective-right_take-2_trimmed-magnet-transparent-peakaboo.mp4,A clear acrylic box suspended from a cord hangs above a tennis ball with a smiley face drawn on it positioned on a wooden table. The box is lowered to cover the ball. Static shot with no camera movement.,Solid Mechanics,0297_perspective-right_trimmed-magnet-transparent-peakaboo.mp4
+0298_perspective-left_take-2_trimmed-magnet-wrench.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small metal wrench is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0298_perspective-left_trimmed-magnet-wrench.mp4
+0299_perspective-center_take-2_trimmed-magnet-wrench.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small metal wrench is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0299_perspective-center_trimmed-magnet-wrench.mp4
+0300_perspective-right_take-2_trimmed-magnet-wrench.mp4,A powerful magnet is placed on the table facing a black rotating platform that rotates clockwise. A small metal wrench is placed on the platform and is rotating towards the magnet. Static shot with no camera movement.,Magnetism,0300_perspective-right_trimmed-magnet-wrench.mp4
+0301_perspective-left_take-2_trimmed-marble-run-x.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0301_perspective-left_trimmed-marble-run-x.mp4
+0302_perspective-center_take-2_trimmed-marble-run-x.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0302_perspective-center_trimmed-marble-run-x.mp4
+0303_perspective-right_take-2_trimmed-marble-run-x.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0303_perspective-right_trimmed-marble-run-x.mp4
+0304_perspective-left_take-2_trimmed-marble-run-y.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0304_perspective-left_trimmed-marble-run-y.mp4
+0305_perspective-center_take-2_trimmed-marble-run-y.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0305_perspective-center_trimmed-marble-run-y.mp4
+0306_perspective-right_take-2_trimmed-marble-run-y.mp4,A few magnetic ramps are attached to a whiteboard for a game of marble run. A yellow marble is released at the top of the ramps and slides down the ramps. Static shot with no camera movement.,Solid Mechanics,0306_perspective-right_trimmed-marble-run-y.mp4
+0307_perspective-left_take-2_trimmed-match.mp4,A lit match is being lowered into a glass of water. Static shot with no camera movement.,Fluid Dynamics,0307_perspective-left_trimmed-match.mp4
+0308_perspective-center_take-2_trimmed-match.mp4,A lit match is being lowered into a glass of water. Static shot with no camera movement.,Fluid Dynamics,0308_perspective-center_trimmed-match.mp4
+0309_perspective-right_take-2_trimmed-match.mp4,A lit match is being lowered into a glass of water. Static shot with no camera movement.,Fluid Dynamics,0309_perspective-right_trimmed-match.mp4
+0310_perspective-left_take-2_trimmed-match-blows-balloon.mp4,A black balloon is sitting on a wooden table next to a small rotating platform with a lit matchstick taped to it. The match rotates clockwise and touches the balloon. Static shot with no camera movement.,Thermodynamics,0310_perspective-left_trimmed-match-blows-balloon.mp4
+0311_perspective-center_take-2_trimmed-match-blows-balloon.mp4,A black balloon is sitting on a wooden table next to a small rotating platform with a lit matchstick taped to it. The match rotates clockwise and touches the balloon. Static shot with no camera movement.,Thermodynamics,0311_perspective-center_trimmed-match-blows-balloon.mp4
+0312_perspective-right_take-2_trimmed-match-blows-balloon.mp4,A black balloon is sitting on a wooden table next to a small rotating platform with a lit matchstick taped to it. The match rotates clockwise and touches the balloon. Static shot with no camera movement.,Thermodynamics,0312_perspective-right_trimmed-match-blows-balloon.mp4
+0313_perspective-left_take-2_trimmed-mirror-ball-fall.mp4,A tennis ball attached to a magnet and string is hanging in front of a mirror and creating an illusion of two tennis balls. The string lowers the ball slowly. Static shot with no camera movement.,Optics,0313_perspective-left_trimmed-mirror-ball-fall.mp4
+0314_perspective-center_take-2_trimmed-mirror-ball-fall.mp4,A tennis ball attached to a magnet and string is hanging in front of a mirror and creating an illusion of two tennis balls. The string lowers the ball slowly. Static shot with no camera movement.,Optics,0314_perspective-center_trimmed-mirror-ball-fall.mp4
+0315_perspective-right_take-2_trimmed-mirror-ball-fall.mp4,A tennis ball attached to a magnet and string is hanging in front of a mirror and creating an illusion of two tennis balls. The string lowers the ball slowly. Static shot with no camera movement.,Optics,0315_perspective-right_trimmed-mirror-ball-fall.mp4
+0316_perspective-left_take-2_trimmed-mirror-ball-rotate.mp4,A tennis ball with a smiley face drawn on it is slowly rotating on a black rotating platform that rotates clockwise in front of a mirror and reflecting the side of the ball that has no smiley face. Static shot with no camera movement.,Optics,0316_perspective-left_trimmed-mirror-ball-rotate.mp4
+0317_perspective-center_take-2_trimmed-mirror-ball-rotate.mp4,A tennis ball with a smiley face drawn on it is slowly rotating on a black rotating platform that rotates clockwise in front of a mirror and reflecting the side of the ball that has no smiley face. Static shot with no camera movement.,Optics,0317_perspective-center_trimmed-mirror-ball-rotate.mp4
+0318_perspective-right_take-2_trimmed-mirror-ball-rotate.mp4,A tennis ball with a smiley face drawn on it is slowly rotating on a black rotating platform that rotates clockwise in front of a mirror and reflecting the side of the ball that has no smiley face. Static shot with no camera movement.,Optics,0318_perspective-right_trimmed-mirror-ball-rotate.mp4
+0319_perspective-left_take-2_trimmed-mirror-teapot-rotate.mp4,A teapot on a rotating display base that rotates clockwise in front of a mirror reflecting the teapot's image. Static shot with no camera movement.,Optics,0319_perspective-left_trimmed-mirror-teapot-rotate.mp4
+0320_perspective-center_take-2_trimmed-mirror-teapot-rotate.mp4,A teapot on a rotating display base that rotates clockwise in front of a mirror reflecting the teapot's image. Static shot with no camera movement.,Optics,0320_perspective-center_trimmed-mirror-teapot-rotate.mp4
+0321_perspective-right_take-2_trimmed-mirror-teapot-rotate.mp4,A teapot on a rotating display base that rotates clockwise in front of a mirror reflecting the teapot's image. Static shot with no camera movement.,Optics,0321_perspective-right_trimmed-mirror-teapot-rotate.mp4
+0322_perspective-left_take-2_trimmed-mug-breaks.mp4,A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug. Static shot with no camera movement.,Solid Mechanics,0322_perspective-left_trimmed-mug-breaks.mp4
+0323_perspective-center_take-2_trimmed-mug-breaks.mp4,A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug. Static shot with no camera movement.,Solid Mechanics,0323_perspective-center_trimmed-mug-breaks.mp4
+0324_perspective-right_take-2_trimmed-mug-breaks.mp4,A yellow mug is held by a grabber tool in front of a white projection screen with a concrete brick positioned beneath it. The grabber releases the mug. Static shot with no camera movement.,Solid Mechanics,0324_perspective-right_trimmed-mug-breaks.mp4
+0325_perspective-left_take-2_trimmed-napkin-soak.mp4,A grabber tool holds a piece of paper towel over a shallow dish of light blue liquid on a wooden table. The grabber releases the paper towel on the dish. Static shot with no camera movement.,Fluid Dynamics,0325_perspective-left_trimmed-napkin-soak.mp4
+0326_perspective-center_take-2_trimmed-napkin-soak.mp4,A grabber tool holds a piece of paper towel over a shallow dish of light blue liquid on a wooden table. The grabber releases the paper towel on the dish. Static shot with no camera movement.,Fluid Dynamics,0326_perspective-center_trimmed-napkin-soak.mp4
+0327_perspective-right_take-2_trimmed-napkin-soak.mp4,A grabber tool holds a piece of paper towel over a shallow dish of light blue liquid on a wooden table. The grabber releases the paper towel on the dish. Static shot with no camera movement.,Fluid Dynamics,0327_perspective-right_trimmed-napkin-soak.mp4
+0328_perspective-left_take-2_trimmed-paint-on-glass.mp4,A clear acrylic sheet placed on a wooden table with a small dollop of red paint. A rotating paintbrush attached to a rotating platform rotates clockwise and goes through the paint. Static shot with no camera movement.,Fluid Dynamics,0328_perspective-left_trimmed-paint-on-glass.mp4
+0329_perspective-center_take-2_trimmed-paint-on-glass.mp4,A clear acrylic sheet placed on a wooden table with a small dollop of red paint. A rotating paintbrush attached to a rotating platform rotates clockwise and goes through the paint. Static shot with no camera movement.,Fluid Dynamics,0329_perspective-center_trimmed-paint-on-glass.mp4
+0330_perspective-right_take-2_trimmed-paint-on-glass.mp4,A clear acrylic sheet placed on a wooden table with a small dollop of red paint. A rotating paintbrush attached to a rotating platform rotates clockwise and goes through the paint. Static shot with no camera movement.,Fluid Dynamics,0330_perspective-right_trimmed-paint-on-glass.mp4
+0331_perspective-left_take-2_trimmed-paper-fall-water.mp4,A grabber tool is holding a crumpled piece of paper over a bowl of water on a wooden table. The grabber then releases the crumpled paper onto the bowl. Static shot with no camera movement.,Fluid Dynamics,0331_perspective-left_trimmed-paper-fall-water.mp4
+0332_perspective-center_take-2_trimmed-paper-fall-water.mp4,A grabber tool is holding a crumpled piece of paper over a bowl of water on a wooden table. The grabber then releases the crumpled paper onto the bowl. Static shot with no camera movement.,Fluid Dynamics,0332_perspective-center_trimmed-paper-fall-water.mp4
+0333_perspective-right_take-2_trimmed-paper-fall-water.mp4,A grabber tool is holding a crumpled piece of paper over a bowl of water on a wooden table. The grabber then releases the crumpled paper onto the bowl. Static shot with no camera movement.,Fluid Dynamics,0333_perspective-right_trimmed-paper-fall-water.mp4
+0334_perspective-left_take-2_trimmed-paper-in-water.mp4,A small piece of crumpled white paper is being lowered into a tall glass containing blue liquid with a green band showing the water level. The crumpled paper is released into the glass. Static shot with no camera movement.,Fluid Dynamics,0334_perspective-left_trimmed-paper-in-water.mp4
+0335_perspective-center_take-2_trimmed-paper-in-water.mp4,A small piece of crumpled white paper is being lowered into a tall glass containing blue liquid with a green band showing the water level. The crumpled paper is released into the glass. Static shot with no camera movement.,Fluid Dynamics,0335_perspective-center_trimmed-paper-in-water.mp4
+0336_perspective-right_take-2_trimmed-paper-in-water.mp4,A small piece of crumpled white paper is being lowered into a tall glass containing blue liquid with a green band showing the water level. The crumpled paper is released into the glass. Static shot with no camera movement.,Fluid Dynamics,0336_perspective-right_trimmed-paper-in-water.mp4
+0337_perspective-left_take-2_trimmed-paper-smoke.mp4,A piece of folded paper is placed on a glass cutting board. The paper is being burnt and white smoke is emitting from it. Static shot with no camera movement.,Thermodynamics,0337_perspective-left_trimmed-paper-smoke.mp4
+0338_perspective-center_take-2_trimmed-paper-smoke.mp4,A piece of folded paper is placed on a glass cutting board. The paper is being burnt and white smoke is emitting from it. Static shot with no camera movement.,Thermodynamics,0338_perspective-center_trimmed-paper-smoke.mp4
+0339_perspective-right_take-2_trimmed-paper-smoke.mp4,A piece of folded paper is placed on a glass cutting board. The paper is being burnt and white smoke is emitting from it. Static shot with no camera movement.,Thermodynamics,0339_perspective-right_trimmed-paper-smoke.mp4
+0340_perspective-left_take-2_trimmed-potato-in-water.mp4,A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass. Static shot with no camera movement.,Fluid Dynamics,0340_perspective-left_trimmed-potato-in-water.mp4
+0341_perspective-center_take-2_trimmed-potato-in-water.mp4,A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass. Static shot with no camera movement.,Fluid Dynamics,0341_perspective-center_trimmed-potato-in-water.mp4
+0342_perspective-right_take-2_trimmed-potato-in-water.mp4,A potato is held by a grabber tool and dropped into a tall glass containing blue liquid with a band of green tape marking a level on the glass. Static shot with no camera movement.,Fluid Dynamics,0342_perspective-right_trimmed-potato-in-water.mp4
+0343_perspective-left_take-2_trimmed-roll-behind-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0343_perspective-left_trimmed-roll-behind-box.mp4
+0344_perspective-center_take-2_trimmed-roll-behind-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0344_perspective-center_trimmed-roll-behind-box.mp4
+0345_perspective-right_take-2_trimmed-roll-behind-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0345_perspective-right_trimmed-roll-behind-box.mp4
+0346_perspective-left_take-2_trimmed-roll-front-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0346_perspective-left_trimmed-roll-front-box.mp4
+0347_perspective-center_take-2_trimmed-roll-front-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0347_perspective-center_trimmed-roll-front-box.mp4
+0348_perspective-right_take-2_trimmed-roll-front-box.mp4,A small white lampshade is on a light wood surface. A grey tennis ball rolls out of the black tube sitting on the table and rolls on the table towards the right. Static shot with no camera movement.,Solid Mechanics,0348_perspective-right_trimmed-roll-front-box.mp4
+0349_perspective-left_take-2_trimmed-roll-in-box.mp4,An olive green fabric box is on a light wood surface. A brown tennis ball rolls out of the black tube sitting on the table and rolls towards the box. Static shot with no camera movement.,Solid Mechanics,0349_perspective-left_trimmed-roll-in-box.mp4
+0350_perspective-center_take-2_trimmed-roll-in-box.mp4,An olive green fabric box is on a light wood surface. A brown tennis ball rolls out of the black tube sitting on the table and rolls towards the box. Static shot with no camera movement.,Solid Mechanics,0350_perspective-center_trimmed-roll-in-box.mp4
+0351_perspective-right_take-2_trimmed-roll-in-box.mp4,An olive green fabric box is on a light wood surface. A brown tennis ball rolls out of the black tube sitting on the table and rolls towards the box. Static shot with no camera movement.,Solid Mechanics,0351_perspective-right_trimmed-roll-in-box.mp4
+0352_perspective-left_take-2_trimmed-rolling-reflection.mp4,A 30lb kettlebell resting on a wooden table next to a mirror. A tennis ball rolls towards the kettlebell. Static shot with no camera movement.,Optics,0352_perspective-left_trimmed-rolling-reflection.mp4
+0353_perspective-center_take-2_trimmed-rolling-reflection.mp4,A 30lb kettlebell resting on a wooden table next to a mirror. A tennis ball rolls towards the kettlebell. Static shot with no camera movement.,Optics,0353_perspective-center_trimmed-rolling-reflection.mp4
+0354_perspective-right_take-2_trimmed-rolling-reflection.mp4,A 30lb kettlebell resting on a wooden table next to a mirror. A tennis ball rolls towards the kettlebell. Static shot with no camera movement.,Optics,0354_perspective-right_trimmed-rolling-reflection.mp4
+0355_perspective-left_take-2_trimmed-silk-cover.mp4,A teapot is placed on a wooden table. a piece of silk fabric is lowered on the teapot to cover it. Static shot with no camera movement.,Solid Mechanics,0355_perspective-left_trimmed-silk-cover.mp4
+0356_perspective-center_take-2_trimmed-silk-cover.mp4,A teapot is placed on a wooden table. a piece of silk fabric is lowered on the teapot to cover it. Static shot with no camera movement.,Solid Mechanics,0356_perspective-center_trimmed-silk-cover.mp4
+0357_perspective-right_take-2_trimmed-silk-cover.mp4,A teapot is placed on a wooden table. a piece of silk fabric is lowered on the teapot to cover it. Static shot with no camera movement.,Solid Mechanics,0357_perspective-right_trimmed-silk-cover.mp4
+0358_perspective-left_take-2_trimmed-single-cradle.mp4,A Newton's cradle device on the table and one of the metal balls is held up by a blue handled grabber tool. The claw releases the ball. Static shot with no camera movement.,Solid Mechanics,0358_perspective-left_trimmed-single-cradle.mp4
+0359_perspective-center_take-2_trimmed-single-cradle.mp4,A Newton's cradle device on the table and one of the metal balls is held up by a blue handled grabber tool. The claw releases the ball. Static shot with no camera movement.,Solid Mechanics,0359_perspective-center_trimmed-single-cradle.mp4
+0360_perspective-right_take-2_trimmed-single-cradle.mp4,A Newton's cradle device on the table and one of the metal balls is held up by a blue handled grabber tool. The claw releases the ball. Static shot with no camera movement.,Solid Mechanics,0360_perspective-right_trimmed-single-cradle.mp4
+0361_perspective-left_take-2_trimmed-siphon.mp4,A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks. Static shot with no camera movement.,Fluid Dynamics,0361_perspective-left_trimmed-siphon.mp4
+0362_perspective-center_take-2_trimmed-siphon.mp4,A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks. Static shot with no camera movement.,Fluid Dynamics,0362_perspective-center_trimmed-siphon.mp4
+0363_perspective-right_take-2_trimmed-siphon.mp4,A bundle of lit matchsticks is placed in a bowl of red liquid. A glass jar gets lowered and covers the matchsticks. Static shot with no camera movement.,Fluid Dynamics,0363_perspective-right_trimmed-siphon.mp4
+0364_perspective-left_take-2_trimmed-smiley-ball-rotates.mp4,A tennis ball with a smiley face drawn on it is placed on a rotating black platform that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0364_perspective-left_trimmed-smiley-ball-rotates.mp4
+0365_perspective-center_take-2_trimmed-smiley-ball-rotates.mp4,A tennis ball with a smiley face drawn on it is placed on a rotating black platform that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0365_perspective-center_trimmed-smiley-ball-rotates.mp4
+0366_perspective-right_take-2_trimmed-smiley-ball-rotates.mp4,A tennis ball with a smiley face drawn on it is placed on a rotating black platform that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0366_perspective-right_trimmed-smiley-ball-rotates.mp4
+0367_perspective-left_take-2_trimmed-solid-ball-peakaboo.mp4,A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again. Static shot with no camera movement.,Solid Mechanics,0367_perspective-left_trimmed-solid-ball-peakaboo.mp4
+0368_perspective-center_take-2_trimmed-solid-ball-peakaboo.mp4,A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again. Static shot with no camera movement.,Solid Mechanics,0368_perspective-center_trimmed-solid-ball-peakaboo.mp4
+0369_perspective-right_take-2_trimmed-solid-ball-peakaboo.mp4,A woven basket is hanging from a rope with a strong magnet attached to the bottom. An orange tennis ball is placed on a table beneath it. The basket is lowered and covers the ball and then the basket starts to lift again. Static shot with no camera movement.,Solid Mechanics,0369_perspective-right_trimmed-solid-ball-peakaboo.mp4
+0370_perspective-left_take-2_trimmed-stable-blocks.mp4,A pink block is being lowered towards a simple structure made of colorful blocks resembling a gate. Static shot with no camera movement.,Solid Mechanics,0370_perspective-left_trimmed-stable-blocks.mp4
+0371_perspective-center_take-2_trimmed-stable-blocks.mp4,A pink block is being lowered towards a simple structure made of colorful blocks resembling a gate. Static shot with no camera movement.,Solid Mechanics,0371_perspective-center_trimmed-stable-blocks.mp4
+0372_perspective-right_take-2_trimmed-stable-blocks.mp4,A pink block is being lowered towards a simple structure made of colorful blocks resembling a gate. Static shot with no camera movement.,Solid Mechanics,0372_perspective-right_trimmed-stable-blocks.mp4
+0373_perspective-left_take-2_trimmed-teapot-rotates.mp4,A teapot is placed on a rotating display that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0373_perspective-left_trimmed-teapot-rotates.mp4
+0374_perspective-center_take-2_trimmed-teapot-rotates.mp4,A teapot is placed on a rotating display that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0374_perspective-center_trimmed-teapot-rotates.mp4
+0375_perspective-right_take-2_trimmed-teapot-rotates.mp4,A teapot is placed on a rotating display that rotates clockwise. Static shot with no camera movement.,Solid Mechanics,0375_perspective-right_trimmed-teapot-rotates.mp4
+0376_perspective-left_take-2_trimmed-two-balls-pass.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0376_perspective-left_trimmed-two-balls-pass.mp4
+0377_perspective-center_take-2_trimmed-two-balls-pass.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0377_perspective-center_trimmed-two-balls-pass.mp4
+0378_perspective-right_take-2_trimmed-two-balls-pass.mp4,A light-colored wooden tabletop with two pipes at the edges. A blue and yellow tennis ball roll out of the pipes and towards each other. Static shot with no camera movement.,Solid Mechanics,0378_perspective-right_trimmed-two-balls-pass.mp4
+0379_perspective-left_take-2_trimmed-unstable-block-stack.mp4,A grabber tool carefully placing a blue wooden block on top of a yellow block which is balanced on a red block forming an L shape. Static shot with no camera movement.,Solid Mechanics,0379_perspective-left_trimmed-unstable-block-stack.mp4
+0380_perspective-center_take-2_trimmed-unstable-block-stack.mp4,A grabber tool carefully placing a blue wooden block on top of a yellow block which is balanced on a red block forming an L shape. Static shot with no camera movement.,Solid Mechanics,0380_perspective-center_trimmed-unstable-block-stack.mp4
+0381_perspective-right_take-2_trimmed-unstable-block-stack.mp4,A grabber tool carefully placing a blue wooden block on top of a yellow block which is balanced on a red block forming an L shape. Static shot with no camera movement.,Solid Mechanics,0381_perspective-right_trimmed-unstable-block-stack.mp4
+0382_perspective-left_take-2_trimmed-water-in-juice.mp4,A glass beverage dispenser is dispensing water into a glass which has some grapefruit juice in it. Static shot with no camera movement.,Fluid Dynamics,0382_perspective-left_trimmed-water-in-juice.mp4
+0383_perspective-center_take-2_trimmed-water-in-juice.mp4,A glass beverage dispenser is dispensing water into a glass which has some grapefruit juice in it. Static shot with no camera movement.,Fluid Dynamics,0383_perspective-center_trimmed-water-in-juice.mp4
+0384_perspective-right_take-2_trimmed-water-in-juice.mp4,A glass beverage dispenser is dispensing water into a glass which has some grapefruit juice in it. Static shot with no camera movement.,Fluid Dynamics,0384_perspective-right_trimmed-water-in-juice.mp4
+0385_perspective-left_take-2_trimmed-weight-on-ceramic.mp4,A 30lb kettlebell is slowly lowered on top of a yellow ceramic coffee mug placed on a wooden table. Static shot with no camera movement.,Solid Mechanics,0385_perspective-left_trimmed-weight-on-ceramic.mp4
+0386_perspective-center_take-2_trimmed-weight-on-ceramic.mp4,A 30lb kettlebell is slowly lowered on top of a yellow ceramic coffee mug placed on a wooden table. Static shot with no camera movement.,Solid Mechanics,0386_perspective-center_trimmed-weight-on-ceramic.mp4
+0387_perspective-right_take-2_trimmed-weight-on-ceramic.mp4,A 30lb kettlebell is slowly lowered on top of a yellow ceramic coffee mug placed on a wooden table. Static shot with no camera movement.,Solid Mechanics,0387_perspective-right_trimmed-weight-on-ceramic.mp4
+0388_perspective-left_take-2_trimmed-weight-on-paper.mp4,A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.,Solid Mechanics,0388_perspective-left_trimmed-weight-on-paper.mp4
+0389_perspective-center_take-2_trimmed-weight-on-paper.mp4,A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.,Solid Mechanics,0389_perspective-center_trimmed-weight-on-paper.mp4
+0390_perspective-right_take-2_trimmed-weight-on-paper.mp4,A 30lb kettlebell is slowly lowered onto a white styrofoam cup placed on a wooden table on its side. Static shot with no camera movement.,Solid Mechanics,0390_perspective-right_trimmed-weight-on-paper.mp4
+0391_perspective-left_take-2_trimmed-weight-on-pillow.mp4,A 30lb kettlebell and a green piece of paper are lowered onto two pillows. Static shot with no camera movement.,Solid Mechanics,0391_perspective-left_trimmed-weight-on-pillow.mp4
+0392_perspective-center_take-2_trimmed-weight-on-pillow.mp4,A 30lb kettlebell and a green piece of paper are lowered onto two pillows. Static shot with no camera movement.,Solid Mechanics,0392_perspective-center_trimmed-weight-on-pillow.mp4
+0393_perspective-right_take-2_trimmed-weight-on-pillow.mp4,A 30lb kettlebell and a green piece of paper are lowered onto two pillows. Static shot with no camera movement.,Solid Mechanics,0393_perspective-right_trimmed-weight-on-pillow.mp4
+0394_perspective-left_take-2_trimmed-weight-protects-duck.mp4,A light beige coffee table with a black kettlebell and a yellow rubber duck on it. A grey tennis ball rolls out of the black tube sitting on the table and towards the duck and kettlebell. Static shot with no camera movement.,Solid Mechanics,0394_perspective-left_trimmed-weight-protects-duck.mp4
+0395_perspective-center_take-2_trimmed-weight-protects-duck.mp4,A light beige coffee table with a black kettlebell and a yellow rubber duck on it. A grey tennis ball rolls out of the black tube sitting on the table and towards the duck and kettlebell. Static shot with no camera movement.,Solid Mechanics,0395_perspective-center_trimmed-weight-protects-duck.mp4
+0396_perspective-right_take-2_trimmed-weight-protects-duck.mp4,A light beige coffee table with a black kettlebell and a yellow rubber duck on it. A grey tennis ball rolls out of the black tube sitting on the table and towards the duck and kettlebell. Static shot with no camera movement.,Solid Mechanics,0396_perspective-right_trimmed-weight-protects-duck.mp4

fastvideo/eval/metrics/physics_iq/metric.py

@@ -0,0 +1,194 @@
+from __future__ import annotations
+
+from typing import Any
+from collections.abc import Iterable, Mapping
+
+import numpy as np
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.metrics.physics_iq.mse.metric import PhysicsIQMSEMetric
+from fastvideo.eval.metrics.physics_iq.spatial_iou.metric import SpatialIoUMetric
+from fastvideo.eval.metrics.physics_iq.spatiotemporal_iou.metric import SpatiotemporalIoUMetric
+from fastvideo.eval.metrics.physics_iq.weighted_spatial_iou.metric import WeightedSpatialIoUMetric
+from fastvideo.eval.metrics.physics_iq.utils import (
+    DEFAULT_DURATION_SECONDS,
+    DEFAULT_TARGET_FPS,
+    mean,
+    prepare_pair_inputs,
+    prepare_triplet_inputs,
+)
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("physics_iq")
+class PhysicsIQMetric(BaseMetric):
+    name = "physics_iq"
+    requires_reference = True
+    higher_is_better = True
+
+    def __init__(
+        self,
+        *,
+        target_fps: int = DEFAULT_TARGET_FPS,
+        duration_seconds: int = DEFAULT_DURATION_SECONDS,
+        video_time_selection: str = "first",
+        threshold: int = 10,
+        alpha: float = 0.3,
+        roundtrip_generated_masks: bool = True,
+    ) -> None:
+        super().__init__()
+        self._prep_kwargs = {
+            "target_fps": target_fps,
+            "duration_seconds": duration_seconds,
+            "video_time_selection": video_time_selection,
+            "threshold": threshold,
+            "alpha": alpha,
+            "roundtrip_generated_masks": roundtrip_generated_masks,
+        }
+        self._mse = PhysicsIQMSEMetric(**self._prep_kwargs)
+        self._spatiotemporal_iou = SpatiotemporalIoUMetric(**self._prep_kwargs)
+        self._spatial_iou = SpatialIoUMetric(**self._prep_kwargs)
+        self._weighted_spatial_iou = WeightedSpatialIoUMetric(**self._prep_kwargs)
+
+    @staticmethod
+    def _extract_payload(result: MetricResult | Mapping[str, Any]) -> Mapping[str, Any]:
+        if isinstance(result, MetricResult):
+            return result.details
+        return result
+
+    def _compute_pair_metrics(self, prepared_pair) -> dict[str, Any]:
+        sample = {"_physics_iq_pair": prepared_pair}
+        mse = self._mse.compute(sample)
+        st = self._spatiotemporal_iou.compute(sample)
+        spatial = self._spatial_iou.compute(sample)
+        weighted = self._weighted_spatial_iou.compute(sample)
+        return {
+            "mse_per_frame": mse.details["per_frame"],
+            "spatiotemporal_iou_per_frame": st.details["per_frame"],
+            "spatial_iou": float(spatial.score),
+            "weighted_spatial_iou": float(weighted.score),
+            "mse_mean": float(mse.score),
+            "spatiotemporal_iou_mean": float(st.score),
+        }
+
+    def compute_single(
+        self,
+        generated: Any,
+        reference: Any,
+        reference_take2: Any,
+        *,
+        generated_mask: Any | None = None,
+        reference_mask: Any | None = None,
+        reference_take2_mask: Any | None = None,
+        scenario: str | None = None,
+        view: str | None = None,
+    ) -> dict[str, Any]:
+        prepared = prepare_triplet_inputs(
+            generated,
+            reference,
+            reference_take2,
+            generated_mask=generated_mask,
+            reference_mask=reference_mask,
+            reference_take2_mask=reference_take2_mask,
+            **self._prep_kwargs,
+        )
+        pair_metrics = self._compute_pair_metrics(prepared)
+        variance_pair = prepare_pair_inputs(
+            reference,
+            reference_take2,
+            generated_mask=reference_mask,
+            reference_mask=reference_take2_mask,
+            **self._prep_kwargs,
+        )
+        variance_metrics = self._compute_pair_metrics(variance_pair)
+        details = {
+            **pair_metrics,
+            "pv_mse_per_frame": variance_metrics["mse_per_frame"],
+            "pv_spatiotemporal_iou_per_frame": variance_metrics["spatiotemporal_iou_per_frame"],
+            "pv_spatial_iou": variance_metrics["spatial_iou"],
+            "pv_weighted_spatial_iou": variance_metrics["weighted_spatial_iou"],
+            "pv_mse_mean": variance_metrics["mse_mean"],
+            "pv_spatiotemporal_iou_mean": variance_metrics["spatiotemporal_iou_mean"],
+        }
+        if scenario is not None:
+            details["scenario"] = scenario
+        if view is not None:
+            details["view"] = view
+        return details
+
+    @classmethod
+    def aggregate(cls, results_list: Iterable[MetricResult | Mapping[str, Any]]) -> float:
+        payloads = [cls._extract_payload(result) for result in results_list]
+        if not payloads:
+            raise ValueError("PhysicsIQMetric.aggregate requires at least one result.")
+
+        a_mse = mean([value for payload in payloads for value in payload["mse_per_frame"]])
+        a_st = mean([value for payload in payloads for value in payload["spatiotemporal_iou_per_frame"]])
+        a_s = mean([float(payload["spatial_iou"]) for payload in payloads])
+        a_ws = mean([float(payload["weighted_spatial_iou"]) for payload in payloads])
+
+        v_mse = mean([value for payload in payloads for value in payload["pv_mse_per_frame"]])
+        v_st = mean([value for payload in payloads for value in payload["pv_spatiotemporal_iou_per_frame"]])
+        v_s = mean([float(payload["pv_spatial_iou"]) for payload in payloads])
+        v_ws = mean([float(payload["pv_weighted_spatial_iou"]) for payload in payloads])
+
+        score = 100.0 * ((((a_st / v_st) + (a_s / v_s) + (a_ws / v_ws)) / 3.0) - (a_mse - v_mse))
+        return round(float(np.clip(score, 0.0, 100.0)), 2)
+
+    @classmethod
+    def aggregate_components(cls, results_list: Iterable[MetricResult | Mapping[str, Any]]) -> dict[str, float]:
+        payloads = [cls._extract_payload(result) for result in results_list]
+        return {
+            "physics_iq": cls.aggregate(payloads),
+            "a_mse": mean([value for payload in payloads for value in payload["mse_per_frame"]]),
+            "a_st": mean([value for payload in payloads for value in payload["spatiotemporal_iou_per_frame"]]),
+            "a_s": mean([float(payload["spatial_iou"]) for payload in payloads]),
+            "a_ws": mean([float(payload["weighted_spatial_iou"]) for payload in payloads]),
+            "v_mse": mean([value for payload in payloads for value in payload["pv_mse_per_frame"]]),
+            "v_st": mean([value for payload in payloads for value in payload["pv_spatiotemporal_iou_per_frame"]]),
+            "v_s": mean([float(payload["pv_spatial_iou"]) for payload in payloads]),
+            "v_ws": mean([float(payload["pv_weighted_spatial_iou"]) for payload in payloads]),
+        }
+
+    def _per_video_score(self, details: Mapping[str, Any]) -> float:
+        score = 100.0 * (
+            ((mean(details["spatiotemporal_iou_per_frame"]) / mean(details["pv_spatiotemporal_iou_per_frame"])) +
+             (float(details["spatial_iou"]) / float(details["pv_spatial_iou"])) +
+             (float(details["weighted_spatial_iou"]) / float(details["pv_weighted_spatial_iou"]))) / 3.0 -
+            (mean(details["mse_per_frame"]) - mean(details["pv_mse_per_frame"])))
+        return round(float(np.clip(score, 0.0, 100.0)), 2)
+
+    def compute(self, sample: dict) -> MetricResult:
+        if "reference" not in sample:
+            raise KeyError("PhysicsIQMetric requires sample['reference'].")
+
+        take2_key = None
+        for candidate in ("reference_take2", "real_take2", "take2"):
+            if candidate in sample:
+                take2_key = candidate
+                break
+        if take2_key is None:
+            raise KeyError("PhysicsIQMetric requires sample['reference_take2'] or an alias.")
+
+        video = sample["video"]
+        reference = sample["reference"]
+        reference_take2 = sample[take2_key]
+        generated_mask = sample.get("video_mask")
+        reference_mask = sample.get("reference_mask")
+        reference_take2_mask = sample.get("reference_take2_mask")
+
+        scenario = sample.get("scenario")
+        view = sample.get("view")
+
+        details = self.compute_single(
+            video,
+            reference,
+            reference_take2,
+            generated_mask=generated_mask,
+            reference_mask=reference_mask,
+            reference_take2_mask=reference_take2_mask,
+            scenario=scenario,
+            view=view,
+        )
+        return MetricResult(name=self.name, score=self._per_video_score(details), details=details)

fastvideo/eval/metrics/physics_iq/mse/metric.py

@@ -0,0 +1,25 @@
+from __future__ import annotations
+
+from typing import Any
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.metrics.physics_iq.utils import compute_mse, prepare_pair
+
+
+@register("physics_iq.mse")
+class PhysicsIQMSEMetric(BaseMetric):
+    name = "physics_iq.mse"
+    requires_reference = True
+    higher_is_better = False
+
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__()
+        self._kwargs = kwargs
+
+    def compute(self, sample: dict) -> MetricResult:
+        prepared = prepare_pair(sample, prep_kwargs=self._kwargs)
+        per_frame = compute_mse(prepared.reference_quarter, prepared.generated_quarter)
+        score = sum(per_frame) / len(per_frame)
+        return MetricResult(name=self.name, score=score, details={"per_frame": per_frame})

fastvideo/eval/metrics/physics_iq/spatial_iou/metric.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from typing import Any
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.metrics.physics_iq.utils import compute_spatial_iou, prepare_pair
+
+
+@register("physics_iq.spatial_iou")
+class SpatialIoUMetric(BaseMetric):
+    name = "physics_iq.spatial_iou"
+    requires_reference = True
+    higher_is_better = True
+
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__()
+        self._kwargs = kwargs
+
+    def compute(self, sample: dict) -> MetricResult:
+        prepared = prepare_pair(sample, prep_kwargs=self._kwargs)
+        score = compute_spatial_iou(prepared.reference_masks, prepared.generated_masks)
+        return MetricResult(name=self.name, score=score, details={})

fastvideo/eval/metrics/physics_iq/spatiotemporal_iou/metric.py

@@ -0,0 +1,25 @@
+from __future__ import annotations
+
+from typing import Any
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.metrics.physics_iq.utils import compute_spatiotemporal_iou, prepare_pair
+
+
+@register("physics_iq.spatiotemporal_iou")
+class SpatiotemporalIoUMetric(BaseMetric):
+    name = "physics_iq.spatiotemporal_iou"
+    requires_reference = True
+    higher_is_better = True
+
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__()
+        self._kwargs = kwargs
+
+    def compute(self, sample: dict) -> MetricResult:
+        prepared = prepare_pair(sample, prep_kwargs=self._kwargs)
+        per_frame = compute_spatiotemporal_iou(prepared.reference_masks, prepared.generated_masks)
+        score = sum(per_frame) / len(per_frame)
+        return MetricResult(name=self.name, score=score, details={"per_frame": per_frame})

fastvideo/eval/metrics/physics_iq/utils.py

@@ -0,0 +1,420 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+import os
+import tempfile
+
+import cv2
+import numpy as np
+import torch
+
+# Default sampling configuration for the Physics-IQ comparison pipeline.
+# Source release ships at 30 FPS / 5 seconds — the metric collapses to those
+# anchors regardless of how the user resampled the input video.
+DEFAULT_TARGET_FPS = 30
+DEFAULT_DURATION_SECONDS = 5
+
+
+@dataclass(frozen=True)
+class PreparedPhysicsIQPair:
+    generated_quarter: np.ndarray
+    reference_quarter: np.ndarray
+    generated_masks: np.ndarray
+    reference_masks: np.ndarray
+
+
+@dataclass(frozen=True)
+class PreparedPhysicsIQTriplet:
+    generated_quarter: np.ndarray
+    reference_quarter: np.ndarray
+    reference_take2_quarter: np.ndarray
+    generated_masks: np.ndarray
+    reference_masks: np.ndarray
+    reference_take2_masks: np.ndarray
+
+
+def tensor_to_uint8_frames(video: torch.Tensor) -> np.ndarray:
+    arr = video.detach().cpu().float().clamp(0, 1).permute(0, 2, 3, 1).numpy()
+    return np.clip(np.rint(arr * 255.0), 0, 255).astype(np.uint8)
+
+
+def read_video_frames(
+    source: str | Path,
+    *,
+    start_frame: int = 0,
+    end_frame: int | None = None,
+) -> np.ndarray:
+    cap = cv2.VideoCapture(str(source))
+    if not cap.isOpened():
+        raise FileNotFoundError(f"Could not open video: {source}")
+
+    frames: list[np.ndarray] = []
+    frame_idx = 0
+    while cap.isOpened():
+        ret, frame = cap.read()
+        if not ret:
+            break
+        if frame_idx >= start_frame and (end_frame is None or frame_idx < end_frame):
+            frames.append(frame)
+        if end_frame is not None and frame_idx >= end_frame:
+            break
+        frame_idx += 1
+    cap.release()
+
+    if not frames:
+        return np.zeros((0, 0, 0, 3), dtype=np.uint8)
+    return np.stack(frames, axis=0)
+
+
+def as_numpy_video(source: Any) -> tuple[np.ndarray, str]:
+    if isinstance(source, torch.Tensor):
+        if source.ndim != 4:
+            raise ValueError(f"Expected 4D video tensor (T,C,H,W), got shape {tuple(source.shape)}")
+        return tensor_to_uint8_frames(source), "rgb"
+    if isinstance(source, np.ndarray):
+        if source.ndim != 4:
+            raise ValueError(f"Expected 4D ndarray video, got shape {source.shape}")
+        if source.shape[-1] == 3:
+            return source.astype(np.uint8), "rgb"
+        if source.shape[1] == 3:
+            return np.transpose(source, (0, 2, 3, 1)).astype(np.uint8), "rgb"
+        raise ValueError(f"Unsupported ndarray video shape: {source.shape}")
+    if isinstance(source, str | Path):
+        return read_video_frames(source), "bgr"
+    raise TypeError(f"Unsupported Physics-IQ video source type: {type(source)!r}")
+
+
+def prepare_pair(
+    sample: dict[str, Any],
+    *,
+    prep_kwargs: dict[str, Any] | None = None,
+) -> PreparedPhysicsIQPair:
+    """Resolve a sample into a prepared (gen, ref) pair.
+
+    Caches the result on ``sample['_physics_iq_pair']`` so other physics_iq
+    sub-metrics on the same sample reuse it instead of re-decoding.
+    """
+    prepared = sample.get("_physics_iq_pair")
+    if prepared is not None:
+        return prepared
+
+    if "reference" not in sample:
+        raise KeyError("Physics-IQ pair metrics require sample['reference'].")
+
+    return prepare_pair_inputs(
+        sample["video"],
+        sample["reference"],
+        generated_mask=sample.get("video_mask"),
+        reference_mask=sample.get("reference_mask"),
+        **(prep_kwargs or {}),
+    )
+
+
+def select_window(frames: np.ndarray, *, target_frames: int, selection: str = "first") -> np.ndarray:
+    if selection != "first":
+        start = max(frames.shape[0] - target_frames, 0)
+        return frames[start:start + target_frames]
+    return frames[:target_frames]
+
+
+def resize_frames(frames: np.ndarray, target_size: tuple[int, int]) -> np.ndarray:
+    if frames.size == 0:
+        return frames
+    resized = [cv2.resize(frame, target_size) for frame in frames]
+    return np.stack(resized, axis=0)
+
+
+def rebinarize_masks(mask_frames: np.ndarray) -> np.ndarray:
+    if mask_frames.ndim == 4 and mask_frames.shape[-1] == 3:
+        mask_frames = mask_frames[..., 0]
+    return (mask_frames > 127).astype(np.uint8)
+
+
+def load_mask_frames(
+    mask_source: Any,
+    *,
+    target_frames: int,
+    target_size: tuple[int, int],
+) -> np.ndarray:
+    if mask_source is None:
+        raise ValueError("mask_source cannot be None when loading mask frames")
+    mask_frames, _ = as_numpy_video(mask_source)
+    mask_frames = select_window(mask_frames, target_frames=target_frames, selection="first")
+    mask_frames = resize_frames(mask_frames, target_size)
+    return rebinarize_masks(mask_frames)
+
+
+def roundtrip_mask_frames(mask_frames: np.ndarray, *, fps: int) -> np.ndarray:
+    if mask_frames.size == 0:
+        return mask_frames
+    fd, tmp_path = tempfile.mkstemp(suffix=".mp4")
+    os.close(fd)
+    output_path = Path(tmp_path)
+    output_path.unlink(missing_ok=True)
+    try:
+        writer = cv2.VideoWriter(
+            str(output_path),
+            cv2.VideoWriter_fourcc(*"mp4v"),
+            fps,
+            (mask_frames.shape[2], mask_frames.shape[1]),
+            isColor=False,
+        )
+        for frame in mask_frames:
+            writer.write(frame)
+        writer.release()
+        return read_video_frames(output_path)
+    finally:
+        output_path.unlink(missing_ok=True)
+
+
+def infer_real_mask_path(video_source: Any) -> str | None:
+    if not isinstance(video_source, str | Path):
+        return None
+
+    video_path = Path(video_source)
+    filename = video_path.name
+    if "_testing-videos_" not in filename:
+        return None
+
+    mask_name = filename.replace("_testing-videos_", "_video-masks_")
+    candidates: list[Path] = []
+    fps_dir = video_path.parent.name
+    for ancestor in video_path.parents:
+        if ancestor.name == "split-videos":
+            candidates.extend([
+                ancestor.parent / "video-masks" / "real" / fps_dir / mask_name,
+                ancestor.parent / "video_masks" / "real" / fps_dir / mask_name,
+            ])
+            break
+
+    for candidate in candidates:
+        if candidate.exists():
+            return str(candidate)
+    return None
+
+
+def prepare_grayscale_frame(frame: np.ndarray, *, color_order: str) -> np.ndarray:
+    if frame.ndim == 2:
+        gray = frame
+    elif color_order == "bgr":
+        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+    elif color_order == "rgb":
+        gray = cv2.cvtColor(frame, cv2.COLOR_RGB2GRAY)
+    else:
+        raise ValueError(f"Unsupported color order: {color_order}")
+    return cv2.GaussianBlur(gray, (5, 5), 0)
+
+
+def generate_motion_mask(
+    video_frames: np.ndarray,
+    *,
+    threshold: int = 10,
+    alpha: float = 0.3,
+    color_order: str = "rgb",
+) -> np.ndarray:
+    if video_frames.size == 0:
+        return np.zeros((0, 0, 0), dtype=np.uint8)
+
+    first_gray = prepare_grayscale_frame(video_frames[0], color_order=color_order)
+    avg_frame = first_gray.astype("float")
+    masks = [np.zeros_like(first_gray, dtype=np.uint8)]
+    kernel = np.ones((5, 5), np.uint8)
+
+    for frame in video_frames[1:]:
+        gray_frame = prepare_grayscale_frame(frame, color_order=color_order)
+        cv2.accumulateWeighted(gray_frame, avg_frame, alpha)
+        avg_gray_frame = cv2.convertScaleAbs(avg_frame)
+        frame_diff = cv2.absdiff(gray_frame, avg_gray_frame)
+        _, binary_frame = cv2.threshold(frame_diff, threshold, 255, cv2.THRESH_BINARY)
+        binary_frame = cv2.morphologyEx(binary_frame, cv2.MORPH_OPEN, kernel)
+        binary_frame = cv2.morphologyEx(binary_frame, cv2.MORPH_CLOSE, kernel)
+        masks.append(binary_frame)
+    return np.stack(masks, axis=0)
+
+
+def compute_iou(mask1: np.ndarray, mask2: np.ndarray) -> float:
+    intersection = np.logical_and(mask1, mask2).sum()
+    union = np.logical_or(mask1, mask2).sum()
+    if union == 0:
+        return 1.0
+    return float(intersection / union)
+
+
+def compute_mse(video1_frames: np.ndarray, video2_frames: np.ndarray) -> list[float]:
+    if len(video1_frames) != len(video2_frames):
+        raise ValueError("Videos must have the same number of frames.")
+    frame_mses: list[float] = []
+    for frame1, frame2 in zip(video1_frames, video2_frames, strict=False):
+        if frame1.shape != frame2.shape:
+            raise ValueError("Frames must have the same dimensions.")
+        mse = np.mean((frame1.astype(np.float32) - frame2.astype(np.float32))**2)
+        frame_mses.append(round(float(mse), 4))
+    return frame_mses
+
+
+def compute_spatiotemporal_iou(mask1_frames: np.ndarray, mask2_frames: np.ndarray) -> list[float]:
+    values: list[float] = []
+    for mask1, mask2 in zip(mask1_frames, mask2_frames, strict=False):
+        values.append(round(compute_iou(mask1, mask2), 4))
+    return values
+
+
+def compute_spatial_iou(mask1_frames: np.ndarray, mask2_frames: np.ndarray) -> float:
+    spatial_mask1 = (np.max(mask1_frames, axis=0) > 0).astype(np.uint8) * 255
+    spatial_mask2 = (np.max(mask2_frames, axis=0) > 0).astype(np.uint8) * 255
+    return compute_iou(spatial_mask1, spatial_mask2)
+
+
+def compute_weighted_spatial_iou(mask1_frames: np.ndarray, mask2_frames: np.ndarray) -> float:
+    weighted_spatial_1 = np.sum(mask1_frames, axis=0, dtype=np.uint16) / len(mask1_frames)
+    weighted_spatial_2 = np.sum(mask2_frames, axis=0, dtype=np.uint16) / len(mask2_frames)
+    intersection = np.minimum(weighted_spatial_1, weighted_spatial_2)
+    union = np.maximum(weighted_spatial_1, weighted_spatial_2)
+    valid_pixels = union > 0
+    if np.sum(valid_pixels) == 0:
+        return 1.0
+    return float(np.sum(intersection[valid_pixels]) / np.sum(union[valid_pixels]))
+
+
+def mean(values: list[float] | tuple[float, ...]) -> float:
+    arr = np.asarray(list(values), dtype=np.float64)
+    if arr.size == 0:
+        raise ValueError("Cannot aggregate empty Physics-IQ values.")
+    return float(arr.mean())
+
+
+def quarter_resolution_target(reference_frames: np.ndarray) -> tuple[int, int]:
+    return (
+        max(reference_frames[0].shape[1] // 4, 1),
+        max(reference_frames[0].shape[0] // 4, 1),
+    )
+
+
+def prepare_pair_inputs(
+    generated: Any,
+    reference: Any,
+    *,
+    generated_mask: Any | None = None,
+    reference_mask: Any | None = None,
+    target_fps: int = DEFAULT_TARGET_FPS,
+    duration_seconds: int = DEFAULT_DURATION_SECONDS,
+    video_time_selection: str = "first",
+    threshold: int = 10,
+    alpha: float = 0.3,
+    roundtrip_generated_masks: bool = True,
+) -> PreparedPhysicsIQPair:
+    generated_frames, generated_color = as_numpy_video(generated)
+    reference_frames, reference_color = as_numpy_video(reference)
+    consider_frames = target_fps * duration_seconds
+
+    generated_frames = select_window(
+        generated_frames,
+        target_frames=consider_frames,
+        selection=video_time_selection,
+    )
+    reference_frames = reference_frames[:consider_frames]
+    if not len(generated_frames) or not len(reference_frames):
+        raise ValueError("Physics-IQ pair metrics require non-empty generated and reference videos.")
+
+    target_size = quarter_resolution_target(reference_frames)
+    reference_mask = reference_mask or infer_real_mask_path(reference)
+
+    generated_quarter = resize_frames(generated_frames, target_size).astype(np.float32) / 255.0
+    reference_quarter = resize_frames(reference_frames, target_size).astype(np.float32) / 255.0
+
+    generated_masks = (load_mask_frames(generated_mask, target_frames=consider_frames, target_size=target_size)
+                       if generated_mask is not None else rebinarize_masks(
+                           resize_frames(
+                               roundtrip_mask_frames(
+                                   generate_motion_mask(
+                                       generated_frames,
+                                       threshold=threshold,
+                                       alpha=alpha,
+                                       color_order=generated_color,
+                                   ),
+                                   fps=target_fps,
+                               ) if roundtrip_generated_masks else generate_motion_mask(
+                                   generated_frames,
+                                   threshold=threshold,
+                                   alpha=alpha,
+                                   color_order=generated_color,
+                               ),
+                               target_size,
+                           )))
+    reference_masks = (load_mask_frames(reference_mask, target_frames=consider_frames, target_size=target_size)
+                       if reference_mask is not None else rebinarize_masks(
+                           resize_frames(
+                               generate_motion_mask(
+                                   reference_frames,
+                                   threshold=threshold,
+                                   alpha=alpha,
+                                   color_order=reference_color,
+                               ),
+                               target_size,
+                           )))
+    return PreparedPhysicsIQPair(
+        generated_quarter=generated_quarter,
+        reference_quarter=reference_quarter,
+        generated_masks=generated_masks,
+        reference_masks=reference_masks,
+    )
+
+
+def prepare_triplet_inputs(
+    generated: Any,
+    reference: Any,
+    reference_take2: Any,
+    *,
+    generated_mask: Any | None = None,
+    reference_mask: Any | None = None,
+    reference_take2_mask: Any | None = None,
+    target_fps: int = DEFAULT_TARGET_FPS,
+    duration_seconds: int = DEFAULT_DURATION_SECONDS,
+    video_time_selection: str = "first",
+    threshold: int = 10,
+    alpha: float = 0.3,
+    roundtrip_generated_masks: bool = True,
+) -> PreparedPhysicsIQTriplet:
+    pair = prepare_pair_inputs(
+        generated,
+        reference,
+        generated_mask=generated_mask,
+        reference_mask=reference_mask,
+        target_fps=target_fps,
+        duration_seconds=duration_seconds,
+        video_time_selection=video_time_selection,
+        threshold=threshold,
+        alpha=alpha,
+        roundtrip_generated_masks=roundtrip_generated_masks,
+    )
+    reference_take2_frames, reference_take2_color = as_numpy_video(reference_take2)
+    consider_frames = target_fps * duration_seconds
+    reference_take2_frames = reference_take2_frames[:consider_frames]
+    if not len(reference_take2_frames):
+        raise ValueError("Physics-IQ requires a non-empty take-2 reference video.")
+
+    target_size = (pair.reference_quarter.shape[2], pair.reference_quarter.shape[1])
+    reference_take2_mask = reference_take2_mask or infer_real_mask_path(reference_take2)
+    reference_take2_quarter = resize_frames(reference_take2_frames, target_size).astype(np.float32) / 255.0
+    reference_take2_masks = (load_mask_frames(
+        reference_take2_mask, target_frames=consider_frames, target_size=target_size)
+                             if reference_take2_mask is not None else rebinarize_masks(
+                                 resize_frames(
+                                     generate_motion_mask(
+                                         reference_take2_frames,
+                                         threshold=threshold,
+                                         alpha=alpha,
+                                         color_order=reference_take2_color,
+                                     ),
+                                     target_size,
+                                 )))
+    return PreparedPhysicsIQTriplet(
+        generated_quarter=pair.generated_quarter,
+        reference_quarter=pair.reference_quarter,
+        reference_take2_quarter=reference_take2_quarter,
+        generated_masks=pair.generated_masks,
+        reference_masks=pair.reference_masks,
+        reference_take2_masks=reference_take2_masks,
+    )

fastvideo/eval/metrics/physics_iq/weighted_spatial_iou/metric.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from typing import Any
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.metrics.physics_iq.utils import compute_weighted_spatial_iou, prepare_pair
+
+
+@register("physics_iq.weighted_spatial_iou")
+class WeightedSpatialIoUMetric(BaseMetric):
+    name = "physics_iq.weighted_spatial_iou"
+    requires_reference = True
+    higher_is_better = True
+
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__()
+        self._kwargs = kwargs
+
+    def compute(self, sample: dict) -> MetricResult:
+        prepared = prepare_pair(sample, prep_kwargs=self._kwargs)
+        score = compute_weighted_spatial_iou(prepared.reference_masks, prepared.generated_masks)
+        return MetricResult(name=self.name, score=score, details={})

fastvideo/eval/metrics/vbench/__init__.py

@@ -0,0 +1,121 @@
+"""VBench metrics. Bootstraps upstream submodule on sys.path and
+installs runtime compat shims for modern torch/transformers/numpy/timm.
+
+The upstream vbench source lives as a git submodule at
+``fastvideo/third_party/eval/vbench`` (pinned to a specific
+Vchitect/VBench SHA). We do not pip-install it — we only need its
+Python modules importable. Its runtime deps (clip, transformers, etc.)
+are already in FastVideo's main env.
+
+Compat with modern dependency versions is achieved at import time, in
+this file, instead of via on-disk patches to upstream files. Each shim
+below corresponds to a specific drift between vbench's pinned-2023 deps
+and FastVideo's current pins. Adding a new shim is preferable to editing
+the submodule.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from typing import Any
+
+# fastvideo/eval/metrics/vbench/__init__.py  →  ../../../third_party/eval/vbench
+# parents[3] is the ``fastvideo/`` package root.
+_UPSTREAM = Path(__file__).resolve().parents[3] / "third_party" / "eval" / "vbench"
+if _UPSTREAM.is_dir() and str(_UPSTREAM) not in sys.path:
+    sys.path.insert(0, str(_UPSTREAM))
+
+
+def _install_compat_shims() -> None:
+    """Apply attribute-level shims that make vbench imports resolve.
+
+    Idempotent and side-effect-free if the targeted modules are already
+    correct (e.g. on older transformers/numpy).
+    """
+    # transformers: apply_chunking_to_forward & friends moved from
+    # ``transformers.modeling_utils`` to ``transformers.pytorch_utils``
+    # (transformers ~= 4.30+). Mirror them back so vbench's legacy
+    # ``from transformers.modeling_utils import (...)`` keeps resolving.
+    try:
+        import transformers.modeling_utils as _mu
+        import transformers.pytorch_utils as _pu
+        for _name in ("apply_chunking_to_forward", "find_pruneable_heads_and_indices", "prune_linear_layer"):
+            if not hasattr(_mu, _name) and hasattr(_pu, _name):
+                setattr(_mu, _name, getattr(_pu, _name))
+    except ImportError:
+        pass
+
+    # numpy.lib.function_base was removed entirely in numpy>=2; vbench's
+    # umt/kinetics still does ``from numpy.lib.function_base import disp``
+    # but never calls disp. Install a stub submodule with a no-op ``disp``
+    # so the legacy import line resolves.
+    try:
+        import types
+        import numpy.lib as _nl
+        if not hasattr(_nl, "function_base"):
+            _stub = types.ModuleType("numpy.lib.function_base")
+            _stub.disp = lambda *a, **k: None  # type: ignore[attr-defined]
+            sys.modules["numpy.lib.function_base"] = _stub
+            _nl.function_base = _stub  # type: ignore[attr-defined]
+    except ImportError:
+        pass
+
+
+def _install_modeling_finetune_hook() -> None:
+    """Wrap vbench's ``vit_large_patch16_224`` to drop the ``cache_dir``
+    kwarg that newer timm passes to model factory functions but the
+    upstream factory doesn't accept. Installed as a meta-path finder so
+    we patch the attribute on the actual module object after it loads,
+    without eagerly importing torch+timm at fastvideo.eval import time.
+    """
+    import importlib.abc
+
+    _target = "vbench.third_party.umt.models.modeling_finetune"
+
+    class _Loader(importlib.abc.Loader):
+
+        def __init__(self, real_loader: Any) -> None:
+            self._real = real_loader
+
+        def create_module(self, spec):
+            return None
+
+        def exec_module(self, module):
+            self._real.exec_module(module)
+            orig = getattr(module, "vit_large_patch16_224", None)
+            if orig is None or getattr(orig, "_fastvideo_patched", False):
+                return
+
+            def patched(pretrained=False, **kwargs):
+                kwargs.pop("cache_dir", None)
+                return orig(pretrained=pretrained, **kwargs)
+
+            patched._fastvideo_patched = True  # type: ignore[attr-defined]
+            module.vit_large_patch16_224 = patched
+
+    class _Finder(importlib.abc.MetaPathFinder):
+        _reentrant = False
+
+        def find_spec(self, fullname, path, target=None):
+            if fullname != _target or self._reentrant:
+                return None
+            self._reentrant = True
+            try:
+                for finder in sys.meta_path:
+                    if finder is self or not hasattr(finder, "find_spec"):
+                        continue
+                    spec = finder.find_spec(fullname, path, target)
+                    if spec is not None and spec.loader is not None:
+                        spec.loader = _Loader(spec.loader)
+                        return spec
+                return None
+            finally:
+                self._reentrant = False
+
+    if not any(isinstance(f, _Finder) for f in sys.meta_path):
+        sys.meta_path.insert(0, _Finder())
+
+
+_install_compat_shims()
+_install_modeling_finetune_hook()

fastvideo/eval/metrics/vbench/_grit_helper.py

@@ -0,0 +1,158 @@
+"""Shared GRiT model loading and detection utilities for VBench metrics.
+
+All 4 GRiT-based metrics (object_class, multiple_objects, color,
+spatial_relationship) use the same model and detection API.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import torch
+
+
+def _patch_detectron2_registries() -> None:
+    """Make detectron2's fvcore registries skip duplicate names instead of raising.
+
+    Both pip-installed ``vbench`` and wm-eval vendor the same GRiT / CenterNet2
+    source with 20+ ``@REGISTRY.register()`` calls. When both are imported in the
+    same process (e.g. the parity test), detectron2's global registries see two
+    different Python classes with the same name and raise ``AssertionError``.
+
+    This one-time patch makes ``_do_register`` silently skip if the name is
+    already registered, which is safe because the classes are identical.
+    """
+    # fvcore Registry (META_ARCH, ROI_HEADS, BACKBONE, PROPOSAL_GENERATOR, ...)
+    try:
+        from fvcore.common.registry import Registry
+        orig = Registry._do_register
+        if not getattr(orig, "_patched_idempotent", False):
+
+            def _safe_do_register(self, name, obj):
+                if name in self._obj_map:
+                    return
+                orig(self, name, obj)
+
+            _safe_do_register._patched_idempotent = True  # type: ignore[attr-defined]
+            Registry._do_register = _safe_do_register
+    except ImportError:
+        pass
+
+    # detectron2 DatasetCatalog (object365_train, vg_train, etc.)
+    try:
+        from detectron2.data import DatasetCatalog
+        orig_ds = DatasetCatalog.register
+        if not getattr(orig_ds, "_patched_idempotent", False):
+
+            def _safe_ds_register(name, func):
+                if name in DatasetCatalog:
+                    return
+                orig_ds(name, func)
+
+            _safe_ds_register._patched_idempotent = True  # type: ignore[attr-defined]
+            DatasetCatalog.register = _safe_ds_register
+    except (ImportError, AttributeError):
+        pass
+
+
+_patch_detectron2_registries()
+
+
+def load_grit_model(device: str | torch.device, task: str = "DenseCap"):
+    """Load the GRiT DenseCaptioning model.
+
+    Parameters
+    ----------
+    task : "DenseCap" | "ObjectDet"
+        VBench uses "ObjectDet" for object_class / multiple_objects /
+        spatial_relationship and "DenseCap" for color (which needs the
+        actual caption text). The two heads return predictions in
+        different formats:
+            - DenseCap → ``[(caption, bbox, [class_label]), ...]``
+            - ObjectDet → ``[(class_label, bbox, [class_label]), ...]``
+        spatial_relationship matches on ``pred[0]`` (the first field),
+        so it must run in ObjectDet mode to compare against class names.
+    """
+    from vbench.third_party.grit_model import DenseCaptioning
+    from fastvideo.eval.models import ensure_checkpoint
+
+    ckpt = ensure_checkpoint(
+        "grit_b_densecap_objectdet.pth",
+        source="OpenGVLab/VBench_Used_Models",
+        filename="grit_b_densecap_objectdet.pth",
+    )
+    # GRiT internals call .type on device, so coerce to torch.device
+    if isinstance(device, str):
+        device = torch.device(device)
+    model = DenseCaptioning(device)
+    if task == "ObjectDet":
+        model.initialize_model_det(ckpt)
+    else:
+        model.initialize_model(ckpt)
+    return model
+
+
+def detect_frames(model, frames_np: list[np.ndarray]) -> list:
+    """Run GRiT detection on a list of (H, W, C) uint8 numpy frames.
+
+    Returns per-frame predictions in the format used by VBench metrics.
+    Each frame's predictions is a list of (description, bbox, object_types).
+    """
+    predictions = []
+    with torch.no_grad():
+        for frame in frames_np:
+            ret = model.run_caption_tensor(frame)
+            predictions.append(ret[0] if len(ret[0]) > 0 else [])
+    return predictions
+
+
+def _vbench_middle_indices(vlen: int, num_frames: int) -> list[int]:
+    """Replicate VBench's get_frame_indices(sample="middle"): split [0, vlen)
+    into num_frames equal intervals and pick the midpoint of each.
+
+    Without this, wm-eval's torch.linspace-based sampler picks different
+    indices than VBench, producing different GRiT predictions and scores
+    on long videos. See vbench/utils.py:get_frame_indices.
+    """
+    acc = min(num_frames, vlen)
+    intervals = np.linspace(0, vlen, acc + 1).astype(int)
+    indices = [(intervals[i] + intervals[i + 1] - 1) // 2 for i in range(acc)]
+    if len(indices) < num_frames:
+        indices = indices + [indices[-1]] * (num_frames - len(indices))
+    return indices
+
+
+def prepare_frames(video_tensor: torch.Tensor, n_frames: int = 16, max_short_side: int = 768) -> list[np.ndarray]:
+    """Convert (T, C, H, W) float [0,1] tensor to list of (H, W, C) numpy frames
+    in VBench's exact format: float32 [0, 255] HWC.
+
+    VBench's load_video casts the decord uint8 buffer with ``torch.Tensor(...)``,
+    yielding **float32 with values in [0, 255]**, then runs torchvision
+    ``Resize`` (which preserves float dtype and produces fractional bilinear
+    outputs), then ``.permute(0,2,3,1).numpy()`` for GRiT. We must replicate
+    this exactly — round-tripping through uint8 truncates the fractional
+    bilinear outputs and shifts GRiT detection counts (e.g. 1 vs 6 persons
+    per frame), which then breaks spatial_relationship/multiple_objects.
+
+    Sampling matches VBench's ``get_frame_indices(sample="middle")``.
+    """
+    from torchvision import transforms
+
+    T = video_tensor.shape[0]
+    indices = _vbench_middle_indices(T, n_frames)
+    # Recover the original uint8 values from the float [0,1] loader
+    # (round, not truncate), then re-cast to float32 [0,255] like VBench's
+    # ``torch.Tensor(decord_uint8_array)`` path.
+    frames_uint8 = (video_tensor[indices] * 255).round().clamp(0, 255).to(torch.uint8)
+    frames_f = frames_uint8.float()
+
+    h, w = frames_f.shape[-2], frames_f.shape[-1]
+    if min(h, w) > max_short_side:
+        scale = 720.0 / min(h, w)
+        new_h, new_w = int(scale * h), int(scale * w)
+        # VBench (object_class.py:55) uses transforms.Resize without
+        # antialias kwarg → torchvision default (BILINEAR, no antialias).
+        # Float input ⇒ fractional bilinear outputs are preserved.
+        frames_f = transforms.Resize(size=(new_h, new_w))(frames_f)
+
+    frames_np = frames_f.permute(0, 2, 3, 1).cpu().numpy().astype(np.float32)
+    return list(frames_np)

fastvideo/eval/metrics/vbench/_utils.py

@@ -0,0 +1,31 @@
+"""Shared utilities for VBench metrics."""
+
+from __future__ import annotations
+
+import torch
+import torch.nn.functional as F
+
+
+def consistency_score(features: torch.Tensor) -> float:
+    """VBench-style temporal consistency from (T, D) L2-normalized features.
+
+    For each frame t > 0, computes:
+        sim = (cos(f[t], f[t-1]) + cos(f[t], f[0])) / 2, clamped >= 0
+
+    Returns the mean similarity across all t > 0.
+    """
+    if features.shape[0] <= 1:
+        return 1.0
+
+    first = features[0:1]  # (1, D)
+    total_sim = 0.0
+    count = 0
+    for t in range(1, features.shape[0]):
+        curr = features[t:t + 1]
+        prev = features[t - 1:t]
+        sim_prev = max(0.0, F.cosine_similarity(prev, curr).item())
+        sim_first = max(0.0, F.cosine_similarity(first, curr).item())
+        total_sim += (sim_prev + sim_first) / 2
+        count += 1
+
+    return total_sim / count

fastvideo/eval/metrics/vbench/aesthetic_quality/metric.py

@@ -0,0 +1,98 @@
+"""VBench Aesthetic Quality — CLIP ViT-L/14 + LAION aesthetic predictor.
+
+Encodes frames through CLIP, passes L2-normalized features through a
+linear aesthetic head (768 → 1), and averages scores / 10.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms.functional import resize, center_crop, normalize
+from torchvision.transforms import InterpolationMode
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+_CLIP_MEAN = [0.48145466, 0.4578275, 0.40821073]
+_CLIP_STD = [0.26862954, 0.26130258, 0.27577711]
+
+_AESTHETIC_URL = "https://raw.githubusercontent.com/LAION-AI/aesthetic-predictor/main/sa_0_4_vit_l_14_linear.pth"
+
+
+def _clip_transform(frames: torch.Tensor) -> torch.Tensor:
+    # antialias=False matches VBench's clip_transform (vbench/utils.py:33)
+    frames = resize(frames, 224, interpolation=InterpolationMode.BICUBIC, antialias=False)
+    frames = center_crop(frames, 224)
+    frames = normalize(frames, mean=_CLIP_MEAN, std=_CLIP_STD)
+    return frames
+
+
+@register("vbench.aesthetic_quality")
+class AestheticQualityMetric(BaseMetric):
+
+    name = "vbench.aesthetic_quality"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["clip"]
+    backbone = "clip_vit_l14"
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._clip_model: Any = None
+        self._aesthetic_head: Any = None
+
+    def to(self, device):
+        super().to(device)
+        if self._clip_model is not None:
+            self._clip_model = self._clip_model.to(self.device)
+        if self._aesthetic_head is not None:
+            self._aesthetic_head = self._aesthetic_head.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._clip_model is not None:
+            return
+
+        import clip
+        from fastvideo.eval.models import ensure_checkpoint, get_cache_dir
+        self._clip_model, _ = clip.load(
+            "ViT-L/14",
+            device=self.device,
+            download_root=str(get_cache_dir() / "clip"),
+        )
+        self._clip_model.eval()
+
+        # Load LAION aesthetic head
+        ckpt_path = ensure_checkpoint(
+            "sa_0_4_vit_l_14_linear.pth",
+            source=_AESTHETIC_URL,
+        )
+        self._aesthetic_head = nn.Linear(768, 1)
+        self._aesthetic_head.load_state_dict(torch.load(ckpt_path, map_location="cpu", weights_only=True))
+        self._aesthetic_head.to(self.device)
+        self._aesthetic_head.eval()
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        video = sample["video"]  # (T, C, H, W)
+        frames = _clip_transform(video.to(self.device))
+
+        chunk = self._chunk_size or 32
+        scores_list = []
+        for i in range(0, frames.shape[0], chunk):
+            feats = self._clip_model.encode_image(frames[i:i + chunk]).float()
+            feats = F.normalize(feats, dim=-1, p=2)
+            scores_list.append(self._aesthetic_head(feats).squeeze(-1))
+
+        all_scores = torch.cat(scores_list, dim=0) / 10.0  # (T,)
+        return MetricResult(
+            name=self.name,
+            score=float(all_scores.mean().item()),
+            details={"per_frame": all_scores.tolist()},
+        )

fastvideo/eval/metrics/vbench/appearance_style/metric.py

@@ -0,0 +1,94 @@
+"""VBench Appearance Style — CLIP ViT-B/32 text-image alignment.
+
+Per-frame cosine similarity between CLIP image features and a text
+prompt describing the expected style.  Requires ``sample["text_prompt"]``.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+from torchvision.transforms.functional import resize, center_crop, normalize
+from torchvision.transforms import InterpolationMode
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+_CLIP_MEAN = [0.48145466, 0.4578275, 0.40821073]
+_CLIP_STD = [0.26862954, 0.26130258, 0.27577711]
+
+
+def _clip_transform(frames: torch.Tensor) -> torch.Tensor:
+    # antialias=False matches VBench's clip_transform (vbench/utils.py:33)
+    frames = resize(frames, 224, interpolation=InterpolationMode.BICUBIC, antialias=False)
+    frames = center_crop(frames, 224)
+    frames = normalize(frames, mean=_CLIP_MEAN, std=_CLIP_STD)
+    return frames
+
+
+@register("vbench.appearance_style")
+class AppearanceStyleMetric(BaseMetric):
+
+    name = "vbench.appearance_style"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["clip"]
+    backbone = "clip_vit_b32"
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+
+    def to(self, device):
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        import clip
+        from fastvideo.eval.models import get_cache_dir
+        self._model, _ = clip.load(
+            "ViT-B/32",
+            device=self.device,
+            download_root=str(get_cache_dir() / "clip"),
+        )
+        self._model.eval()
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        import clip
+
+        video = sample["video"]  # (T, C, H, W)
+        text_prompt = sample.get("text_prompt")
+        if text_prompt is None:
+            return self._skip(sample, "missing text_prompt")
+
+        frames = _clip_transform(video.to(self.device))
+
+        chunk = self._chunk_size or 64
+        img_feats = []
+        for i in range(0, frames.shape[0], chunk):
+            f = self._model.encode_image(frames[i:i + chunk]).float()
+            f = F.normalize(f, dim=-1, p=2)
+            img_feats.append(f)
+        img_feats = torch.cat(img_feats, dim=0)  # (T, D)
+
+        # truncate=True: CLIP context length is 77 tokens; long prompts
+        # truncate instead of raising. Matches CLIP's documented convention.
+        text_tokens = clip.tokenize([text_prompt], truncate=True).to(self.device)
+        text_feat = self._model.encode_text(text_tokens).float()
+        text_feat = F.normalize(text_feat, dim=-1, p=2)  # (1, D)
+
+        sims = (img_feats @ text_feat.T).squeeze(-1)  # (T,)
+        return MetricResult(
+            name=self.name,
+            score=float(sims.mean().item()),
+            details={"per_frame": sims.tolist()},
+        )

fastvideo/eval/metrics/vbench/background_consistency/metric.py

@@ -0,0 +1,83 @@
+"""VBench Background Consistency — CLIP ViT-B/32 temporal feature similarity.
+
+Measures background stability via cosine similarity of CLIP features
+between consecutive frames and the first frame.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+from torchvision.transforms.functional import resize, center_crop, normalize
+from torchvision.transforms import InterpolationMode
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.metrics.vbench._utils import consistency_score
+
+_CLIP_MEAN = [0.48145466, 0.4578275, 0.40821073]
+_CLIP_STD = [0.26862954, 0.26130258, 0.27577711]
+
+
+def _clip_transform(frames: torch.Tensor) -> torch.Tensor:
+    """Apply CLIP preprocessing to (N, C, H, W) float [0,1] tensors."""
+    # antialias=False matches VBench's clip_transform (vbench/utils.py:33)
+    frames = resize(frames, 224, interpolation=InterpolationMode.BICUBIC, antialias=False)
+    frames = center_crop(frames, 224)
+    frames = normalize(frames, mean=_CLIP_MEAN, std=_CLIP_STD)
+    return frames
+
+
+@register("vbench.background_consistency")
+class BackgroundConsistencyMetric(BaseMetric):
+
+    name = "vbench.background_consistency"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["clip"]
+    backbone = "clip_vit_b32"
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+
+    def to(self, device):
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        import clip
+        from fastvideo.eval.models import get_cache_dir
+        model, _ = clip.load(
+            "ViT-B/32",
+            device=self.device,
+            download_root=str(get_cache_dir() / "clip"),
+        )
+        model.eval()
+        self._model = model
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        video = sample["video"]  # (T, C, H, W)
+        frames = _clip_transform(video.to(self.device))
+
+        chunk = self._chunk_size or 64
+        feats = []
+        for i in range(0, frames.shape[0], chunk):
+            f = self._model.encode_image(frames[i:i + chunk]).float()
+            f = F.normalize(f, dim=-1, p=2)
+            feats.append(f)
+        all_feats = torch.cat(feats, dim=0)  # (T, D)
+        return MetricResult(
+            name=self.name,
+            score=consistency_score(all_feats),
+            details={},
+        )

fastvideo/eval/metrics/vbench/color/metric.py

@@ -0,0 +1,106 @@
+"""VBench Color — GRiT dense captioning for color accuracy.
+
+Detects the target object via GRiT and checks if the expected color
+keyword appears in the object's caption. Score = frames_with_correct_color
+/ frames_with_object_detected.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+_COLOR_KEYWORDS = [
+    "white",
+    "red",
+    "pink",
+    "blue",
+    "silver",
+    "purple",
+    "orange",
+    "green",
+    "gray",
+    "yellow",
+    "black",
+    "grey",
+]
+
+
+@register("vbench.color")
+class ColorMetric(BaseMetric):
+
+    name = "vbench.color"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["detectron2"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        from fastvideo.eval.metrics.vbench._grit_helper import load_grit_model
+        self._model = load_grit_model(self.device)
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        from fastvideo.eval.metrics.vbench._grit_helper import prepare_frames
+
+        video = sample["video"]  # (T, C, H, W)
+        aux = sample.get("auxiliary_info") or {}
+        if "color" not in aux:
+            return self._skip(sample, "missing 'color' in auxiliary_info")
+
+        prompt = sample.get("text_prompt") or ""
+        color_key = aux["color"]
+        # Parse object name: remove "a ", "an ", and the color word
+        object_key = prompt.replace("a ", "").replace("an ", "").replace(color_key, "").strip()
+
+        frames_np = prepare_frames(video)
+
+        preds = []
+        for frame in frames_np:
+            ret = self._model.run_caption_tensor(frame)
+            cur_pred = []
+            if len(ret[0]) < 1:
+                cur_pred.append(["", ""])
+            else:
+                for cap_det in ret[0]:
+                    cur_pred.append([cap_det[0], cap_det[2][0]])
+            preds.append(cur_pred)
+
+        # Score: matching VBench's check_generate logic
+        cur_object = 0
+        cur_object_color = 0
+        for frame_pred in preds:
+            object_flag = False
+            color_flag = False
+            for pred in frame_pred:
+                if object_key == pred[1]:
+                    for cq in _COLOR_KEYWORDS:
+                        if cq in pred[0]:
+                            object_flag = True
+                    if color_key in pred[0]:
+                        color_flag = True
+            if color_flag:
+                cur_object_color += 1
+            if object_flag:
+                cur_object += 1
+
+        score = cur_object_color / cur_object if cur_object > 0 else 0.0
+        return MetricResult(
+            name=self.name,
+            score=float(score),
+            details={
+                "object_detected": cur_object,
+                "color_correct": cur_object_color
+            },
+        )

fastvideo/eval/metrics/vbench/dynamic_degree/metric.py

@@ -0,0 +1,134 @@
+"""VBench Dynamic Degree — RAFT optical flow motion detection.
+
+For each consecutive frame pair, computes optical flow via RAFT and takes
+the mean of the top 5% flow magnitudes.  If enough pairs exceed an
+adaptive threshold, the video is classified as dynamic (1.0) vs static (0.0).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import numpy as np
+import torch
+from easydict import EasyDict
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("vbench.dynamic_degree")
+class DynamicDegreeMetric(BaseMetric):
+
+    name = "vbench.dynamic_degree"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["easydict"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+        self._chunk_size = 16
+
+    def to(self, device):
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        from vbench.third_party.RAFT.core.raft import RAFT
+
+        args = EasyDict(small=False, mixed_precision=False, alternate_corr=False, dropout=0.0)
+        model = torch.nn.DataParallel(RAFT(args))
+
+        from fastvideo.eval.models import ensure_checkpoint
+        ckpt_path = ensure_checkpoint(
+            "raft-things.pth",
+            source="sbalani/raft-things",
+            filename="raft-things.pth",
+        )
+        model.load_state_dict(torch.load(ckpt_path, map_location="cpu"))
+        model = model.module
+        model.to(self.device)
+        model.eval()
+        self._model = model
+
+    def _get_score(self, flow: torch.Tensor) -> float:
+        """Top-5% mean flow magnitude (matching VBench dynamic_degree.get_score)."""
+        flo = flow.permute(1, 2, 0).cpu().numpy()
+        rad = np.sqrt(flo[..., 0]**2 + flo[..., 1]**2)
+        h, w = rad.shape
+        cut = max(1, int(h * w * 0.05))
+        rad_flat = rad.flatten()
+        return float(np.mean(np.sort(rad_flat)[-cut:]))
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        from vbench.third_party.RAFT.core.utils_core.utils import InputPadder
+
+        video = sample["video"]  # (T, C, H, W) [0, 1]
+        T, _, H, W = video.shape
+
+        # fps controls the temporal sampling stride for optical flow.
+        # vbench computes flow at 8fps (interval = round(fps/8)). The metric
+        # cannot auto-derive fps from a tensor, so a missing fps would silently
+        # use a wrong stride and produce a wrong score. Skip explicitly.
+        if "fps" not in sample:
+            return self._skip(sample, "missing 'fps' (required to set the "
+                              "8fps optical-flow sampling stride)")
+        fps = float(sample["fps"])
+        interval = max(1, round(fps / 8.0))
+
+        video_255 = video * 255.0
+        chunk = self._chunk_size or 16
+
+        # Cap chunk so the RAFT correlation volume doesn't overflow int32.
+        # RAFT downsamples 8x in the feature encoder; CorrBlock's tensor is
+        # shape (B*H1*W1, 1, H2, W2) with H1=H2=H/8, W1=W2=W/8. Its element
+        # count is B*(H/8)^2*(W/8)^2 — F.avg_pool2d's index space starts to
+        # overflow int32 around 2^31. Safety factor 2x.
+        h_red = max(1, H // 8)
+        w_red = max(1, W // 8)
+        max_chunk = max(1, (1 << 30) // (h_red * h_red * w_red * w_red))
+        chunk = min(chunk, max_chunk)
+
+        indices = list(range(0, T, interval))
+        n = len(indices)
+        all_img1 = [video_255[indices[i]] for i in range(n - 1)]
+        all_img2 = [video_255[indices[i + 1]] for i in range(n - 1)]
+
+        scores: list[float] = []
+        for start in range(0, len(all_img1), chunk):
+            end = min(start + chunk, len(all_img1))
+            img1_batch = torch.stack(all_img1[start:end]).to(self.device)
+            img2_batch = torch.stack(all_img2[start:end]).to(self.device)
+            padder = InputPadder(img1_batch.shape)
+            img1p, img2p = padder.pad(img1_batch, img2_batch)
+            _, flow = self._model(img1p, img2p, iters=20, test_mode=True)
+            for i in range(flow.shape[0]):
+                scores.append(self._get_score(flow[i]))
+
+        scale = min(H, W)
+        thres = 6.0 * (scale / 256.0)
+        count_needed = round(4 * (n / 16.0))
+        count_above = sum(1 for s in scores if s > thres)
+        is_dynamic = 1.0 if count_above >= count_needed else 0.0
+
+        return MetricResult(
+            name=self.name,
+            score=is_dynamic,
+            details={
+                "per_pair_magnitude": scores,
+                "threshold": thres,
+                "count_above": count_above,
+                "count_needed": count_needed,
+                "fps": fps,
+                "interval": interval,
+                "n_frames_used": n
+            },
+        )

fastvideo/eval/metrics/vbench/human_action/metric.py

@@ -0,0 +1,131 @@
+"""VBench Human Action — UMT ViT-L/16 action classification (Kinetics-400).
+
+Classifies human actions in 16-frame clips. Top-5 predictions with
+confidence >= 0.85 are compared against the ground-truth action label.
+Score = 1.0 if match found, 0.0 otherwise.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import torch
+from torchvision.transforms.functional import resize, center_crop, normalize
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+from fastvideo.eval.io.video import extract_frames
+
+# Kinetics-400 class names (loaded lazily). The label file ships inside
+# the upstream vbench submodule.
+_CAT_DICT: dict[str, str] | None = None
+
+
+def _load_cat_dict() -> dict[str, str]:
+    global _CAT_DICT
+    if _CAT_DICT is not None:
+        return _CAT_DICT
+    import vbench.third_party.umt as _umt_pkg
+    cat_path = (Path(_umt_pkg.__file__).resolve().parent / "kinetics_400_categories.txt")
+    out: dict[str, str] = {}
+    with cat_path.open() as f:
+        for line in f:
+            parts = line.strip().split("\t")
+            if len(parts) == 2:
+                cat, idx = parts
+                out[idx] = cat.lower()
+    _CAT_DICT = out
+    return _CAT_DICT
+
+
+@register("vbench.human_action")
+class HumanActionMetric(BaseMetric):
+
+    name = "vbench.human_action"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["timm"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+
+    def to(self, device):
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        from timm.models import create_model
+        from fastvideo.eval.models import ensure_checkpoint
+
+        ckpt_path = ensure_checkpoint(
+            "umt_l16_kinetics400.pth",
+            source="OpenGVLab/VBench_Used_Models",
+            filename="l16_ptk710_ftk710_ftk400_f16_res224.pth",
+        )
+
+        import vbench.third_party.umt.models.modeling_finetune  # noqa: F401
+
+        self._model = create_model(
+            "vit_large_patch16_224",
+            pretrained=False,
+            num_classes=400,
+            all_frames=16,
+            tubelet_size=1,
+            use_learnable_pos_emb=False,
+            fc_drop_rate=0.0,
+            drop_rate=0.0,
+            drop_path_rate=0.2,
+            attn_drop_rate=0.0,
+            drop_block_rate=None,
+            use_checkpoint=False,
+            checkpoint_num=16,
+            use_mean_pooling=True,
+            init_scale=0.001,
+        )
+        state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+        self._model.load_state_dict(state_dict, strict=False)
+        self._model.to(self.device)
+        self._model.eval()
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        video = sample["video"]  # (T, C, H, W) [0, 1]
+        text_prompt = sample.get("text_prompt")
+        if text_prompt is None:
+            return self._skip(sample, "missing text_prompt with action labels")
+
+        cat_dict = _load_cat_dict()
+
+        frames = extract_frames(video, 16)  # (16, C, H, W)
+        frames = resize(frames, 256, antialias=True)
+        frames = center_crop(frames, 224)
+        frames = normalize(frames, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+        # UMT expects (C, T, H, W); add a leading batch dim of 1.
+        clip_in = frames.permute(1, 0, 2, 3).unsqueeze(0).to(self.device)
+
+        logits = torch.sigmoid(self._model(clip_in))  # (1, 400)
+        top_scores, top_indices = torch.topk(logits[0], 5)
+        top_indices = top_indices.tolist()
+        top_scores = top_scores.tolist()
+
+        predictions = [
+            cat_dict.get(str(idx), "") for idx, score in zip(top_indices, top_scores, strict=False) if score >= 0.85
+        ]
+        gt_label = text_prompt.lower().strip()
+        match = any(pred == gt_label for pred in predictions)
+        return MetricResult(
+            name=self.name,
+            score=1.0 if match else 0.0,
+            details={
+                "predictions": predictions,
+                "ground_truth": gt_label
+            },
+        )

fastvideo/eval/metrics/vbench/imaging_quality/metric.py

@@ -0,0 +1,71 @@
+"""VBench Imaging Quality — MUSIQ-based per-frame technical quality.
+
+Uses MUSIQ (Multi-Scale Image Quality) from pyiqa.  Frames are resized
+so the longer side is at most 512px.  Score = mean(MUSIQ_scores) / 100.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+from torchvision.transforms.functional import resize
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("vbench.imaging_quality")
+class ImagingQualityMetric(BaseMetric):
+
+    name = "vbench.imaging_quality"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["pyiqa"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+
+    def to(self, device):
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        import pyiqa
+        self._model = pyiqa.create_metric("musiq-spaq", device=self.device)
+        self._model.eval()
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        video = sample["video"]  # (T, C, H, W)
+        T, _, H, W = video.shape
+
+        if max(H, W) > 512:
+            scale = 512.0 / max(H, W)
+            new_h, new_w = int(H * scale), int(W * scale)
+        else:
+            new_h, new_w = H, W
+
+        frames = video.to(self.device)
+        if (new_h, new_w) != (H, W):
+            # antialias=False matches VBench's imaging_quality.transform
+            frames = resize(frames, [new_h, new_w], antialias=False)
+
+        chunk = self._chunk_size or 32
+        chunks: list[torch.Tensor] = []
+        for i in range(0, T, chunk):
+            scores = self._model(frames[i:i + chunk])
+            chunks.append(scores.squeeze(-1))
+        per_frame = torch.cat(chunks, dim=0)  # (T,)
+        return MetricResult(
+            name=self.name,
+            score=float(per_frame.mean().item()) / 100.0,
+            details={"per_frame_raw": per_frame.tolist()},
+        )

fastvideo/eval/metrics/vbench/motion_smoothness/metric.py

@@ -0,0 +1,200 @@
+"""VBench Motion Smoothness — AMT-S frame interpolation quality.
+
+Takes every-other frame, uses AMT-S to interpolate the missing middle
+frames, then compares interpolated vs actual frames.
+Score = (255 - mean_pixel_diff) / 255.  Higher = smoother motion.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import os
+
+import cv2
+import numpy as np
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("vbench.motion_smoothness")
+class MotionSmoothnessMetric(BaseMetric):
+
+    name = "vbench.motion_smoothness"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["omegaconf"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model: Any = None
+        self._embt: Any = None
+        self._chunk_size = 8
+
+    def to(self, device):
+        super().to(device)
+        if self._model is not None:
+            self._model = self._model.to(self.device)
+        if self._embt is not None:
+            self._embt = self._embt.to(self.device)
+        return self
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        from omegaconf import OmegaConf
+        import vbench.third_party.amt as _amt_pkg
+        from vbench.third_party.amt.utils.build_utils import build_from_cfg
+        from fastvideo.eval.models import ensure_checkpoint
+
+        amt_dir = os.path.dirname(_amt_pkg.__file__)
+        cfg_path = os.path.join(amt_dir, "cfgs", "AMT-S.yaml")
+
+        ckpt_path = ensure_checkpoint(
+            "amt-s.pth",
+            source="https://huggingface.co/lalala125/AMT/resolve/main/amt-s.pth",
+        )
+
+        network_cfg = OmegaConf.load(cfg_path).network
+        self._model = build_from_cfg(network_cfg)
+        ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+        self._model.load_state_dict(ckpt["state_dict"])
+        self._model.to(self.device)
+        self._model.eval()
+
+        self._embt = torch.tensor(1 / 2).float().view(1, 1, 1, 1).to(self.device)
+
+    def _get_scale(self, h: int, w: int) -> float:
+        """Pick a downscale factor that keeps AMT's correlation volume
+        within free GPU memory.
+
+        Re-queries free memory on every call (rather than caching at setup
+        time) so the scale adapts to whatever's actually available — other
+        metric replicas already loaded, residual generator allocations,
+        another process sharing the GPU, etc. The upstream version cached
+        ``total_memory`` at setup, which on a shared/loaded GPU lets AMT
+        attempt a 30+ GB correlation volume reshape and OOM.
+        """
+        if self.device.type != "cuda":
+            return 1.0
+        # Free memory that won't be claimed by other allocations during this
+        # forward pass. min(free, total) is conservative against transient
+        # spikes; mem_get_info returns (free, total) in bytes.
+        free_bytes, _ = torch.cuda.mem_get_info(self.device)
+        anchor_resolution = 1024 * 512
+        anchor_memory = 1500 * 1024**2
+        anchor_memory_bias = 2500 * 1024**2
+        if free_bytes <= anchor_memory_bias:
+            # Less than the model + scratch overhead is free; force the
+            # most aggressive downscale we support.
+            return 1 / 16
+        scale = anchor_resolution / (h * w) * np.sqrt((free_bytes - anchor_memory_bias) / anchor_memory)
+        if scale >= 1.0:
+            return 1.0
+        scale = 1 / np.floor(1 / np.sqrt(scale) * 16) * 16
+        return float(scale)
+
+    _MIN_AMT_SCALE = 1 / 16
+
+    def _safe_amt_forward(self, in0: Any, in1: Any, embt: Any, scale: Any) -> Any:
+        """Run one chunk through AMT with OOM-retry on two axes.
+
+        Recovery strategy on ``CUDA out of memory``:
+
+        1. **Halve the batch** until batch=1. AMT's per-pair memory
+           dominates; splitting helps until each pair is on its own.
+        2. **Halve the scale_factor** passed to AMT (which controls its
+           internal feature-map resolution and therefore the correlation
+           volume size). Bottoms out at ``_MIN_AMT_SCALE`` — beyond that
+           the feature maps are too coarse to produce meaningful
+           interpolation and we re-raise.
+
+        Self-tunes under memory pressure: the upstream autoscale formula
+        in :meth:`_get_scale` mis-extrapolates at large resolutions
+        (treats memory as linear in pixel count, but AMT's correlation
+        volume grows quadratically). This retry path makes the metric
+        robust to that without rewriting the formula.
+        """
+        try:
+            return self._model(in0, in1, embt, scale_factor=scale, eval=True)["imgt_pred"]
+        except torch.cuda.OutOfMemoryError:
+            torch.cuda.empty_cache()
+            bs = in0.shape[0]
+            if bs > 1:
+                half = bs // 2
+                a = self._safe_amt_forward(in0[:half], in1[:half], embt[:half], scale)
+                b = self._safe_amt_forward(in0[half:], in1[half:], embt[half:], scale)
+                return torch.cat([a, b], dim=0)
+            if scale > self._MIN_AMT_SCALE:
+                return self._safe_amt_forward(in0, in1, embt, scale / 2)
+            raise
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        from vbench.third_party.amt.utils.utils import (
+            img2tensor,
+            tensor2img,
+            check_dim_and_resize,
+            InputPadder,
+        )
+
+        video = sample["video"]  # (T, C, H, W) [0, 1]
+        chunk = self._chunk_size or 8
+
+        frames_np = (video * 255).to(torch.uint8).cpu().numpy()
+        frames_np = [f.transpose(1, 2, 0) for f in frames_np]  # list of (H,W,C)
+
+        even_indices = list(range(0, len(frames_np), 2))
+        if len(even_indices) <= 1:
+            return MetricResult(name=self.name, score=1.0, details={})
+
+        even_frames = [frames_np[i] for i in even_indices]
+        inputs = [img2tensor(f).to(self.device) for f in even_frames]
+        inputs = check_dim_and_resize(inputs)
+
+        h, w = inputs[0].shape[-2:]
+        scale = self._get_scale(h, w)
+        padding = int(16 / scale)
+        padder = InputPadder(inputs[0].shape, padding)
+        inputs = padder.pad(*inputs)
+
+        n_pairs = len(inputs) - 1
+        all_in0 = [inputs[i] for i in range(n_pairs)]
+        all_in1 = [inputs[i + 1] for i in range(n_pairs)]
+        all_gt = [
+            frames_np[even_indices[i] + 1] if even_indices[i] + 1 < len(frames_np) else frames_np[-1]
+            for i in range(n_pairs)
+        ]
+
+        all_preds = []
+        for start in range(0, len(all_in0), chunk):
+            end = min(start + chunk, len(all_in0))
+            in0_batch = torch.cat(all_in0[start:end], dim=0).to(self.device)
+            in1_batch = torch.cat(all_in1[start:end], dim=0).to(self.device)
+            embt = self._embt.expand(in0_batch.shape[0], -1, -1, -1)
+            pred = self._safe_amt_forward(in0_batch, in1_batch, embt, scale)
+            all_preds.append(pred.cpu())
+        all_preds = torch.cat(all_preds, dim=0)
+
+        diffs: list[float] = []
+        for i in range(n_pairs):
+            pred = all_preds[i:i + 1]
+            pred_unpadded = padder.unpad(pred)[0]
+            pred_np = tensor2img(pred_unpadded)
+            gt_np = all_gt[i]
+            # gt comes from frames_np; pred goes through check_dim_and_resize +
+            # AMT pad/unpad, which can reshape. Match shapes before absdiff.
+            if gt_np.shape[:2] != pred_np.shape[:2]:
+                gt_np = cv2.resize(gt_np, (pred_np.shape[1], pred_np.shape[0]), interpolation=cv2.INTER_AREA)
+            diffs.append(float(np.mean(cv2.absdiff(gt_np, pred_np))))
+
+        vfi_score = float(np.mean(diffs)) if diffs else 0.0
+        return MetricResult(
+            name=self.name,
+            score=(255.0 - vfi_score) / 255.0,
+            details={"vfi_score": vfi_score},
+        )

fastvideo/eval/metrics/vbench/multiple_objects/metric.py

@@ -0,0 +1,73 @@
+"""VBench Multiple Objects — GRiT detection for dual-object presence.
+
+Checks if BOTH target objects are detected in each of 16 sampled frames.
+Score = matching_frames / total_frames.
+"""
+
+from __future__ import annotations
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("vbench.multiple_objects")
+class MultipleObjectsMetric(BaseMetric):
+
+    name = "vbench.multiple_objects"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["detectron2"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model = None
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        from fastvideo.eval.metrics.vbench._grit_helper import load_grit_model
+        # VBench's multiple_objects uses ObjectDet head
+        self._model = load_grit_model(self.device, task="ObjectDet")
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        from fastvideo.eval.metrics.vbench._grit_helper import prepare_frames, detect_frames
+
+        video = sample["video"]  # (T, C, H, W)
+        aux = sample.get("auxiliary_info") or {}
+        if "object" not in aux:
+            return self._skip(sample, "missing 'object' in auxiliary_info")
+
+        object_info = aux["object"]
+        if " and " not in object_info:
+            # multiple_objects expects "<a> and <b>"; single objects are
+            # the object_class metric's territory — skip this row.
+            return self._skip(sample, "'object' lacks ' and ' separator")
+        key_a, key_b = [k.strip() for k in object_info.split(" and ")]
+
+        frames_np = prepare_frames(video)
+        preds = detect_frames(self._model, frames_np)
+
+        matching = 0
+        for frame_pred in preds:
+            try:
+                obj_set = set(frame_pred[0][2]) if frame_pred else set()
+            except (IndexError, TypeError):
+                obj_set = set()
+            if key_a in obj_set and key_b in obj_set:
+                matching += 1
+
+        total = len(preds)
+        score = matching / total if total > 0 else 0.0
+        return MetricResult(
+            name=self.name,
+            score=float(score),
+            details={
+                "matching_frames": matching,
+                "total_frames": total
+            },
+        )

fastvideo/eval/metrics/vbench/object_class/metric.py

@@ -0,0 +1,71 @@
+"""VBench Object Class — GRiT object detection for class matching.
+
+Checks if a target object class is detected in each of 16 sampled frames.
+Score = matching_frames / total_frames.
+"""
+
+from __future__ import annotations
+
+import torch
+
+from fastvideo.eval.metrics.base import BaseMetric
+from fastvideo.eval.registry import register
+from fastvideo.eval.types import MetricResult
+
+
+@register("vbench.object_class")
+class ObjectClassMetric(BaseMetric):
+
+    name = "vbench.object_class"
+    requires_reference = False
+    higher_is_better = True
+    needs_gpu = True
+    dependencies = ["detectron2"]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self._model = None
+
+    def setup(self) -> None:
+        if self._model is not None:
+            return
+        from fastvideo.eval.metrics.vbench._grit_helper import load_grit_model
+        # VBench's object_class uses ObjectDet head (init_submodules → "ObjectDet")
+        self._model = load_grit_model(self.device, task="ObjectDet")
+
+    @torch.no_grad()
+    def compute(self, sample: dict) -> MetricResult:
+        from fastvideo.eval.metrics.vbench._grit_helper import prepare_frames, detect_frames
+
+        video = sample["video"]  # (T, C, H, W)
+        aux = sample.get("auxiliary_info") or {}
+        if "object" not in aux:
+            return self._skip(sample, "missing 'object' in auxiliary_info")
+
+        object_key = aux["object"]
+        if " and " in object_key:
+            # multiple_objects' territory; skip this row for object_class.
+            return self._skip(sample, "'object' contains ' and ' (multi-object)")
+
+        frames_np = prepare_frames(video)
+        preds = detect_frames(self._model, frames_np)
+
+        matching = 0
+        for frame_pred in preds:
+            try:
+                obj_set = set(frame_pred[0][2]) if frame_pred else set()
+            except (IndexError, TypeError):
+                obj_set = set()
+            if object_key in obj_set:
+                matching += 1
+
+        total = len(preds)
+        score = matching / total if total > 0 else 0.0
+        return MetricResult(
+            name=self.name,
+            score=float(score),
+            details={
+                "matching_frames": matching,
+                "total_frames": total
+            },
+        )