chore(deps): drop <3.13 Python cap, require coreml-diffusion>=0.1.6 (#68)

Mirror the library: coreml-diffusion 0.1.6 dropped its stale <3.13 cap, so this
package can too (requires-python >=3.12). Bump the dependency floor to 0.1.6 to
keep the cap consistent across both — a consumer pinned to <3.13 with a library
allowing 3.13+ would be unresolvable for the 3.13 range. Pin the dev/CI
interpreter to 3.12 via .python-version so CI runs on a fixed baseline rather
than floating to a newer Python without a pinned-torch wheel.

.github/workflows/publish.yml

@@ -0,0 +1,25 @@
+name: Publish to Comfy registry
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - "pyproject.toml"
+
+permissions:
+  issues: write
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    if: ${{ github.repository_owner == 'aszc-dev' }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@v1
+        with:
+          ## Add your own personal access token to your Github Repository secrets and reference it here.
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

.github/workflows/tier0.yml

@@ -0,0 +1,27 @@
+name: Tier 0 — Unit (Linux)
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+# Deps are resolved from pyproject.toml via uv, so the toolchain pins live in
+# one place. Tier 0 must run without ComfyUI; the in-tree purity gate
+# (tests/unit/test_tier0_purity.py) enforces that the suite hasn't started
+# leaking framework imports.
+jobs:
+  unit:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+
+      - name: uv sync
+        run: uv sync --no-install-project
+
+      - name: Run Tier 0
+        run: uv run pytest -m unit tests/ -v

.github/workflows/tier2.yml

@@ -0,0 +1,134 @@
+name: Tier 2 — M2 / ANE (self-hosted)
+
+on:
+  pull_request:
+    # `labeled` fires when run-m2 is first added; `synchronize`/`reopened`
+    # re-run on every subsequent push while the label is present, so the
+    # result tracks the PR head instead of going stale. The `if` below keeps
+    # the run gated on the run-m2 label for all pull_request events.
+    types: [labeled, synchronize, reopened]
+  schedule:
+    # Nightly at 04:00 UTC (~05/06 in PL). Keeps the M2 path honest
+    # without burning the runner on every PR.
+    - cron: "0 4 * * *"
+  workflow_dispatch:
+
+jobs:
+  m2:
+    if: |
+      github.event_name == 'schedule' ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' &&
+       contains(github.event.pull_request.labels.*.name, 'run-m2'))
+    # Self-hosted Apple Silicon runner. Prerequisites: COMFY_DIR pointing at
+    # a runner-owned ComfyUI clone, plus a cached SD1.5 checkpoint.
+    runs-on: [self-hosted, macOS, ARM64, coreml]
+    timeout-minutes: 90
+    steps:
+      - uses: actions/checkout@v4
+
+      # Hybrid ComfyUI strategy:
+      #   - schedule (nightly)  -> latest origin/master + ComfyUI's own
+      #     requirements.txt (constrained). Canary for upstream API breakage.
+      #   - PR label / dispatch -> the requires-comfyui version tag + the frozen
+      #     `comfy` uv group. Reproducible merge gate, immune to overnight drift.
+      - name: Resolve ComfyUI ref + mode
+        run: |
+          if [ "$GITHUB_EVENT_NAME" = "schedule" ]; then
+            echo "COMFY_MODE=latest" >> "$GITHUB_ENV"
+            echo "COMFY_REF=master"  >> "$GITHUB_ENV"
+          else
+            # requires-comfyui is a semver constraint (e.g. ">=0.3.27"); pin the
+            # gate to the matching ComfyUI release tag (vX.Y.Z).
+            VERSION="$(sed -nE 's/^requires-comfyui *= *"[^0-9]*([0-9]+\.[0-9]+\.[0-9]+).*/\1/p' pyproject.toml)"
+            if [ -z "$VERSION" ]; then echo "could not parse requires-comfyui from pyproject.toml"; exit 1; fi
+            echo "COMFY_MODE=pinned"      >> "$GITHUB_ENV"
+            echo "COMFY_REF=v$VERSION"    >> "$GITHUB_ENV"
+          fi
+
+      - name: Set up ComfyUI checkout
+        # COMFY_DIR is exported by the self-hosted runner's .env and MUST be a
+        # runner-owned ComfyUI clone (never your dev checkout — this step does
+        # git reset --hard and rewrites custom_nodes). Cloned on first run.
+        run: |
+          set -euo pipefail
+          if [ -z "${COMFY_DIR:-}" ]; then echo "COMFY_DIR unset"; exit 1; fi
+          # Init-in-place rather than `git clone`: COMFY_DIR may already hold the
+          # cached checkpoint (models/checkpoints) or converted .mlmodelc, and
+          # `git clone` refuses a non-empty target. init + fetch + `checkout -f`
+          # populates the ComfyUI tree while leaving untracked files (the
+          # checkpoint, the cached models) untouched — so setup order is free.
+          if [ ! -d "$COMFY_DIR/.git" ]; then
+            echo "initialising ComfyUI repo in $COMFY_DIR"
+            mkdir -p "$COMFY_DIR"
+            git -C "$COMFY_DIR" init -q
+          fi
+          git -C "$COMFY_DIR" remote get-url origin >/dev/null 2>&1 \
+            || git -C "$COMFY_DIR" remote add origin https://github.com/comfyanonymous/ComfyUI.git
+          git -C "$COMFY_DIR" fetch --quiet origin
+          if [ "$COMFY_MODE" = "latest" ]; then
+            git -C "$COMFY_DIR" checkout -f -B master origin/master
+          else
+            git -C "$COMFY_DIR" checkout -f "$COMFY_REF"
+          fi
+          COMFY_SHA="$(git -C "$COMFY_DIR" rev-parse HEAD)"
+          echo "COMFY_SHA=$COMFY_SHA" >> "$GITHUB_ENV"
+          echo "Tier 2 mode=$COMFY_MODE, ComfyUI \`$COMFY_SHA\`" >> "$GITHUB_STEP_SUMMARY"
+
+          # Point ComfyUI's custom-node loader at this checkout. Refresh the
+          # symlink only; refuse to clobber a real directory (guards against a
+          # COMFY_DIR that is accidentally a dev checkout).
+          NODE_LINK="$COMFY_DIR/custom_nodes/ComfyUI-CoreMLSuite"
+          if [ -e "$NODE_LINK" ] && [ ! -L "$NODE_LINK" ]; then
+            echo "ERROR: $NODE_LINK is a real directory, not a symlink."
+            echo "COMFY_DIR must be a runner-owned ComfyUI, not your dev checkout."
+            exit 1
+          fi
+          mkdir -p "$COMFY_DIR/custom_nodes"
+          ln -sfn "$GITHUB_WORKSPACE" "$NODE_LINK"
+
+      - name: Install dependencies
+        run: |
+          set -euo pipefail
+          if [ "$COMFY_MODE" = "latest" ]; then
+            # Node deps (our coremltools-9 toolchain), then ComfyUI's own
+            # requirements for the pulled SHA, capped by the toolchain ceiling.
+            uv sync
+            uv pip install -r "$COMFY_DIR/requirements.txt" \
+              -c constraints/comfy-ceiling.txt
+          else
+            # Pinned gate: the frozen group mirrors the known-good pinned SHA.
+            uv sync --group comfy
+          fi
+
+      - name: Start ComfyUI server (background)
+        run: |
+          cd "$COMFY_DIR"
+          nohup "$GITHUB_WORKSPACE/.venv/bin/python" main.py --port 8188 --cpu-vae > /tmp/comfyui-ci.log 2>&1 &
+          # Poll the HTTP endpoint for readiness — robust to startup-banner
+          # wording / colored-log changes in a floating-latest ComfyUI.
+          for _ in $(seq 1 90); do
+            if curl -sf -o /dev/null http://127.0.0.1:8188/system_stats; then
+              echo "comfy ready (ComfyUI ${COMFY_SHA:-unknown})"; exit 0
+            fi
+            sleep 2
+          done
+          echo "comfy failed to start"; tail -100 /tmp/comfyui-ci.log; exit 1
+
+      - name: Purge cached Core ML UNets (force fresh conversion)
+        # The converter skips when a model of the same name already exists. That
+        # cache key is conversion *parameters* only, not the conversion code or
+        # toolchain — so a stale model would let a conversion regression pass.
+        # Clear it so every Tier 2 run exercises the full convert -> compile ->
+        # sample path end to end.
+        run: |
+          rm -rf "$COMFY_DIR"/models/unet/*.mlpackage "$COMFY_DIR"/models/unet/*.mlmodelc || true
+
+      - name: Run Tier 2 (m2 marker)
+        # Drives the Core ML Converter node, which converts the UNet from the
+        # checkpoint on every run (cache purged above).
+        run: uv run --no-sync pytest -m m2 tests/ -v
+
+      - name: Stop ComfyUI server
+        if: always()
+        run: pkill -f "main.py.*8188" || true

.gitignore

@@ -1,3 +1,6 @@
 playground/
-experiments/
 __pycache__/
+models/
+.venv/
+test_results/
+.claude/

.python-version

@@ -0,0 +1 @@
+3.12

CONVERTER_EXTRACTION_SPEC.md

@@ -0,0 +1,618 @@
+# ComfyUI-CoreMLSuite — Converter Extraction Spec for Claude Code
+
+> **Companion to `MODERNIZATION_SPEC.md`.** That spec hardens the repo and (Phase 3)
+> splits the *inference* math from the framework. **This** spec splits the *conversion*
+> path (`safetensors → CoreML`) out into a standalone, `comfy`-free, pip-installable
+> package that CoreMLSuite then depends on — and that other projects (incl. on-device
+> iOS tooling) can reuse.
+>
+> **Same discipline as the modernization spec:** safety-net first, behavior-preserving
+> until told otherwise, one phase = one branch = one PR, `STOP — VALIDATE` gate between
+> every phase, golden-latent as the regression anchor. `[M2]` = needs macOS/Apple Silicon;
+> `[M2-ANE]` = needs the Neural Engine. Everything else must run on plain Linux/CI.
+
+---
+
+## 0. How to work (read first — non-negotiable)
+
+1. **Behavior-preserving until Phase E6.** Phases E1–E5 must not change image output, node
+   names, `INPUT_TYPES` field names, or `NODE_CLASS_MAPPINGS` keys. The node graph is the
+   public contract; saved user-workflow JSON breaks if these change.
+2. **The conversion package produces an artifact and stops there.** Its job ends at a written
+   `.mlpackage` / `.mlmodelc` on disk. It must NOT import `comfy`, `folder_paths`, or
+   `comfy_extras`, and must NOT know ComfyUI's `models/unet` layout. Paths are *inputs*.
+3. **The runtime loader stays in the suite.** The loader is the **local** `coreml_suite.coreml_model.CoreMLModel`
+   — a thin wrapper over `coremltools.models.MLModel` (NOT Apple's
+   `python_coreml_stable_diffusion.coreml_model.CoreMLModel`, which is no longer used; see #58).
+   It *runs* a compiled model in Python — a desktop/Python inference concern, not a conversion
+   concern. It is NOT moved into the package. (On iOS the `.mlmodelc` is loaded natively; the
+   package's output is the deliverable, not a Python runner.)
+4. **Decouple in-repo before splitting repos.** Phases E1–E4 create the package *inside this
+   repo* and prove equivalence. The physical second-repo split is Phase E5, only after the
+   golden latent is proven identical. Do not create a second repository before Gate E4 passes.
+5. **Reuse the existing regression anchor.** The golden latent / PSNR anchor from
+   `MODERNIZATION_SPEC.md` Phase 2 is the cross-cutting proof for every gate here. If it is not
+   yet captured, capture it first (it is a prerequisite for E2 onward).
+6. **No new runtime dependencies** without flagging in the gate report (name, why, license, size).
+7. **A failing gate means stop and report**, not work around into the next phase.
+8. **Tooling is `uv`, not bare `pip`/`venv`.** Every environment/install/lock step uses the
+   project's `uv` toolchain: `uv venv`, `uv pip install`, `uv pip install -e .`, `uv lock`,
+   `uv run pytest`, `uv export`/`uv pip freeze` for baselines. Where this spec says "fresh venv",
+   read "`uv venv` + `uv pip install`". Reserve `uv pip` (not `pip`) inside that venv too.
+9. **The package is the single source of truth for *what is possible*; the node is a thin,
+   discovery-driven frontend.** See the "Interface contract" pillar below — this is the
+   maintainer's hard requirement and it overrides the earlier (now-rescinded) "freeze the
+   dropdown list" instruction.
+
+---
+
+## Interface contract (the maintainer's hard requirement) — read before any phase
+
+Two coupled guarantees must hold once the package is split out:
+
+**(A) Updating the converter must NOT require updating CoreMLSuite.**
+This is satisfied by treating the package's public surface as a versioned contract:
+- `convert(...)` and `compile_model(...)` are **keyword-only with defaults** for everything
+  past the genuinely-required positionals (`ckpt_path`, `model_version`, `out_path`). New
+  capabilities are added as new keyword args with defaults, so an old Suite's call still
+  validates against a newer package. **Never** reorder or rename existing parameters.
+- `compose_out_name` (the `.mlpackage` filename = the cache key) **moves into the package** and
+  is versioned with it. The Suite must not carry its own copy; if the package changes the naming
+  scheme that is a **major** bump (old cached artifacts stop resolving).
+
+**(B) CoreMLSuite must be able to list *new* conversion types WITHOUT a Suite code change or
+version bump.** Today the node hardcodes its dropdowns:
+```python
+"model_version": ([ModelVersion.SD15.name, ModelVersion.SDXL.name],),  # hand-typed, also INCOMPLETE (no LCM / SDXL_REFINER)
+"attention_implementation": (list(ATTENTION_IMPLEMENTATIONS),),         # from coreml_suite.attention
+"quantize_nbits": (list(QUANT_NBITS_VALUES), {"default": "none"}),      # from coreml_suite.core.naming
+```
+These are replaced by **runtime discovery calls into the package**, evaluated inside
+`INPUT_TYPES` (ComfyUI re-evaluates `INPUT_TYPES` on every plugin load):
+```python
+import coreml_diffusion
+"model_version":            (coreml_diffusion.list_model_versions(),),
+"attention_implementation": (coreml_diffusion.list_attention_impls(),),
+"quantize_nbits":           (coreml_diffusion.list_quant_modes(), {"default": "none"}),
+```
+Effect: `uv pip install -U coreml_diffusion` + ComfyUI restart surfaces any newly-added type in the old
+plugin's dropdown — **no Suite edit, no Suite version bump.** This is the requirement.
+
+**The cost, stated honestly (accept this trade-off explicitly at Gate E0):**
+- The Suite becomes a "dumb" frontend; the package is the sole authority on what conversions
+  exist. The Suite can no longer guarantee its saved workflows are valid against *arbitrary*
+  future package versions.
+- Therefore the package's discovery identifiers (`ModelVersion` values, attn-impl strings, quant
+  modes) are an **ADDITIVE-ONLY contract**: the package may *add* identifiers freely (minor bump,
+  no Suite change); **removing or renaming an identifier is a breaking change requiring a MAJOR
+  bump and a migration note**, because a saved workflow JSON references these strings verbatim.
+  Without this rule, "no version bump" silently becomes "randomly broken workflows."
+- `INPUT_TYPES` must **fail soft** when the package is missing/old: wrap the discovery calls so a
+  missing `coreml_diffusion` (or an old one lacking a `list_*` function) yields a sane fallback list and a
+  logged warning, instead of the node failing to register and disappearing from the menu.
+
+**Discovery API the package must expose (stable names):**
+```python
+coreml_diffusion.list_model_versions() -> list[str]   # VERIFIED ones only, e.g. ["SD15","SDXL"] today (.name — see seam.md)
+coreml_diffusion.list_attention_impls() -> list[str]  # ["SPLIT_EINSUM","SPLIT_EINSUM_V2","ORIGINAL"]
+coreml_diffusion.list_quant_modes() -> list[str]      # ["none","8","6","4"]
+coreml_diffusion.CONTRACT_VERSION: str                # bump rules above; Suite may log/compare it
+```
+These return the *display strings already used today*, so existing workflows keep validating.
+
+**Verification status is a PACKAGE property, not a node hardcode (maintainer's intent).**
+The Suite wants to expose *every model the converter can verifiably convert*. Today `lcm` and
+`sdxl_refiner` are absent from the converter node not because the Suite chooses to hide them, but
+because they lack a full golden/PSNR verification. So the gating lives in the package as a status:
+```python
+from enum import Enum
+class Status(Enum):
+    VERIFIED = "verified"          # has a golden anchor + passing [M2-ANE] check
+    EXPERIMENTAL = "experimental"  # convertible but not yet anchored/verified
+
+# internal registry, single source of truth.
+# KEY by ModelVersion enum MEMBER so list_* can emit .name. Keying by the lowercase
+# .value string returns ["sd15",...], which the node reverses via ModelVersion[...] -> KeyError.
+_MODEL_STATUS = {ModelVersion.SD15: Status.VERIFIED, ModelVersion.SDXL: Status.VERIFIED,
+                 ModelVersion.SDXL_REFINER: Status.EXPERIMENTAL, ModelVersion.LCM: Status.EXPERIMENTAL}
+
+def list_model_versions(include_experimental: bool = False) -> list[str]:
+    return [v.name for v, s in _MODEL_STATUS.items()   # .name -> "SD15","SDXL"; node reverses with ModelVersion[...]
+            if s is Status.VERIFIED or (include_experimental and s is Status.EXPERIMENTAL)]
+```
+Consequence: **promoting a model to VERIFIED in the package expands the Suite's dropdown with no
+Suite change and no Suite bump** — exactly the requirement. The act of verification (E-LCM
+produces an LCM golden anchor; same later for refiner) is what flips the status. The Suite's
+converter node calls `list_model_versions()` (verified-only); a power-user/CLI path may pass
+`include_experimental=True`. Promotion VERIFIED-from-EXPERIMENTAL is additive (minor bump);
+demotion or removal is breaking (major bump + note).
+
+---
+
+## Naming & layout (chosen — frozen at Gate E0)
+
+**Distribution name (PyPI):** `coreml-diffusion`. **Import name (Python):** `coreml_diffusion`.
+(PyPI normalizes `-`/`_`; the distribution uses the hyphen, the importable module the underscore.)
+Availability checked: both `coreml-diffusion` and the near variants were free on PyPI at E0.
+
+**Why this name (the positioning it encodes):** the project's niche is *diffusion models on Apple
+Neural Engine via CoreML, inside ComfyUI and on-device* — **not** Stable Diffusion specifically.
+`sd*` was rejected because it falsely narrows scope to SD; `coreml-diffusion` keeps `coreml` on the
+front for discoverability while `diffusion` honestly states the scope (SD/SDXL/LCM today, Flux and
+other diffusion architectures later) **without** promising arbitrary non-diffusion torch models,
+whose tracing/shape/sample-input pipeline differs. The name must not be re-narrowed to SD in
+future docs. ANE is the *differentiator* (documented in the README), but `coreml` was chosen over
+`ane` in the name for search discoverability per maintainer decision.
+
+Target package layout (framework-free — zero `comfy` imports):
+
+```
+coreml_diffusion/
+  __init__.py        # public API surface (see "Public API" below)
+  model_version.py   # ModelVersion enum — the SINGLE source of truth, no comfy
+  attention.py       # ATTENTION_IMPLEMENTATIONS tuple (from coreml_suite/attention.py) + apply_attention_implementation
+  pipeline.py        # get_pipeline (from_single_file), get_unet (cml UNet from ref unet)
+  unet.py            # UNet2DConditionModelLCM (moved from coreml_suite/lcm/unet.py)
+  inputs.py          # get_sample_input, lcm_inputs, sdxl_inputs,
+                     #   get_encoder_hidden_states_shape, get_coreml_inputs, get_inputs_spec
+  controlnet.py      # add_cnet_support (conversion-side residual SHAPE calc only)
+  convert.py         # convert_unet, convert (orchestration), convert_to_coreml, load_coreml_model
+  compile.py         # compile_coreml_model
+  quantize.py        # (Phase E6 / MODERNIZATION Phase 6 lands here) palettization 4/6/8-bit
+  cli.py             # console entry point: `coreml-diffusion convert ...`
+  pyproject.toml     # standalone packaging (at E5)
+```
+
+What stays in `coreml_suite/` (the ComfyUI side, thinned):
+- `nodes.py` — still owns **name-encoding** (`out_name` construction), path resolution via
+  `folder_paths`, the node `INPUT_TYPES`/mappings, and wrapping the result in `CoreMLModel`.
+- `models.py`, `latents.py`, `controlnet.py` (inference parts), `lcm/utils.py`, `config.py`
+  (inference config build) — untouched by this spec except the import-source of `ModelVersion`.
+
+### Public API (the contract `coreml_diffusion` exposes)
+```python
+from coreml_diffusion import ModelVersion, convert, compile_model, compose_out_name
+from coreml_diffusion import list_model_versions, list_attention_impls, list_quant_modes, CONTRACT_VERSION
+
+# Mirror the CURRENT converter.py signature, made keyword-only past the required positionals
+# and with paths/device injected (no folder_paths, no comfy.model_management):
+# convert(ckpt_path, model_version, out_path, *,
+#         batch_size=1, sample_size=(64, 64), controlnet_support=False,
+#         lora_weights=None, attn_impl=list_attention_impls()[0], config_path=None,
+#         quantize_nbits="none", device=None) -> None   # side effect: writes out_path
+#   (current convert() returns None and writes via convert_unet → coreml_unet.save; keep that,
+#    or change to `return out_path` as a deliberate, documented improvement — pick one at E0.)
+# compile_model(src_path, out_dir, final_name) -> str   # returns compiled .mlmodelc path
+```
+Note: `convert` takes an **explicit `out_path`** — no `folder_paths`. `device` is injected
+(defaults to torch's default device). `compose_out_name` lives here (cache-key contract) and the
+node imports it from the package. The `list_*` discovery functions back the node's dropdowns.
+
+---
+
+## The import chains to cut (root cause inventory) — REVISED against current code
+
+> **State note (verified):** the code moved on since the original draft. Several chains are
+> already cut. Re-verify each line by `grep` before acting; do not assume the original draft.
+
+**Already done (verify, then skip):**
+- ✅ `converter.py` already imports `from coreml_suite.model_version import ModelVersion`, and
+  `model_version.py` is **clean** (`from enum import Enum` only — zero comfy). The old
+  "converter → config → comfy" chain is **already broken**. `config.py` still imports comfy, but
+  it is **inference-side** (`get_model_config` via `supported_models_base`/`latent_formats`) —
+  *not* on the conversion path. Do **not** treat `config.py` as a converter dependency.
+- ✅ `converter.py` now uses `diffusers.UNet2DConditionModel.from_single_file` and a local
+  `CoreMLUNetWrapper` (in `coreml_suite/conversion/unet.py`) — it is **no longer** importing the
+  Apple `python_coreml_stable_diffusion.unet.UNet2DConditionModel*` internals on the main path.
+  A `coreml_suite/conversion/` subpackage already exists (`attention`, `shapes`, `trace`, `unet`).
+- ✅ Name-encoding already extracted to `coreml_suite/core/naming.py` (`compose_out_name`,
+  `lora_names_from_params`, `ATTN_SUFFIX`, `QUANT_NBITS_VALUES`) **with characterization tests**
+  (`tests/unit/test_characterization_out_name.py`). The pure-naming split is done.
+- ✅ Quantization is **already implemented** in `converter.py` (`quantize_nbits`, k-means
+  `palettize_weights`) and surfaced as an optional node input. Phase E6 is therefore *move*, not
+  *build* (see revised E6).
+
+**Still to cut (the real remaining work):**
+1. `coreml_suite/converter.py::get_out_path` → `from folder_paths import get_folder_paths`.
+   Main converter still reaches into ComfyUI's model dir. **Cut: `out_path` is an injected arg;
+   `folder_paths` resolution moves up into the node** (the node already computes `out_name`).
+2. `coreml_suite/lcm/converter.py` → still has its **own** `from folder_paths import
+   get_folder_paths` (`get_out_path`) and (per original draft) `comfy.model_management`. Verify
+   the current LCM file and cut both: inject `out_path` and `device`.
+3. Global mutation of the attention impl: confirm where it now lives. Main path appears to route
+   through `coreml_suite/conversion/attention.apply_attention_implementation` (cleaner than the
+   old global), but `lcm/converter.py` may still set a module global at import. **Ensure the
+   package sets attention per-call, never at import time.**
+4. **Duplication LCM vs main:** `lcm/converter.py` still carries its own copies of
+   `convert_to_coreml`, `load_coreml_model`, `get_out_path`, `get_sample_input` (the LCM variant
+   takes a `scheduler` arg), and hardcodes `SimianLuo/LCM_Dreamshaper_v7`. **Dedupe into the
+   single `coreml_diffusion` implementation;** the HF-hardcode consolidation is the *behavior-changing*
+   part → deferred to optional **E-LCM**, not E1–E5.
+5. **`compose_out_name` ownership:** currently in `coreml_suite/core/naming.py` and called by the
+   node. Per the Interface-contract pillar it must **move into the package** (it is the cache-key
+   contract) and the node must import it from `coreml_diffusion`, not keep a copy.
+
+---
+
+## Phase E0 — Seam decision & inventory (no code change)
+
+**Objective:** lock the cut line, the interface contract, and naming so later phases don't drift.
+
+### Tasks
+1. Produce `docs/extraction/seam.md`: a table of every symbol in `converter.py`,
+   `lcm/converter.py`, `lcm/unet.py`, **plus the already-extracted `conversion/` subpackage
+   (`attention`, `shapes`, `trace`, `unet`) and `core/naming.py`**, classified
+   **CONVERSION → coreml_diffusion** vs **STAYS (comfy/node)**. Note which are already framework-free.
+2. ~~Confirm the current `python_coreml_stable_diffusion` footprint.~~ **DONE (seam.md §6):
+   footprint is ZERO** — no runtime imports anywhere; only a docstring mention in
+   `core/__init__.py:4`. Main path uses `diffusers` + local `CoreMLUNetWrapper`; the runtime
+   `CoreMLModel` (STAYS in suite) is a local coremltools wrapper, not Apple's. No shape/attn helper
+   comes from Apple (local `conversion/shapes.py`, `conversion/attention.py`).
+3. **Decide the interface contract concretely (the maintainer's hard requirement):**
+   - Discovery functions `list_model_versions / list_attention_impls / list_quant_modes` live in
+     the package and return today's display strings verbatim. Node `INPUT_TYPES` calls them.
+   - `ModelVersion` values, attn-impl strings, quant modes are **ADDITIVE-ONLY** across package
+     versions; removal/rename = MAJOR bump + migration note. Write this into the package's
+     versioning policy doc now.
+   - `compose_out_name` moves to the package; node imports it (no copy). Confirm the
+     characterization tests in `test_characterization_out_name.py` will be re-pointed, not
+     duplicated.
+   - **Resolve the `model_version` dropdown question (maintainer decided):** the Suite exposes
+     *every model the converter can verifiably convert*. `lcm` and `sdxl_refiner` are absent today
+     only because they lack a golden/PSNR verification — **not** because the node hardcodes a
+     short list. Encode this as a **status registry in the package** (`VERIFIED` vs
+     `EXPERIMENTAL`); `list_model_versions()` returns VERIFIED-only by default. The converter node
+     calls it plainly. Promoting LCM/refiner to VERIFIED (after E-LCM / a refiner anchor) expands
+     the dropdown with **no Suite change**. Do NOT add permanent per-node filtering — the gate is
+     verification status, owned by the package.
+4. ~~Confirm the `ml-stable-diffusion` git dep is pinned.~~ **N/A — already removed (#58).** Verified:
+   zero `python_coreml_stable_diffusion` imports in the repo; `CoreMLModel` is now a local
+   coremltools wrapper; the dep is absent from `pyproject.toml`/`requirements.txt`. No SHA to pin.
+
+### STOP — VALIDATE (Gate E0)
+```
+## Gate E0 report
+- seam.md committed: <path>; symbol counts (move / stay / already-framework-free)
+- python_coreml_stable_diffusion usage (verified by grep): conversion=<list> runtime=<list>
+- Discovery API signatures frozen: list_model_versions (verified-only) / list_attention_impls / list_quant_modes
+- Status registry decided: sd15+sdxl=VERIFIED, lcm+sdxl_refiner=EXPERIMENTAL (gated, not hidden)
+- Additive-only contract policy doc written (incl. promotion=minor, demotion/removal=major): <path>
+- model_version dropdown: expose all (incl. LCM/REFINER) / filtered per node — DECISION: <...>
+- compose_out_name move-not-copy confirmed; tests re-point plan: <...>
+- LCM consolidation deferred to optional E-LCM: YES/NO
+- ml-stable-diffusion: N/A — already removed (#58), not a dependency (was: pin-or-BLOCKER)
+- Package name in-repo: coreml_diffusion (final PyPI name deferred to E5)
+```
+
+---
+
+## Phase E1 — Establish `coreml_diffusion` package + discovery API (mostly verification)
+
+**Objective:** stand up the package namespace and the discovery surface. Much of the comfy-chain
+cut is **already done** — this phase mostly *verifies* that and adds the discovery functions.
+
+### Tasks
+1. **Verify (don't redo):** `coreml_suite/model_version.py` is already clean (`Enum` only). Confirm
+   `import coreml_suite.model_version` works with **no comfy** (`uv run python -c "..."` in a
+   comfy-free `uv venv`). If true, E1's original "extract ModelVersion" task is already satisfied.
+2. Create the `coreml_diffusion/` package skeleton with `__init__.py` exporting the **discovery API**
+   backed by the *existing* sources of truth for now (re-export `ModelVersion`, the
+   `ATTENTION_IMPLEMENTATIONS` tuple, and `QUANT_NBITS_VALUES`) so values are byte-identical:
+   ```python
+   def list_model_versions(): return [v.name for v in ModelVersion]   # .name -> "SD15" (node reverses via ModelVersion[...]; .value KeyErrors)
+   def list_attention_impls(): return list(ATTENTION_IMPLEMENTATIONS)
+   def list_quant_modes(): return list(QUANT_NBITS_VALUES)
+   CONTRACT_VERSION = "1.0"
+   ```
+   (At this stage `coreml_diffusion` may live inside the repo and import from `coreml_suite.*`; the
+   physical move of implementation happens in E2. The point of E1 is to freeze the *contract*.)
+3. **Decided (`.name`):** the node renders `ModelVersion.SD15.name` (`"SD15"`) and reverses the
+   dropdown string via `ModelVersion[model_version]` (name lookup, `nodes.py:286`). Discovery API
+   therefore returns `.name`; `.value` (`"sd15"`) would `KeyError`. Recorded in `seam.md` §5.
+
+### Acceptance criteria
+- `uv run python -c "import coreml_diffusion; print(coreml_diffusion.list_model_versions(), coreml_diffusion.list_quant_modes())"`
+  works in a **comfy-free** `uv venv` and prints today's exact strings.
+- Existing characterization tests pass unchanged.
+- No node behavior change yet (node still uses its current hardcoded lists in E1).
+
+### STOP — VALIDATE (Gate E1)
+```
+## Gate E1 report
+- model_version.py confirmed comfy-free (uv, no comfy): PASS/FAIL
+- coreml_diffusion.list_* returns byte-identical strings to current dropdowns: YES/NO (show values)
+- .name vs .value decision for model_version discovery: <...>
+- CONTRACT_VERSION set; additive-only policy linked: <path>
+- Characterization tests unchanged & green (uv run pytest): YES/NO
+```
+
+---
+
+## Phase E2 — Move conversion code into `coreml_diffusion` (in-repo, dedup, behavior-preserving)
+
+**Objective:** physically relocate the conversion mechanics into the framework-free package,
+collapsing the two duplicate converters into one, with paths/device injected.
+
+### Tasks
+1. Move into `coreml_diffusion/`: `pipeline.py` (`get_pipeline`, `get_unet`), `unet.py`
+   (`UNet2DConditionModelLCM`), `inputs.py` (sample/lcm/sdxl input builders +
+   `get_encoder_hidden_states_shape` + `get_coreml_inputs` + `get_inputs_spec`),
+   `controlnet.py` (`add_cnet_support`), `convert.py` (`convert_unet`, `convert`,
+   `convert_to_coreml`, `load_coreml_model`), `compile.py` (`compile_coreml_model`).
+2. **Dedupe LCM vs main** (the real remaining duplication): delete `lcm/converter.py`'s copies of
+   `convert_to_coreml` / `load_coreml_model` / `get_out_path` / `get_sample_input` (LCM variant
+   carries a `scheduler` arg — fold that into the shared `get_sample_input` as an optional param)
+   in favor of the single `coreml_diffusion` implementation. The main path's helpers
+   (`get_unet`/`get_encoder_hidden_states_shape`/`get_coreml_inputs`/`convert_unet`/`convert`) and
+   the `conversion/` subpackage (`attention`, `shapes`, `trace`, `unet`) move as-is.
+3. **Inject paths**: replace `get_out_path`'s `folder_paths` reach-in with an injected `out_path`
+   argument on `convert(...)`; `folder_paths` resolution moves up into the node (which already
+   computes `out_name`). No `folder_paths` import anywhere in `coreml_diffusion`.
+4. **Inject device** where the LCM path used `comfy.model_management` (verify it still does):
+   `convert(..., device=None)`, default to torch's default device.
+5. **Attention per-call, never at import:** main path already routes through
+   `conversion/attention.apply_attention_implementation` — keep that. If `lcm/converter.py` still
+   sets any module global at import, remove it; the package sets attention from the `attn_impl`
+   arg inside `convert`.
+6. **Move `compose_out_name` into the package** (`coreml_diffusion/naming.py`); re-point
+   `test_characterization_out_name.py` imports to `coreml_diffusion.naming` — assertions and values
+   unchanged. The node will import it from the package in E3.
+7. Leave **thin shims** in `coreml_suite/converter.py` and `coreml_suite/lcm/converter.py` that
+   re-export from `coreml_diffusion`, preserving the old call signatures the nodes use (nodes untouched
+   this phase). Shims map comfy `folder_paths`/device into package args.
+
+### Acceptance criteria
+- `uv run pytest -m unit` (Tier 0) imports `coreml_diffusion.*` with **no comfy / no MPS** and is green on Linux.
+- The dedup leaves exactly one implementation of each previously-duplicated function.
+- Characterization tests pass unchanged after the `compose_out_name` re-point.
+- `[M2]` A real SD1.5 conversion via the shim still produces a loadable model.
+- `[M2-ANE]` **Golden latent identical / within tolerance** to the MODERNIZATION Phase 2 anchor
+  (same seed/prompt) — proves the move + dedup changed nothing.
+
+### STOP — VALIDATE (Gate E2 — first regression gate)
+```
+## Gate E2 report
+- Tier 0 import of coreml_diffusion without comfy/MPS (uv run): PASS/FAIL
+- LCM/main duplicated funcs collapsed to one (list old→new): <map>
+- compose_out_name moved to package; char-tests re-pointed & green: YES/NO
+- Paths injected (no folder_paths in package): confirmed
+- Device injected (no comfy.model_management in package): confirmed
+- Attention set per-call, not at import (both main & lcm): confirmed
+- [M2-ANE] Golden latent vs Phase-2 anchor: identical / within tol <x> / DIVERGED (STOP)
+- Node INPUT_TYPES / mappings untouched: confirmed (diff)
+```
+**If the golden latent diverged at all, STOP and report — do not continue.**
+
+---
+
+## Phase E3 — Thin the nodes onto the package (behavior-preserving)
+
+**Objective:** remove the shims; have the ComfyUI nodes call `coreml_diffusion` directly, keeping the
+node contract byte-identical.
+
+### Tasks
+1. `CoreMLConverter.convert` (in `coreml_suite/nodes.py`): keep the `folder_paths`-based path
+   resolution **in the node**; import `compose_out_name` from `coreml_diffusion` (not `coreml_suite.core`);
+   call `coreml_diffusion.convert(...)` and `coreml_diffusion.compile_model(...)` directly; wrap the compiled path
+   in `CoreMLModel`.
+2. **Wire the dropdowns to discovery (the maintainer's hard requirement).** Replace the hardcoded
+   `INPUT_TYPES` lists with fail-soft discovery calls:
+   ```python
+   def _discover(fn, fallback):
+       try:
+           import coreml_diffusion
+           return getattr(coreml_diffusion, fn)()
+       except Exception as e:        # missing/old package, or import error
+           logger.warning(f"coreml_diffusion.{fn} unavailable ({e}); using fallback {fallback}")
+           return fallback
+   ...
+   "model_version":            (_discover("list_model_versions", ["SD15", "SDXL"]),),
+   "attention_implementation": (_discover("list_attention_impls", ["SPLIT_EINSUM","SPLIT_EINSUM_V2","ORIGINAL"]),),
+   "quantize_nbits":           (_discover("list_quant_modes", ["none","8","6","4"]), {"default": "none"}),
+   ```
+   This is what makes "update the package → new types appear in the old node, no Suite bump" true.
+3. `COREML_CONVERT_LCM` (in `coreml_suite/lcm/nodes.py`): route through `coreml_diffusion` for the shared
+   mechanics. **Keep the existing LCM behavior/HF-hardcode for now** — consolidation is optional E-LCM.
+4. Delete the now-dead `coreml_suite/converter.py` / `coreml_suite/lcm/converter.py` shims (or
+   reduce to a one-line re-export if anything external imports them — grep first).
+
+### Acceptance criteria
+- `NODE_CLASS_MAPPINGS` / `NODE_DISPLAY_NAME_MAPPINGS` keys: **unchanged** (diff `__init__.py`).
+- Every `INPUT_TYPES` **field name** unchanged. Dropdown **values**: the discovery calls must
+  return **a superset of** today's values, with every previously-present value still present and
+  spelled identically (additive-only). *(This deliberately replaces the original spec's
+  "values must be byte-identical/frozen" criterion — the maintainer requires the list be
+  extensible at runtime. Frozen-field-names + additive-only-values is the new contract.)*
+- With `coreml_diffusion` **absent**, the node still registers and shows the fallback lists (fail-soft).
+- `[M2-ANE]` Golden latent still identical to the Phase-2 anchor.
+- `[M2-ANE]` The committed e2e workflow `tests/integration/...` still passes (PSNR > 25).
+
+### STOP — VALIDATE (Gate E3)
+```
+## Gate E3 report
+- Node mappings diff: empty (confirmed)
+- INPUT_TYPES field-names diff: empty (confirmed)
+- Dropdown values: superset of prior, all prior values still present & identical: YES/NO (show)
+- Fail-soft with coreml_diffusion absent (node still registers): PASS/FAIL
+- compose_out_name now imported from coreml_diffusion (no node-side copy): confirmed
+- [M2-ANE] Golden latent vs anchor: identical / within tol / DIVERGED (STOP)
+- [M2-ANE] e2e workflow PSNR: <value> (> 25?)
+- Dead converter shims removed / reduced: <list>
+```
+
+---
+
+## Phase E4 — Standalone packaging & CLI (still in-repo)
+
+**Objective:** make `coreml_diffusion` independently installable and usable without ComfyUI, with a CLI
+suitable for the planned article and for on-device/iOS conversion workflows.
+
+### Tasks
+1. Add `coreml_diffusion/pyproject.toml`: name (working `coreml_diffusion`), `requires-python`, dependencies
+   = `coremltools` (pinned to the MODERNIZATION-validated version), `diffusers`, `transformers`,
+   `peft`, `omegaconf`, `numpy`, `torch`. **No `ml-stable-diffusion`** (already removed in #58, see
+   §0.3) and **no comfy**. Suite pins `transformers>=4.44`/`peft>=0.13`/`omegaconf>=2.3` today;
+   grep-confirm each is on the conversion path before listing it. A `[project.scripts]` entry:
+   `coreml-diffusion = "coreml_diffusion.cli:main"`.
+2. `coreml_diffusion/cli.py`: `coreml-diffusion convert --ckpt PATH --model-version sd15 --out PATH
+   [--height --width --batch-size --attn-impl --controlnet --lora NAME:STRENGTH ... --config PATH]`
+   and `coreml-diffusion compile --src PATH --out-dir DIR --name NAME`. Mirrors `convert()`/`compile_model()`.
+3. Tier-0 Linux tests for the CLI **arg→call mapping** (mock the heavy `convert`); the real
+   convert remains `[M2]`. Add a `[M2]` smoke test: convert a tiny synthetic UNet end-to-end.
+4. README for the package: install, CLI usage, "produce a `.mlpackage`/`.mlmodelc` for use in a
+   Swift/iOS app", and the ANE positioning note (low-power, GPU-free, embeddable; SD1.5/SDXL on
+   ANE, **not** a Flux-speed claim).
+
+### Acceptance criteria
+- Fresh `python -m venv` + `uv pip install ./coreml-diffusion` (no ComfyUI present) imports and runs
+  `coreml-diffusion --help` and the arg-mapping tests on Linux.
+- `[M2]` `coreml-diffusion convert` produces a model file identical (golden) to the node path.
+
+### STOP — VALIDATE (Gate E4)
+```
+## Gate E4 report
+- uv pip install ./coreml-diffusion in comfy-free venv: PASS/FAIL (log)
+- CLI arg→call tests (Tier 0, Linux): green
+- [M2] CLI-produced model golden vs node-produced model: identical / DIVERGED
+- Package deps list (with pinned SHAs/versions + licenses):
+- New runtime deps vs suite before: <none / list>
+```
+**This is the gate that proves the package stands alone. Do not split repos before it passes.**
+
+---
+
+## Phase E5 — Physical split into a second repository
+
+**Objective:** move `coreml_diffusion/` to its own repo; CoreMLSuite depends on it by pinned version.
+
+### Tasks
+1. Create the new repo (maintainer action — agent prepares the tree, not the GitHub repo).
+   Choose final distributable name; rename imports if changed (single sweep, recorded).
+2. CoreMLSuite `pyproject.toml` / `requirements.txt`: replace the conversion-only deps with a
+   pinned dependency on the new package (`coreml_diffusion==<version>` from PyPI, or `git+...@<tag>`
+   until first PyPI release). (There is no `git+...ml-stable-diffusion` line to remove — already
+   gone since #58.)
+3. ~~Keep `python_coreml_stable_diffusion` for the loader.~~ **Void.** The loader is the local
+   `coreml_suite/coreml_model.py` over `coremltools`; the suite keeps `coremltools` as a direct dep
+   for it. No Apple lib involved.
+4. Set up the new repo's CI: Tier 0 on Linux (import + arg-mapping + input-shape math),
+   `[M2]`/`[M2-ANE]` on a self-hosted/macOS-ARM runner reusing the golden-latent anchor.
+5. Versioning: SemVer; first release `0.1.0`. Document the compatibility matrix
+   (coreml_diffusion ↔ coremltools version ↔ diffusers version). No ml-stable-diffusion axis.
+
+### Acceptance criteria
+- CoreMLSuite installs in a fresh venv pulling the new package; e2e workflow still passes `[M2-ANE]`.
+- New repo CI green on Linux (Tier 0) and `[M2-ANE]` golden latent matches the anchor.
+- No conversion code remains in CoreMLSuite (grep: no `ct.convert`, no `from_single_file`,
+  no `torch.jit.trace`).
+
+### STOP — VALIDATE (Gate E5)
+```
+## Gate E5 report
+- New repo tree prepared: <path/branch>; final package name: <name>
+- Suite depends on package by pinned version: <spec>
+- Suite e2e [M2-ANE] PSNR after split: <value> (> 25?)
+- Conversion code fully absent from suite: confirmed (grep output)
+- Compatibility matrix documented: <link>
+- First release tag: 0.1.0
+```
+
+---
+
+## Phase E6 — Quantization travels WITH the conversion code (already implemented → move)
+
+**Objective:** quantization is **already implemented** (k-means `palettize_weights` in
+`converter.py`, `quantize_nbits` node input, `_q<bits>` filename suffix, README tradeoff table).
+There is nothing to *build*. It simply **moves with the conversion code in E2** as part of
+`convert_unet`. This phase is a checkpoint that it survived the extraction intact, plus exposing
+it through the CLI.
+
+### Tasks
+1. Confirm the palettization block moved cleanly into `coreml_diffusion` (lives in `convert.py` or a
+   `quantize.py` helper called from `convert_unet`). Default `"none"` stays byte-identical.
+2. Expose via CLI flag `--quantize {none,8,6,4}` (E4 already lists this) and via
+   `list_quant_modes()` discovery (E1/E3).
+3. The existing README tradeoff table (SD1.5 1×512×512 SPLIT_EINSUM: none/8/6/4 → size/ms/PSNR)
+   moves to the package README. Re-confirm one row `[M2-ANE]` so the article can cite a live number.
+
+### Acceptance criteria
+- Default (`none`) output byte-identical to pre-extraction (covered by the E2/E3 golden latent).
+- `coreml-diffusion convert --quantize 4` produces a `_q4` artifact matching the node's `_q4` artifact `[M2]`.
+- `list_quant_modes()` drives the node dropdown (no hardcoded copy remains).
+
+### STOP — VALIDATE (Gate E6)
+```
+## Gate E6 report
+- Palettization relocated into coreml_diffusion, called from convert_unet: confirmed
+- Default none output identical (golden): YES/NO
+- [M2] CLI --quantize {8,6,4} artifacts match node artifacts: YES/NO
+- Tradeoff table in package README with at least one re-confirmed [M2-ANE] row: <link>
+```
+
+---
+
+## Phase E-LCM — FIRST task after the split: clean up LCM + verify → promote (behavior-changing, gated)
+
+> Promoted from "optional, someday" to **the first thing after E5**, per maintainer intent: the
+> Suite should expose every verifiably-convertible model, and LCM is the obvious first cleanup.
+
+Two coupled goals:
+1. **Consolidate the LCM path.** Make the LCM node use the unified `from_single_file` path in
+   `coreml_diffusion.convert(model_version=LCM, ...)` instead of the hardcoded `SimianLuo/LCM_Dreamshaper_v7`
+   HF download; drop the duplicated LCM helpers (already deduped in E2). **Behavior change** ⇒
+   capture an LCM golden anchor *before* the change, then prove within-tolerance after.
+2. **Verify → promote.** Once the LCM conversion has a passing `[M2-ANE]` golden anchor, flip
+   `_MODEL_STATUS["lcm"] = Status.VERIFIED` **in the package** (minor bump). The Suite's dropdown
+   gains `lcm` automatically — no Suite change, no Suite bump. This is the end-to-end proof that
+   the discovery contract works as designed.
+
+Repeat the same recipe for `sdxl_refiner` when it gets an anchor (separate small gate). Do NOT
+bundle E-LCM into E1–E5; it changes behavior and must stand on its own golden.
+
+### STOP — VALIDATE (Gate E-LCM)
+```
+## Gate E-LCM report
+- LCM golden anchor captured BEFORE change: <path/hash>
+- LCM node now uses unified from_single_file path; HF hardcode removed: confirmed
+- [M2-ANE] LCM golden after change: identical / within tol <x> / DIVERGED (STOP)
+- Status flipped lcm→VERIFIED in package (minor bump <ver>): confirmed
+- Suite dropdown now lists lcm with NO Suite code change / NO Suite bump: confirmed (diff empty)
+- LCM node accepts a checkpoint arg now (documented breaking-ish UI note): <link>
+```
+
+---
+
+## Article deliverable (after E4)
+
+Once the CLI exists and stands alone, the "convert a Comfy/A1111 workflow into an on-device iOS
+app" write-up becomes a clean tutorial: `coreml-diffusion convert` → `.mlmodelc` → load in Swift/CoreML.
+Frame the niche honestly per the README note above (ANE feasibility & power, not raw Flux speed).
+
+---
+
+## Quick reference: extraction gate discipline
+
+```
+E0  Seam decision, interface contract, discovery API frozen → Gate E0  (cut line + additive-only policy?)
+E1  Stand up coreml_diffusion + discovery API (mostly verify)        → Gate E1  (list_* byte-identical, comfy-free?)
+E2  Move conversion code, dedup LCM/main, inject paths/device→ Gate E2  (golden identical? duplicates gone?) ← first regression gate
+E3  Thin nodes onto package + wire discovery dropdowns       → Gate E3  (field-names frozen, values additive, fail-soft, golden identical?)
+E4  Standalone packaging + CLI (uv)                          → Gate E4  (uv pip install w/o comfy? CLI golden?) ← proves it stands alone
+E5  Physical second-repo split                               → Gate E5  (suite depends on pkg? conversion absent?)
+E-LCM  FIRST post-split: clean up LCM, verify → promote      → Gate E-LCM (LCM golden? dropdown gains lcm w/ no Suite bump?)
+E6  Quantization checkpoint (already built → moved in E2)    → Gate E6  (default identical? CLI quant matches?)
+(refiner)  same recipe as E-LCM when an anchor exists        → own small gate (promote sdxl_refiner→VERIFIED)
+```
+
+**Interface-contract invariants (the maintainer's hard requirement), restated:**
+- Package API is keyword-only-with-defaults past the required positionals → converter updates
+  don't force Suite updates.
+- Node dropdowns are discovery-driven (`coreml_diffusion.list_*`) + fail-soft → new conversion types
+  appear in the old plugin with `uv pip install -U coreml_diffusion`, **no Suite code change, no bump**.
+- Discovery identifiers are **additive-only**; removal/rename = MAJOR bump + migration note.
+- `compose_out_name` (cache key) lives in the package, single copy.
+
+
+**Golden rule (inherited): never cross a gate with a failing acceptance criterion.
+Stop, report, wait. The golden latent is the single source of truth that the extraction
+changed nothing.**

LICENSE

@@ -1,674 +1,21 @@
-                    GNU GENERAL PUBLIC LICENSE
-                       Version 3, 29 June 2007
-
- Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
- Everyone is permitted to copy and distribute verbatim copies
- of this license document, but changing it is not allowed.
-
-                            Preamble
-
-  The GNU General Public License is a free, copyleft license for
-software and other kinds of works.
-
-  The licenses for most software and other practical works are designed
-to take away your freedom to share and change the works.  By contrast,
-the GNU General Public License is intended to guarantee your freedom to
-share and change all versions of a program--to make sure it remains free
-software for all its users.  We, the Free Software Foundation, use the
-GNU General Public License for most of our software; it applies also to
-any other work released this way by its authors.  You can apply it to
-your programs, too.
-
-  When we speak of free software, we are referring to freedom, not
-price.  Our General Public Licenses are designed to make sure that you
-have the freedom to distribute copies of free software (and charge for
-them if you wish), that you receive source code or can get it if you
-want it, that you can change the software or use pieces of it in new
-free programs, and that you know you can do these things.
-
-  To protect your rights, we need to prevent others from denying you
-these rights or asking you to surrender the rights.  Therefore, you have
-certain responsibilities if you distribute copies of the software, or if
-you modify it: responsibilities to respect the freedom of others.
-
-  For example, if you distribute copies of such a program, whether
-gratis or for a fee, you must pass on to the recipients the same
-freedoms that you received.  You must make sure that they, too, receive
-or can get the source code.  And you must show them these terms so they
-know their rights.
-
-  Developers that use the GNU GPL protect your rights with two steps:
-(1) assert copyright on the software, and (2) offer you this License
-giving you legal permission to copy, distribute and/or modify it.
-
-  For the developers' and authors' protection, the GPL clearly explains
-that there is no warranty for this free software.  For both users' and
-authors' sake, the GPL requires that modified versions be marked as
-changed, so that their problems will not be attributed erroneously to
-authors of previous versions.
-
-  Some devices are designed to deny users access to install or run
-modified versions of the software inside them, although the manufacturer
-can do so.  This is fundamentally incompatible with the aim of
-protecting users' freedom to change the software.  The systematic
-pattern of such abuse occurs in the area of products for individuals to
-use, which is precisely where it is most unacceptable.  Therefore, we
-have designed this version of the GPL to prohibit the practice for those
-products.  If such problems arise substantially in other domains, we
-stand ready to extend this provision to those domains in future versions
-of the GPL, as needed to protect the freedom of users.
-
-  Finally, every program is threatened constantly by software patents.
-States should not allow patents to restrict development and use of
-software on general-purpose computers, but in those that do, we wish to
-avoid the special danger that patents applied to a free program could
-make it effectively proprietary.  To prevent this, the GPL assures that
-patents cannot be used to render the program non-free.
-
-  The precise terms and conditions for copying, distribution and
-modification follow.
-
-                       TERMS AND CONDITIONS
-
-  0. Definitions.
-
-  "This License" refers to version 3 of the GNU General Public License.
-
-  "Copyright" also means copyright-like laws that apply to other kinds of
-works, such as semiconductor masks.
-
-  "The Program" refers to any copyrightable work licensed under this
-License.  Each licensee is addressed as "you".  "Licensees" and
-"recipients" may be individuals or organizations.
-
-  To "modify" a work means to copy from or adapt all or part of the work
-in a fashion requiring copyright permission, other than the making of an
-exact copy.  The resulting work is called a "modified version" of the
-earlier work or a work "based on" the earlier work.
-
-  A "covered work" means either the unmodified Program or a work based
-on the Program.
-
-  To "propagate" a work means to do anything with it that, without
-permission, would make you directly or secondarily liable for
-infringement under applicable copyright law, except executing it on a
-computer or modifying a private copy.  Propagation includes copying,
-distribution (with or without modification), making available to the
-public, and in some countries other activities as well.
-
-  To "convey" a work means any kind of propagation that enables other
-parties to make or receive copies.  Mere interaction with a user through
-a computer network, with no transfer of a copy, is not conveying.
-
-  An interactive user interface displays "Appropriate Legal Notices"
-to the extent that it includes a convenient and prominently visible
-feature that (1) displays an appropriate copyright notice, and (2)
-tells the user that there is no warranty for the work (except to the
-extent that warranties are provided), that licensees may convey the
-work under this License, and how to view a copy of this License.  If
-the interface presents a list of user commands or options, such as a
-menu, a prominent item in the list meets this criterion.
-
-  1. Source Code.
-
-  The "source code" for a work means the preferred form of the work
-for making modifications to it.  "Object code" means any non-source
-form of a work.
-
-  A "Standard Interface" means an interface that either is an official
-standard defined by a recognized standards body, or, in the case of
-interfaces specified for a particular programming language, one that
-is widely used among developers working in that language.
-
-  The "System Libraries" of an executable work include anything, other
-than the work as a whole, that (a) is included in the normal form of
-packaging a Major Component, but which is not part of that Major
-Component, and (b) serves only to enable use of the work with that
-Major Component, or to implement a Standard Interface for which an
-implementation is available to the public in source code form.  A
-"Major Component", in this context, means a major essential component
-(kernel, window system, and so on) of the specific operating system
-(if any) on which the executable work runs, or a compiler used to
-produce the work, or an object code interpreter used to run it.
-
-  The "Corresponding Source" for a work in object code form means all
-the source code needed to generate, install, and (for an executable
-work) run the object code and to modify the work, including scripts to
-control those activities.  However, it does not include the work's
-System Libraries, or general-purpose tools or generally available free
-programs which are used unmodified in performing those activities but
-which are not part of the work.  For example, Corresponding Source
-includes interface definition files associated with source files for
-the work, and the source code for shared libraries and dynamically
-linked subprograms that the work is specifically designed to require,
-such as by intimate data communication or control flow between those
-subprograms and other parts of the work.
-
-  The Corresponding Source need not include anything that users
-can regenerate automatically from other parts of the Corresponding
-Source.
-
-  The Corresponding Source for a work in source code form is that
-same work.
-
-  2. Basic Permissions.
-
-  All rights granted under this License are granted for the term of
-copyright on the Program, and are irrevocable provided the stated
-conditions are met.  This License explicitly affirms your unlimited
-permission to run the unmodified Program.  The output from running a
-covered work is covered by this License only if the output, given its
-content, constitutes a covered work.  This License acknowledges your
-rights of fair use or other equivalent, as provided by copyright law.
-
-  You may make, run and propagate covered works that you do not
-convey, without conditions so long as your license otherwise remains
-in force.  You may convey covered works to others for the sole purpose
-of having them make modifications exclusively for you, or provide you
-with facilities for running those works, provided that you comply with
-the terms of this License in conveying all material for which you do
-not control copyright.  Those thus making or running the covered works
-for you must do so exclusively on your behalf, under your direction
-and control, on terms that prohibit them from making any copies of
-your copyrighted material outside their relationship with you.
-
-  Conveying under any other circumstances is permitted solely under
-the conditions stated below.  Sublicensing is not allowed; section 10
-makes it unnecessary.
-
-  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
-
-  No covered work shall be deemed part of an effective technological
-measure under any applicable law fulfilling obligations under article
-11 of the WIPO copyright treaty adopted on 20 December 1996, or
-similar laws prohibiting or restricting circumvention of such
-measures.
-
-  When you convey a covered work, you waive any legal power to forbid
-circumvention of technological measures to the extent such circumvention
-is effected by exercising rights under this License with respect to
-the covered work, and you disclaim any intention to limit operation or
-modification of the work as a means of enforcing, against the work's
-users, your or third parties' legal rights to forbid circumvention of
-technological measures.
-
-  4. Conveying Verbatim Copies.
-
-  You may convey verbatim copies of the Program's source code as you
-receive it, in any medium, provided that you conspicuously and
-appropriately publish on each copy an appropriate copyright notice;
-keep intact all notices stating that this License and any
-non-permissive terms added in accord with section 7 apply to the code;
-keep intact all notices of the absence of any warranty; and give all
-recipients a copy of this License along with the Program.
-
-  You may charge any price or no price for each copy that you convey,
-and you may offer support or warranty protection for a fee.
-
-  5. Conveying Modified Source Versions.
-
-  You may convey a work based on the Program, or the modifications to
-produce it from the Program, in the form of source code under the
-terms of section 4, provided that you also meet all of these conditions:
-
-    a) The work must carry prominent notices stating that you modified
-    it, and giving a relevant date.
-
-    b) The work must carry prominent notices stating that it is
-    released under this License and any conditions added under section
-    7.  This requirement modifies the requirement in section 4 to
-    "keep intact all notices".
-
-    c) You must license the entire work, as a whole, under this
-    License to anyone who comes into possession of a copy.  This
-    License will therefore apply, along with any applicable section 7
-    additional terms, to the whole of the work, and all its parts,
-    regardless of how they are packaged.  This License gives no
-    permission to license the work in any other way, but it does not
-    invalidate such permission if you have separately received it.
-
-    d) If the work has interactive user interfaces, each must display
-    Appropriate Legal Notices; however, if the Program has interactive
-    interfaces that do not display Appropriate Legal Notices, your
-    work need not make them do so.
-
-  A compilation of a covered work with other separate and independent
-works, which are not by their nature extensions of the covered work,
-and which are not combined with it such as to form a larger program,
-in or on a volume of a storage or distribution medium, is called an
-"aggregate" if the compilation and its resulting copyright are not
-used to limit the access or legal rights of the compilation's users
-beyond what the individual works permit.  Inclusion of a covered work
-in an aggregate does not cause this License to apply to the other
-parts of the aggregate.
-
-  6. Conveying Non-Source Forms.
-
-  You may convey a covered work in object code form under the terms
-of sections 4 and 5, provided that you also convey the
-machine-readable Corresponding Source under the terms of this License,
-in one of these ways:
-
-    a) Convey the object code in, or embodied in, a physical product
-    (including a physical distribution medium), accompanied by the
-    Corresponding Source fixed on a durable physical medium
-    customarily used for software interchange.
-
-    b) Convey the object code in, or embodied in, a physical product
-    (including a physical distribution medium), accompanied by a
-    written offer, valid for at least three years and valid for as
-    long as you offer spare parts or customer support for that product
-    model, to give anyone who possesses the object code either (1) a
-    copy of the Corresponding Source for all the software in the
-    product that is covered by this License, on a durable physical
-    medium customarily used for software interchange, for a price no
-    more than your reasonable cost of physically performing this
-    conveying of source, or (2) access to copy the
-    Corresponding Source from a network server at no charge.
-
-    c) Convey individual copies of the object code with a copy of the
-    written offer to provide the Corresponding Source.  This
-    alternative is allowed only occasionally and noncommercially, and
-    only if you received the object code with such an offer, in accord
-    with subsection 6b.
-
-    d) Convey the object code by offering access from a designated
-    place (gratis or for a charge), and offer equivalent access to the
-    Corresponding Source in the same way through the same place at no
-    further charge.  You need not require recipients to copy the
-    Corresponding Source along with the object code.  If the place to
-    copy the object code is a network server, the Corresponding Source
-    may be on a different server (operated by you or a third party)
-    that supports equivalent copying facilities, provided you maintain
-    clear directions next to the object code saying where to find the
-    Corresponding Source.  Regardless of what server hosts the
-    Corresponding Source, you remain obligated to ensure that it is
-    available for as long as needed to satisfy these requirements.
-
-    e) Convey the object code using peer-to-peer transmission, provided
-    you inform other peers where the object code and Corresponding
-    Source of the work are being offered to the general public at no
-    charge under subsection 6d.
-
-  A separable portion of the object code, whose source code is excluded
-from the Corresponding Source as a System Library, need not be
-included in conveying the object code work.
-
-  A "User Product" is either (1) a "consumer product", which means any
-tangible personal property which is normally used for personal, family,
-or household purposes, or (2) anything designed or sold for incorporation
-into a dwelling.  In determining whether a product is a consumer product,
-doubtful cases shall be resolved in favor of coverage.  For a particular
-product received by a particular user, "normally used" refers to a
-typical or common use of that class of product, regardless of the status
-of the particular user or of the way in which the particular user
-actually uses, or expects or is expected to use, the product.  A product
-is a consumer product regardless of whether the product has substantial
-commercial, industrial or non-consumer uses, unless such uses represent
-the only significant mode of use of the product.
-
-  "Installation Information" for a User Product means any methods,
-procedures, authorization keys, or other information required to install
-and execute modified versions of a covered work in that User Product from
-a modified version of its Corresponding Source.  The information must
-suffice to ensure that the continued functioning of the modified object
-code is in no case prevented or interfered with solely because
-modification has been made.
-
-  If you convey an object code work under this section in, or with, or
-specifically for use in, a User Product, and the conveying occurs as
-part of a transaction in which the right of possession and use of the
-User Product is transferred to the recipient in perpetuity or for a
-fixed term (regardless of how the transaction is characterized), the
-Corresponding Source conveyed under this section must be accompanied
-by the Installation Information.  But this requirement does not apply
-if neither you nor any third party retains the ability to install
-modified object code on the User Product (for example, the work has
-been installed in ROM).
-
-  The requirement to provide Installation Information does not include a
-requirement to continue to provide support service, warranty, or updates
-for a work that has been modified or installed by the recipient, or for
-the User Product in which it has been modified or installed.  Access to a
-network may be denied when the modification itself materially and
-adversely affects the operation of the network or violates the rules and
-protocols for communication across the network.
-
-  Corresponding Source conveyed, and Installation Information provided,
-in accord with this section must be in a format that is publicly
-documented (and with an implementation available to the public in
-source code form), and must require no special password or key for
-unpacking, reading or copying.
-
-  7. Additional Terms.
-
-  "Additional permissions" are terms that supplement the terms of this
-License by making exceptions from one or more of its conditions.
-Additional permissions that are applicable to the entire Program shall
-be treated as though they were included in this License, to the extent
-that they are valid under applicable law.  If additional permissions
-apply only to part of the Program, that part may be used separately
-under those permissions, but the entire Program remains governed by
-this License without regard to the additional permissions.
-
-  When you convey a copy of a covered work, you may at your option
-remove any additional permissions from that copy, or from any part of
-it.  (Additional permissions may be written to require their own
-removal in certain cases when you modify the work.)  You may place
-additional permissions on material, added by you to a covered work,
-for which you have or can give appropriate copyright permission.
-
-  Notwithstanding any other provision of this License, for material you
-add to a covered work, you may (if authorized by the copyright holders of
-that material) supplement the terms of this License with terms:
-
-    a) Disclaiming warranty or limiting liability differently from the
-    terms of sections 15 and 16 of this License; or
-
-    b) Requiring preservation of specified reasonable legal notices or
-    author attributions in that material or in the Appropriate Legal
-    Notices displayed by works containing it; or
-
-    c) Prohibiting misrepresentation of the origin of that material, or
-    requiring that modified versions of such material be marked in
-    reasonable ways as different from the original version; or
-
-    d) Limiting the use for publicity purposes of names of licensors or
-    authors of the material; or
-
-    e) Declining to grant rights under trademark law for use of some
-    trade names, trademarks, or service marks; or
-
-    f) Requiring indemnification of licensors and authors of that
-    material by anyone who conveys the material (or modified versions of
-    it) with contractual assumptions of liability to the recipient, for
-    any liability that these contractual assumptions directly impose on
-    those licensors and authors.
-
-  All other non-permissive additional terms are considered "further
-restrictions" within the meaning of section 10.  If the Program as you
-received it, or any part of it, contains a notice stating that it is
-governed by this License along with a term that is a further
-restriction, you may remove that term.  If a license document contains
-a further restriction but permits relicensing or conveying under this
-License, you may add to a covered work material governed by the terms
-of that license document, provided that the further restriction does
-not survive such relicensing or conveying.
-
-  If you add terms to a covered work in accord with this section, you
-must place, in the relevant source files, a statement of the
-additional terms that apply to those files, or a notice indicating
-where to find the applicable terms.
-
-  Additional terms, permissive or non-permissive, may be stated in the
-form of a separately written license, or stated as exceptions;
-the above requirements apply either way.
-
-  8. Termination.
-
-  You may not propagate or modify a covered work except as expressly
-provided under this License.  Any attempt otherwise to propagate or
-modify it is void, and will automatically terminate your rights under
-this License (including any patent licenses granted under the third
-paragraph of section 11).
-
-  However, if you cease all violation of this License, then your
-license from a particular copyright holder is reinstated (a)
-provisionally, unless and until the copyright holder explicitly and
-finally terminates your license, and (b) permanently, if the copyright
-holder fails to notify you of the violation by some reasonable means
-prior to 60 days after the cessation.
-
-  Moreover, your license from a particular copyright holder is
-reinstated permanently if the copyright holder notifies you of the
-violation by some reasonable means, this is the first time you have
-received notice of violation of this License (for any work) from that
-copyright holder, and you cure the violation prior to 30 days after
-your receipt of the notice.
-
-  Termination of your rights under this section does not terminate the
-licenses of parties who have received copies or rights from you under
-this License.  If your rights have been terminated and not permanently
-reinstated, you do not qualify to receive new licenses for the same
-material under section 10.
-
-  9. Acceptance Not Required for Having Copies.
-
-  You are not required to accept this License in order to receive or
-run a copy of the Program.  Ancillary propagation of a covered work
-occurring solely as a consequence of using peer-to-peer transmission
-to receive a copy likewise does not require acceptance.  However,
-nothing other than this License grants you permission to propagate or
-modify any covered work.  These actions infringe copyright if you do
-not accept this License.  Therefore, by modifying or propagating a
-covered work, you indicate your acceptance of this License to do so.
-
-  10. Automatic Licensing of Downstream Recipients.
-
-  Each time you convey a covered work, the recipient automatically
-receives a license from the original licensors, to run, modify and
-propagate that work, subject to this License.  You are not responsible
-for enforcing compliance by third parties with this License.
-
-  An "entity transaction" is a transaction transferring control of an
-organization, or substantially all assets of one, or subdividing an
-organization, or merging organizations.  If propagation of a covered
-work results from an entity transaction, each party to that
-transaction who receives a copy of the work also receives whatever
-licenses to the work the party's predecessor in interest had or could
-give under the previous paragraph, plus a right to possession of the
-Corresponding Source of the work from the predecessor in interest, if
-the predecessor has it or can get it with reasonable efforts.
-
-  You may not impose any further restrictions on the exercise of the
-rights granted or affirmed under this License.  For example, you may
-not impose a license fee, royalty, or other charge for exercise of
-rights granted under this License, and you may not initiate litigation
-(including a cross-claim or counterclaim in a lawsuit) alleging that
-any patent claim is infringed by making, using, selling, offering for
-sale, or importing the Program or any portion of it.
-
-  11. Patents.
-
-  A "contributor" is a copyright holder who authorizes use under this
-License of the Program or a work on which the Program is based.  The
-work thus licensed is called the contributor's "contributor version".
-
-  A contributor's "essential patent claims" are all patent claims
-owned or controlled by the contributor, whether already acquired or
-hereafter acquired, that would be infringed by some manner, permitted
-by this License, of making, using, or selling its contributor version,
-but do not include claims that would be infringed only as a
-consequence of further modification of the contributor version.  For
-purposes of this definition, "control" includes the right to grant
-patent sublicenses in a manner consistent with the requirements of
-this License.
-
-  Each contributor grants you a non-exclusive, worldwide, royalty-free
-patent license under the contributor's essential patent claims, to
-make, use, sell, offer for sale, import and otherwise run, modify and
-propagate the contents of its contributor version.
-
-  In the following three paragraphs, a "patent license" is any express
-agreement or commitment, however denominated, not to enforce a patent
-(such as an express permission to practice a patent or covenant not to
-sue for patent infringement).  To "grant" such a patent license to a
-party means to make such an agreement or commitment not to enforce a
-patent against the party.
-
-  If you convey a covered work, knowingly relying on a patent license,
-and the Corresponding Source of the work is not available for anyone
-to copy, free of charge and under the terms of this License, through a
-publicly available network server or other readily accessible means,
-then you must either (1) cause the Corresponding Source to be so
-available, or (2) arrange to deprive yourself of the benefit of the
-patent license for this particular work, or (3) arrange, in a manner
-consistent with the requirements of this License, to extend the patent
-license to downstream recipients.  "Knowingly relying" means you have
-actual knowledge that, but for the patent license, your conveying the
-covered work in a country, or your recipient's use of the covered work
-in a country, would infringe one or more identifiable patents in that
-country that you have reason to believe are valid.
-
-  If, pursuant to or in connection with a single transaction or
-arrangement, you convey, or propagate by procuring conveyance of, a
-covered work, and grant a patent license to some of the parties
-receiving the covered work authorizing them to use, propagate, modify
-or convey a specific copy of the covered work, then the patent license
-you grant is automatically extended to all recipients of the covered
-work and works based on it.
-
-  A patent license is "discriminatory" if it does not include within
-the scope of its coverage, prohibits the exercise of, or is
-conditioned on the non-exercise of one or more of the rights that are
-specifically granted under this License.  You may not convey a covered
-work if you are a party to an arrangement with a third party that is
-in the business of distributing software, under which you make payment
-to the third party based on the extent of your activity of conveying
-the work, and under which the third party grants, to any of the
-parties who would receive the covered work from you, a discriminatory
-patent license (a) in connection with copies of the covered work
-conveyed by you (or copies made from those copies), or (b) primarily
-for and in connection with specific products or compilations that
-contain the covered work, unless you entered into that arrangement,
-or that patent license was granted, prior to 28 March 2007.
-
-  Nothing in this License shall be construed as excluding or limiting
-any implied license or other defenses to infringement that may
-otherwise be available to you under applicable patent law.
-
-  12. No Surrender of Others' Freedom.
-
-  If conditions are imposed on you (whether by court order, agreement or
-otherwise) that contradict the conditions of this License, they do not
-excuse you from the conditions of this License.  If you cannot convey a
-covered work so as to satisfy simultaneously your obligations under this
-License and any other pertinent obligations, then as a consequence you may
-not convey it at all.  For example, if you agree to terms that obligate you
-to collect a royalty for further conveying from those to whom you convey
-the Program, the only way you could satisfy both those terms and this
-License would be to refrain entirely from conveying the Program.
-
-  13. Use with the GNU Affero General Public License.
-
-  Notwithstanding any other provision of this License, you have
-permission to link or combine any covered work with a work licensed
-under version 3 of the GNU Affero General Public License into a single
-combined work, and to convey the resulting work.  The terms of this
-License will continue to apply to the part which is the covered work,
-but the special requirements of the GNU Affero General Public License,
-section 13, concerning interaction through a network will apply to the
-combination as such.
-
-  14. Revised Versions of this License.
-
-  The Free Software Foundation may publish revised and/or new versions of
-the GNU General Public License from time to time.  Such new versions will
-be similar in spirit to the present version, but may differ in detail to
-address new problems or concerns.
-
-  Each version is given a distinguishing version number.  If the
-Program specifies that a certain numbered version of the GNU General
-Public License "or any later version" applies to it, you have the
-option of following the terms and conditions either of that numbered
-version or of any later version published by the Free Software
-Foundation.  If the Program does not specify a version number of the
-GNU General Public License, you may choose any version ever published
-by the Free Software Foundation.
-
-  If the Program specifies that a proxy can decide which future
-versions of the GNU General Public License can be used, that proxy's
-public statement of acceptance of a version permanently authorizes you
-to choose that version for the Program.
-
-  Later license versions may give you additional or different
-permissions.  However, no additional obligations are imposed on any
-author or copyright holder as a result of your choosing to follow a
-later version.
-
-  15. Disclaimer of Warranty.
-
-  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
-APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
-HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
-OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
-THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
-PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
-IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
-ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
-
-  16. Limitation of Liability.
-
-  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
-WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
-THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
-GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
-USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
-DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
-PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
-EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
-SUCH DAMAGES.
-
-  17. Interpretation of Sections 15 and 16.
-
-  If the disclaimer of warranty and limitation of liability provided
-above cannot be given local legal effect according to their terms,
-reviewing courts shall apply local law that most closely approximates
-an absolute waiver of all civil liability in connection with the
-Program, unless a warranty or assumption of liability accompanies a
-copy of the Program in return for a fee.
-
-                     END OF TERMS AND CONDITIONS
-
-            How to Apply These Terms to Your New Programs
-
-  If you develop a new program, and you want it to be of the greatest
-possible use to the public, the best way to achieve this is to make it
-free software which everyone can redistribute and change under these terms.
-
-  To do so, attach the following notices to the program.  It is safest
-to attach them to the start of each source file to most effectively
-state the exclusion of warranty; and each file should have at least
-the "copyright" line and a pointer to where the full notice is found.
-
-    <one line to give the program's name and a brief idea of what it does.>
-    Copyright (C) <year>  <name of author>
-
-    This program is free software: you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation, either version 3 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License
-    along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-Also add information on how to contact you by electronic and paper mail.
-
-  If the program does terminal interaction, make it output a short
-notice like this when it starts in an interactive mode:
-
-    <program>  Copyright (C) <year>  <name of author>
-    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
-    This is free software, and you are welcome to redistribute it
-    under certain conditions; type `show c' for details.
-
-The hypothetical commands `show w' and `show c' should show the appropriate
-parts of the General Public License.  Of course, your program's commands
-might be different; for a GUI interface, you would use an "about box".
-
-  You should also get your employer (if you work as a programmer) or school,
-if any, to sign a "copyright disclaimer" for the program, if necessary.
-For more information on this, and how to apply and follow the GNU GPL, see
-<https://www.gnu.org/licenses/>.
-
-  The GNU General Public License does not permit incorporating your program
-into proprietary programs.  If your program is a subroutine library, you
-may consider it more useful to permit linking proprietary applications with
-the library.  If this is what you want to do, use the GNU Lesser General
-Public License instead of this License.  But first, please read
-<https://www.gnu.org/licenses/why-not-lgpl.html>.
+MIT License
+
+Copyright (c) 2023-2026 Adrian Szczepański
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -2,14 +2,14 @@
 
 ## Overview
 
-Welcome! I've developed a set of custom nodes for ComfyUI that allows you to use Core ML models in your ComfyUI
-workflows.
+Welcome! In this repository you'll find a set of custom nodes for [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
+that allows you to use Core ML models in your ComfyUI workflows.
 These models are designed to leverage the Apple Neural Engine (ANE) on Apple Silicon (M1/M2) machines,
 thereby enhancing your workflows and improving performance.
 
 If you're not sure how to obtain these models, you can download them
-[here](https://huggingface.co/coreml-community) or convert your own models using
-[coremltools](https://github.com/apple/ml-stable-diffusion).
+[here](https://huggingface.co/coreml-community) or convert your own checkpoints
+directly with the conversion nodes in this suite (see [How to use](#how-to-use)).
 
 In simple terms, think of Core ML models as a tool that can help your ComfyUI work faster and more efficiently.
 For instance, during my tests on an M2 Pro 32GB machine,
@@ -48,6 +48,8 @@ That's it! You're now ready to start enhancing your ComfyUI workflows with Core
 - **VAE**: Variational Autoencoder. A model that learns a latent representation of images. It's used as a prior in
   Stable Diffusion.
 - **Checkpoint**: A file that contains the weights of a model. It's used to load models in Stable Diffusion.
+- **LCM**: [Latent Consistency Model](https://latent-consistency-models.github.io/). A type of model designed to
+  generate images with as few steps as possible.
 
 > [!NOTE]
 > Note on Compute Units:
@@ -64,6 +66,12 @@ These custom nodes come with a host of features, including:
 - Support for ANE (Apple Neural Engine)
 - Support for CPU and GPU
 - Support for `mlmodelc` and `mlpackage` files
+- Support for SDXL models
+- Support for LCM models
+- Support for LoRAs
+- SD1.5 -> Core ML conversion
+- SDXL -> Core ML conversion
+- LCM -> Core ML conversion
 
 > [!NOTE]
 > Please note that using Core ML models can take a bit longer to load initially.
@@ -73,9 +81,43 @@ These custom nodes come with a host of features, including:
 > [!NOTE]  
 > This repository will continue to be updated with more nodes and features over time.
 
+## Conversion & Acknowledgements
+
+The Core ML conversion pipeline in this repository began as an adaptation of
+Apple's [ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion),
+which pioneered running Stable Diffusion on the Apple Neural Engine. The
+implementation has since diverged and no longer depends on that package:
+
+- UNet conversion runs natively on `diffusers`' `UNet2DConditionModel`.
+- The ANE-friendly attention path (`SPLIT_EINSUM`, `SPLIT_EINSUM_V2`) is
+  reimplemented as standalone `diffusers` attention processors.
+- The toolchain tracks current ComfyUI (NumPy 2, Torch 2.7, coremltools 9,
+  Python 3.12).
+
+The goal is to keep iterating on these methods independently and to explore
+support beyond SD1.5.
+
+> [!IMPORTANT]
+> **Breaking change in 2.0.0.** The converted Core ML UNet now takes
+> `encoder_hidden_states` in the native `diffusers` layout
+> `(batch, tokens, hidden)` instead of the previous
+> `(batch, hidden, 1, tokens)`. Core ML models converted with earlier versions
+> are not compatible with 2.0.0 and must be re-converted.
+
 ## Installation
 
-The installation process is simple!
+### Using ComfyUI-Manager
+
+The easiest way to install the custom nodes is to use the ComfyUI-Manager. You can find the installation instructions
+[here](https://github.com/ltdrdata/ComfyUI-Manager#installation). Once you've installed the ComfyUI-Manager, you can
+install the custom nodes by following these steps:
+
+- Open the ComfyUI-Manager by clicking the `Manager` button in the ComfyUI toolbar.
+- Click the `Install Custom Nodes` button.
+- Search for `Core ML` and click the `Install` button.
+- Restart ComfyUI.
+
+### Manual Installation
 
 1. Clone this repository into the custom_nodes directory of your ComfyUI. If you're not sure how to do this, you can
    download the repository as a zip file and extract it into the same directory.
@@ -121,10 +163,6 @@ node is a `coreml_model` object that can be used with the Core ML Sampler.
 - **Outputs**:
     - **coreml_model**: A Core ML model that can be used with the Core ML Sampler.
 
-> [!NOTE]  
-> Some models are designed to support ControlNet. If you're using such a model,
-> make sure to provide a ControlNet input; otherwise, the model will use random noise as ControlNet input.
-
 #### Core ML Sampler (`CoreMLSampler`)
 
 ![CoreMLSampler](./assets/sampler.png?raw=true)
@@ -143,6 +181,118 @@ resulting latent as you normally would in your workflow.
     - **LATENT**: The latent image output by the Core ML model. This can be decoded using a VAE Decoder or used as input
       to the next node in your workflow.
 
+#### Checkpoint Converter
+
+![CoreMLConverter](./assets/checkpoint_converter.png?raw=true)
+
+You can use this node to convert any **SD1.5** based checkpoint to a Core ML model. The converted model is stored in the
+`models/unet` directory and can be used with the `Core ML UNet Loader`. The conversion parameters are encoded in
+the node name, so if the model already exists, the node will not convert it again.
+
+- **Inputs**:
+    - **ckpt_name**: The name of the checkpoint to convert. This should be the name of the checkpoint file stored in the
+      `models/checkpoints` directory.
+    - **model_version**: Whether the model is based on SD1.5 or SDXL.
+    - **height**: The desired height of the image generated by the model. The default is 512. Any positive multiple of 8 is accepted.
+    - **width**: The desired width of the image generated by the model. The default is 512. Any positive multiple of 8 is accepted.
+    - **batch_size**: The batch size of generated images. If you're planning to generate batches of images, you can try
+      increasing this value to speed up the generation process. The default is 1.
+    - **attention_implementation**: The attention implementation used when converting the model. Choose SPLIT_EINSUM or
+      SPLIT_EINSUM_V2 for better ANE support. Choose ORIGINAL for better GPU support.
+    - **compute_unit**: The hardware on which the model should run. This is used only when loading the model and doesn't
+      affect the conversion process.
+    - **controlnet_support**: For the model to support ControlNet, it must be converted with this option set to True.
+      The
+      default is False.
+    - **lora_params** [optional]: Optional LoRA names and weights. If provided, the model will be converted with LoRA(s)
+      baked in. More on loading LoRAs below.
+- **Outputs**:
+    - **coreml_model**: The converted Core ML model that can be used with Core ML Sampler.
+
+> [!NOTE]
+> Some models use a custom config .yaml file. If you're using such a model, you'll need to place the config file in the
+> `models/configs` directory. The config file should be named the same as the checkpoint file. For example, if the
+> checkpoint file is named `juggernaut_aftermath.safetensors`, the config file should be
+> named `juggernaut_aftermath.yaml`.
+> The config file will be automatically loaded during conversion.
+
+> [!NOTE]
+> For now, the converter relies heavilty on the model name to determine the conversion parameters. This means that if
+> you change the model name, the node will convert the model again. Other than that, if you find the name too long or
+> confusing, you can change it to anything you want.
+
+#### LoRA Loader
+
+![LoRALoader](./assets/lora_loader.png?raw=true)
+
+This node allows you to load LoRAs and bake them into a model. Since this is a workaround (as model weights can't be
+modified
+after conversion), there are a few caveats to keep in mind:
+
+- The LoRA weights and _strength_model_ parameter are baked into the model. This means that you can't change them
+  after conversion. This also means that you need to convert the model again if you want to change the LoRA weights.
+- Loading LoRA affects CLIP, which is not a part of Core ML workflow, so you'll need to load CLIP separately,
+  either using `CLIPLoader` or `CheckpointLoaderSimple`. (See [example workflows](#example-workflows) for more details.)
+- After conversion, if you want to load the model using `CoreMLUnetLoader`, you'll need to apply the same LoRAs to
+  CLIP manually. (See [example workflows](#example-workflows) for more details.)
+- The LoRA names are encoded in the model name. This means that if you change the name of the LoRA file,
+  you'll need to change the model name as well, or the node will convert the model again. (Model strength is not
+  encoded, so if you want to change it, you'll need to delete the converted model manually)
+- _strength_clip_ parameter only affects the CLIP model and is not baked into the converted model. This means that
+  you can change it after conversion.
+
+- **Inputs**:
+    - **lora_name**: The name of the LoRA to load.
+    - **strength_model**: The strength of the LoRA model.
+    - **strength_clip**: The strength of the LoRA CLIP.
+    - **lora_params** [optional]: Optional output from other LoRA Loaders.
+    - **clip**: The CLIP model to use with the LoRA. This can be either output of the
+      `CLIPLoader`/`CheckpointLoaderSimple` or other LoRA Loaders.
+- **Outputs**:
+    - **lora_params**: The LoRA parameters that can be passed to the Core ML Converter or other LoRA Loaders.
+    - **CLIP**: The CLIP model with LoRA applied.
+
+#### LCM Converter
+
+![LCMConverter](./assets/lcm_converter.png?raw=true)
+
+This node converts [SimianLuo/LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7) model to Core
+ML. The converted model is stored in the `models/unet` directory and can be used with the Core ML UNet Loader. The
+conversion parameteres are encoded in the node name, so if the model already exists, the node will not convert it again.
+
+- **Inputs**:
+    - **height**: The desired height of the image generated by the model. The default is 512. Must be a multiple of 8.
+    - **width**: The desired width of the image generated by the model. The default is 512. Must be a multiple of 8.
+    - **batch_size**: The batch size of generated images. If you're planning to generate batches of images, you can try
+      increasing this value to speed up the generation process. The default is 1.
+    - **compute_unit**: The hardware on which the model should run. This is used only when loading the model and
+      doesn't affect the conversion process.
+    - **controlnet_support**: For the model to support ControlNet, it must be converted with this option set to True.
+      The default is False.
+
+> [!NOTE]
+> The conversion process can take a while, so please be patient.
+
+> [!NOTE]
+> When using the LCM model with Core ML Sampler, please set _sampler_name_ to `lcm` and _scheduler_ to `sgm_uniform`.
+
+#### Core ML Adapter (Experimental) (`CoreMLModelAdapter`)
+
+![CoreMLModelAdapter](./assets/adapter.png?raw=true)
+
+This node allows you to use a Core ML as a standard ComfyUI model. This is an experimental node and may not work with
+all models and nodes. Please use with caution and pay attention to the expected inputs of the model.
+
+- **Input**:
+    - **coreml_model**: The Core ML model to use as a ComfyUI model.
+- **Output**:
+    - **MODEL**: The Core ML model wrapped in a ComfyUI model.
+
+> [!NOTE]
+> While this approach allows you to use Core ML models with many ComfyUI nodes (both standard and custom), the
+> expected inputs of the model will not be checked, which may cause errors. Please make sure to use a model compatible
+> with the expected parameters.
+
 ### Example Workflows
 
 > [!NOTE]
@@ -157,8 +307,8 @@ can use any CLIP or VAE model as long as it's compatible with Stable Diffusion v
 
 1. **Loading text encoder (CLIP) and VAE models separately**
     - This workflow uses CLIP and VAE models available
-      [here](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/text_encoder/model.safetensors) and
-      [here](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/vae/diffusion_pytorch_model.safetensors).
+      [here](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/text_encoder/model.safetensors) and
+      [here](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/vae/diffusion_pytorch_model.safetensors).
       Once downloaded, place the models in the`models/clip` and `models/vae` directories respectively.
     - The Core ML UNet model is available
       [here](https://huggingface.co/coreml-community/coreml-stable-diffusion-v1-5_cn/blob/main/split_einsum/stable-diffusion-_v1-5_split-einsum_cn.zip).
@@ -166,7 +316,7 @@ can use any CLIP or VAE model as long as it's compatible with Stable Diffusion v
       ![coreml-unet+clip+vae](./assets/unet+sampler+clip+vae.png?raw=true)
 2. **Loading text encoder (CLIP) and VAE models from checkpoint file**
     - This workflow loads the CLIP and VAE models from the checkpoint file available
-      [here](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.safetensors).
+      [here](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors).
       Once downloaded, place the model in the`models/checkpoints` directory.
     - The Core ML UNet model is available
       [here](https://huggingface.co/coreml-community/coreml-stable-diffusion-v1-5_cn/blob/main/split_einsum/stable-diffusion-_v1-5_split-einsum_cn.zip).
@@ -180,19 +330,117 @@ being loaded using the standard ComfyUI nodes. Please refer to
 the [basic txt2img workflow](#basic-txt2img-with-core-ml-unet-loader) for more details on how to load the CLIP and VAE
 models.
 The ControlNet model used in this workflow is available
-[here](https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_lineart.pth).
+[here](https://huggingface.co/lllyasviel/control_v11p_sd15_scribble/blob/main/diffusion_pytorch_model.fp16.safetensors).
 Once downloaded, place the model in the `models/controlnet` directory.
 ![coreml-unet+controlnet](./assets/unet+sampler+controlnet.png?raw=true)
 
+#### Checkpoint conversion
+
+This workflow uses the Checkpoint Converter to convert the checkpoint file. See
+[Checkpoint Converter](#checkpoint-converter) description for more details.
+
+![checkpoint-converter](./assets/basic_conversion.png?raw=true)
+
+#### Checkpoint conversion with LoRA
+
+This workflow uses the Checkpoint Converter to convert the checkpoint file with LoRA. See
+[LoRA Loader](#lora-loader) description to read more about the caveats of using LoRA.
+
+![checkpoint-converter+lora](./assets/conversion+lora.png?raw=true)
+
+#### LCM LoRA conversion
+
+Please note that you can use multiple LoRAs with the same model. To do this, you'll need to use multiple LoRA Loaders.
+> [!IMPORTANT]
+> In this example, the model is passed through the adapter and `ModelSamplingDiscrete` nodes to a standard ComfyUI's
+> KSampler (not Core ML Sampler). ModelSamplingDiscrete needs to be used to sample models with LCM LoRAs properly.
+
+![multiple-loras](./assets/conversion+lcm_lora.png?raw=true)
+
+#### Loader with LoRAs
+
+This workflow uses the Core ML UNet Loader to load a model with LoRAs. The CLIP must be loaded separately and passed
+through the same LoRA nodes as during conversion. See [LoRA Loader](#lora-loader) description to read more about the
+caveats of using LoRA. Since _lora_name_ and _strength_model_ are baked into the model, it is not necessary to pass
+them as inputs to the loader.
+> [!IMPORTANT]
+> In this example, the model is passed through the adapter and `ModelSamplingDiscrete` nodes to a standard ComfyUI's
+> KSampler (not Core ML Sampler). ModelSamplingDiscrete needs to be used to sample models with LCM LoRAs properly.
+
+![loader+lora](./assets/loader+lcm_lora.png?raw=true)
+
+#### LCM conversion with ControlNet
+
+This workflow uses LCM converter to
+convert [SimianLuo/LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7)
+model to Core ML. The converted model can then be used with or without ControlNet to generate images.
+![lcm+controlnet](./assets/lcm+controlnet.png?raw=true)
+
+#### SDXL Base + Refiner conversion
+
+This is a basic workflow for SDXL. You add LoRAs and ControlNets the same way as in the previous examples.
+You can also skip the refiner step.
+
+The models used in this workflow are available at the following links:
+
+- [Base model + text_encoder (clip) + text_encoder_2 (clip2)](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+- [Refiner model](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)
+- [VAE](https://huggingface.co/stabilityai/sdxl-vae)
+
+> [!IMPORTANT]
+> SDXL on ANE is not supported. If loading of the model gets stuck, please try using CPU_AND_GPU or CPU_ONLY.  
+> For best results, use ORIGINAL attention implementation.
+
+![sdxl](./assets/sdxl_conversion.png?raw=true)
+
+## Quantization (opt-in)
+
+The `Core ML Converter` and `Core ML LCM Converter` nodes accept an
+optional `quantize_nbits` dropdown that runs k-means weight palettization
+(`coremltools.optimize.coreml.palettize_weights`) on the UNet before save.
+
+Values: `none` (default — no quantization, identical to unquantized
+behavior and filenames), `8`, `6`, `4`. The number is appended to the
+.mlpackage stem as `_q<bits>` so quantized and unquantized variants
+coexist on disk and in cache.
+
+### SD1.5 1×512×512 SPLIT_EINSUM tradeoffs (M2 Pro, ANE)
+
+Measured with 20 UNet forward passes at a fixed seed for the PSNR
+comparison:
+
+| nbits | size (MB) | size vs none | fwd median (ms) | PSNR vs `none` (dB) |
+|---|---:|---:|---:|---:|
+| none | 1641 | 1.000 | 197.1 | — |
+| 8    |  822 | 0.501 | 186.6 | 53.5 |
+| 6    |  617 | 0.376 | 183.0 | 40.2 |
+| 4    |  412 | 0.251 | 179.8 | 27.5 |
+
+PSNR here is computed on the raw `noise_pred` output of a single UNet
+forward at a fixed seed, not on the final decoded image — it isolates
+the quantization-induced drift from sampler / VAE noise. Final-image
+PSNR is comfortably higher (the sampler averages over 20 steps).
+
+### Recommended settings per chip / RAM
+
+- **8 GB RAM (M1 base, M2 base):** `nbits=4`. ~4× smaller model, still
+  loads, PSNR 27 dB is visually identical at SD1.5 sizes.
+- **16 GB RAM (M1/M2/M3 Pro):** `nbits=6` is the sweet spot — ~2.7×
+  smaller, PSNR 40 dB, no perceptible quality drop.
+- **32 GB+ RAM (Max / Ultra):** `nbits=8` if you want the safety
+  margin, `none` if you want bit-identical output for golden testing.
+
+The default stays `none` so existing workflows produce byte-for-byte
+identical output.
+
 ## Limitations
 
 - Core ML models are fixed in terms of their inputs and outputs.
   This means you'll need to use latent images of the same size as the input of the model (512x512 is the default for
   SD1.5).
-  However, you can convert the model to a different input size using tools available
-  in the [apple/ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion) repository.
-- For now, only Stable Diffusion v1.5 is supported.
-- LoRA is not supported yet.
+  However, you can re-convert the model to a different input size using the
+  conversion nodes in this suite (set the desired width and height).
+- SD2.1 models are not supported.
 
 [^1]:
 Unless [EnumeratedShapes](https://apple.github.io/coremltools/docs-guides/source/flexible-inputs.html#select-from-predetermined-shapes)

__init__.py

@@ -3,13 +3,28 @@ import sys
 
 sys.path.append(os.path.dirname(__file__))
 
-from coreml_suite import CoreMLLoaderUNet, CoreMLSampler
+from coreml_suite.nodes import (
+    CoreMLLoaderUNet,
+    CoreMLSampler,
+    CoreMLSamplerAdvanced,
+    CoreMLModelAdapter,
+    CoreMLConverter,
+    COREML_LOAD_LORA,
+)
 
 NODE_CLASS_MAPPINGS = {
     "CoreMLUNetLoader": CoreMLLoaderUNet,
     "CoreMLSampler": CoreMLSampler,
+    "CoreMLSamplerAdvanced": CoreMLSamplerAdvanced,
+    "CoreMLModelAdapter": CoreMLModelAdapter,
+    "Core ML LoRA Loader": COREML_LOAD_LORA,
+    "Core ML Converter": CoreMLConverter,
 }
 NODE_DISPLAY_NAME_MAPPINGS = {
     "CoreMLUNetLoader": "Load Core ML UNet",
     "CoreMLSampler": "Core ML Sampler",
+    "CoreMLSamplerAdvanced": "Core ML Sampler (Advanced)",
+    "CoreMLModelAdapter": "Core ML Adapter (Experimental)",
+    "Core ML LoRA Loader": "Load LoRA to use with Core ML",
+    "Core ML Converter": "Convert Checkpoint to Core ML",
 }

conftest.py

@@ -0,0 +1,4 @@
+"""Top-level conftest: prevent pytest from importing the repo-root
+__init__.py (the ComfyUI custom-node entry point pulls in comfy + nodes,
+which breaks the Tier-0 'no-framework' promise)."""
+collect_ignore = ["__init__.py"]

constraints/comfy-ceiling.txt

@@ -0,0 +1,18 @@
+# Toolchain ceiling for installing a floating-latest ComfyUI's requirements.txt
+# in the Tier 2 nightly canary (.github/workflows/tier2.yml, latest mode).
+#
+# ComfyUI's requirements.txt requests bare `torch`/`torchvision`/`torchaudio`
+# and `numpy>=1.25.0`, which would float past the versions coremltools 9 /
+# apple-ml-stable-diffusion have been validated against.
+# These constraints cap the resolution so the canary keeps testing the same
+# toolchain the suite actually ships.
+#
+# If upstream ComfyUI ever hard-requires something beyond these bounds, the
+# install FAILS — and that failure is the signal we want: it means the host
+# outgrew the pinned toolchain and coremltools / ml-stable-diffusion need a
+# deliberate bump, not a silent float.
+torch>=2.7,<2.8
+torchvision>=0.22,<0.23
+torchaudio>=2.7,<2.8
+numpy>=1.25,<2
+coremltools>=9,<10

coreml_suite/__init__.py

@@ -1,4 +1,2 @@
-from coreml_suite.loaders import CoreMLLoaderUNet
-from coreml_suite.samplers import CoreMLSampler
-
-__all__ = ["CoreMLLoaderUNet", "CoreMLSampler"]
+class COREML_NODE:
+    CATEGORY = "Core ML Suite"

coreml_suite/config.py

@@ -0,0 +1,115 @@
+import torch
+
+from comfy import supported_models_base
+from comfy import latent_formats
+from comfy.model_detection import convert_config
+
+from coreml_diffusion import ModelVersion
+
+
+config_map = {
+    ModelVersion.SD15: {
+        "use_checkpoint": False,
+        "image_size": 32,
+        "out_channels": 4,
+        "use_spatial_transformer": True,
+        "legacy": False,
+        "adm_in_channels": None,
+        "dtype": torch.float16,
+        "in_channels": 4,
+        "model_channels": 320,
+        "num_res_blocks": 2,
+        "attention_resolutions": [1, 2, 4],
+        "transformer_depth": [1, 1, 1, 0],
+        "channel_mult": [1, 2, 4, 4],
+        "transformer_depth_middle": 1,
+        "use_linear_in_transformer": False,
+        "context_dim": 768,
+        "num_heads": 8,
+        "disable_unet_model_creation": True,
+    },
+    ModelVersion.SDXL: {
+        "use_checkpoint": False,
+        "image_size": 32,
+        "out_channels": 4,
+        "use_spatial_transformer": True,
+        "legacy": False,
+        "num_classes": "sequential",
+        "adm_in_channels": 2816,
+        "dtype": torch.float16,
+        "in_channels": 4,
+        "model_channels": 320,
+        "num_res_blocks": 2,
+        "attention_resolutions": [2, 4],
+        "transformer_depth": [0, 2, 10],
+        "channel_mult": [1, 2, 4],
+        "transformer_depth_middle": 10,
+        "use_linear_in_transformer": True,
+        "context_dim": 2048,
+        "num_head_channels": 64,
+        "disable_unet_model_creation": True,
+    },
+    ModelVersion.SDXL_REFINER: {
+        "use_checkpoint": False,
+        "image_size": 32,
+        "out_channels": 4,
+        "use_spatial_transformer": True,
+        "legacy": False,
+        "num_classes": "sequential",
+        "adm_in_channels": 2560,
+        "dtype": torch.float16,
+        "in_channels": 4,
+        "model_channels": 384,
+        "num_res_blocks": 2,
+        "attention_resolutions": [2, 4],
+        "transformer_depth": [0, 4, 4, 0],
+        "channel_mult": [1, 2, 4, 4],
+        "transformer_depth_middle": 4,
+        "use_linear_in_transformer": True,
+        "context_dim": 1280,
+        "num_head_channels": 64,
+        "disable_unet_model_creation": True,
+    },
+}
+
+latent_format_map = {
+    ModelVersion.SD15: latent_formats.SD15,
+    ModelVersion.SDXL: latent_formats.SDXL,
+    ModelVersion.SDXL_REFINER: latent_formats.SDXL,
+}
+
+
+def get_model_config(model_version: ModelVersion):
+    unet_config = convert_config(config_map[model_version])
+    config = supported_models_base.BASE(unet_config)
+    config.latent_format = latent_format_map[model_version]()
+    return config
+
+
+def unet_config_from_diffusers_unet(state_dict):
+    match = {}
+    attention_resolutions = []
+
+    attn_res = 1
+    for i in range(5):
+        k = "down_blocks.{}.attentions.1.transformer_blocks.0.attn2.to_k.weight".format(
+            i
+        )
+        if k in state_dict:
+            match["context_dim"] = state_dict[k].shape[1]
+            attention_resolutions.append(attn_res)
+        attn_res *= 2
+
+    match["attention_resolutions"] = attention_resolutions
+
+    match["model_channels"] = state_dict["conv_in.weight"].shape[0]
+    match["in_channels"] = state_dict["conv_in.weight"].shape[1]
+    match["adm_in_channels"] = None
+    if "class_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[
+            1
+        ]
+    elif "add_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1]
+
+    print(match)

coreml_suite/controlnet.py

@@ -0,0 +1,14 @@
+"""Compatibility shim — re-exports from coreml_suite.core.controlnet."""
+from coreml_suite.core.controlnet import (
+    chunk_control,
+    expand_inputs,
+    extract_residual_kwargs,
+    no_control,
+)
+
+__all__ = [
+    "chunk_control",
+    "expand_inputs",
+    "extract_residual_kwargs",
+    "no_control",
+]

coreml_suite/core/__init__.py

@@ -0,0 +1,10 @@
+"""Framework-free pure-logic core of ComfyUI-CoreMLSuite.
+
+Modules under this package must NOT import `comfy`, `coremltools`,
+`python_coreml_stable_diffusion`, `folder_paths`, `nodes`, or any other
+ComfyUI / Apple runtime. Only `numpy` and `torch` are allowed.
+
+The thin adapters in `coreml_suite.{latents,controlnet,models}` keep the
+old public import paths working so `coreml_suite/nodes.py` and downstream
+ComfyUI workflows are unchanged.
+"""

coreml_suite/core/controlnet.py

@@ -0,0 +1,67 @@
+"""Pure helpers around the ControlNet residual inputs of the Core ML UNet.
+
+Re-exported by coreml_suite.controlnet. Characterization tests cover
+shapes, dtype (fp16), and zero-fill fallback.
+"""
+from itertools import chain
+from math import ceil
+
+import numpy as np
+import torch
+
+from coreml_suite.core.latents import chunk_batch
+
+
+def expand_inputs(inputs):
+    expanded = inputs.copy()
+    for k, v in inputs.items():
+        if isinstance(v, np.ndarray):
+            expanded[k] = np.concatenate([v] * 2) if v.shape[0] == 1 else v
+        elif isinstance(v, torch.Tensor):
+            expanded[k] = torch.cat([v] * 2) if v.shape[0] == 1 else v
+        elif isinstance(v, list):
+            expanded[k] = v * 2 if len(v) == 1 else v
+        elif isinstance(v, dict):
+            expand_inputs(v)
+    return expanded
+
+
+def extract_residual_kwargs(expected_inputs, control):
+    if "additional_residual_0" not in expected_inputs.keys():
+        return {}
+    if control is None:
+        return no_control(expected_inputs)
+
+    residual_kwargs = {
+        "additional_residual_{}".format(i): r.cpu().numpy().astype(np.float16)
+        for i, r in enumerate(chain(control["output"], control["middle"]))
+    }
+    return residual_kwargs
+
+
+def no_control(expected_inputs):
+    shapes_dict = {
+        k: v["shape"] for k, v in expected_inputs.items() if k.startswith("additional")
+    }
+    residual_kwargs = {
+        k: torch.zeros(*shape).cpu().numpy().astype(dtype=np.float16)
+        for k, shape in shapes_dict.items()
+    }
+    return residual_kwargs
+
+
+def chunk_control(cn, target_size):
+    if cn is None:
+        return [None] * target_size
+
+    num_chunks = ceil(cn["output"][0].shape[0] / target_size)
+
+    out = [{"output": [], "middle": []} for _ in range(num_chunks)]
+
+    for k, v in cn.items():
+        for i, x in enumerate(v):
+            chunks = chunk_batch(x, (target_size, *x.shape[1:]))
+            for j, chunk in enumerate(chunks):
+                out[j][k].append(chunk)
+
+    return out

coreml_suite/core/inputs.py

@@ -0,0 +1,111 @@
+"""Pure transform from torch sampler inputs to Core ML UNet kwargs.
+
+Characterization tests cover SD1.5 / SDXL base / SDXL refiner / LCM
+variants and the chunked-batch fan-out.
+"""
+import numpy as np
+import torch
+
+from coreml_suite.core.controlnet import extract_residual_kwargs, chunk_control
+from coreml_suite.core.latents import chunk_batch
+
+
+class CoreMLInputs:
+    def __init__(self, x, t, context, control, **kwargs):
+        self.x = x
+        self.t = t
+        self.context = context
+        self.control = control
+        self.time_ids = kwargs.get("time_ids")
+        self.text_embeds = kwargs.get("text_embeds")
+        self.ts_cond = kwargs.get("timestep_cond")
+
+    def coreml_kwargs(self, expected_inputs):
+        sample = self.x.cpu().numpy().astype(np.float16)
+
+        context = self.context.cpu().numpy().astype(np.float16)
+
+        t = self.t.cpu().numpy().astype(np.float16)
+
+        model_input_kwargs = {
+            "sample": sample,
+            "encoder_hidden_states": context,
+            "timestep": t,
+        }
+        residual_kwargs = extract_residual_kwargs(expected_inputs, self.control)
+        model_input_kwargs |= residual_kwargs
+
+        # LCM
+        if self.ts_cond is not None:
+            model_input_kwargs["timestep_cond"] = (
+                self.ts_cond.cpu().numpy().astype(np.float16)
+            )
+
+        # SDXL
+        if "text_embeds" in expected_inputs:
+            model_input_kwargs["text_embeds"] = (
+                self.text_embeds.cpu().numpy().astype(np.float16)
+            )
+        if "time_ids" in expected_inputs:
+            model_input_kwargs["time_ids"] = (
+                self.time_ids.cpu().numpy().astype(np.float16)
+            )
+
+        return model_input_kwargs
+
+    def chunks(self, expected_inputs):
+        sample_shape = expected_inputs["sample"]["shape"]
+        timestep_shape = expected_inputs["timestep"]["shape"]
+        context_shape = expected_inputs["encoder_hidden_states"]["shape"]
+
+        chunked_x = chunk_batch(self.x, sample_shape)
+        ts = list(torch.full((len(chunked_x), timestep_shape[0]), self.t[0]))
+        chunked_context = chunk_batch(self.context, context_shape)
+
+        chunked_control = [None] * len(chunked_x)
+        if self.control is not None:
+            chunked_control = chunk_control(self.control, sample_shape[0])
+
+        chunked_ts_cond = [None] * len(chunked_x)
+        if self.ts_cond is not None:
+            ts_cond_shape = expected_inputs["timestep_cond"]["shape"]
+            chunked_ts_cond = chunk_batch(self.ts_cond, ts_cond_shape)
+
+        chunked_time_ids = [None] * len(chunked_x)
+        if expected_inputs.get("time_ids") is not None:
+            time_ids_shape = expected_inputs["time_ids"]["shape"]
+            if self.time_ids is None:
+                self.time_ids = torch.zeros(len(chunked_x), *time_ids_shape[1:]).to(
+                    self.x.device
+                )
+            chunked_time_ids = chunk_batch(self.time_ids, time_ids_shape)
+
+        chunked_text_embeds = [None] * len(chunked_x)
+        if expected_inputs.get("text_embeds") is not None:
+            text_embeds_shape = expected_inputs["text_embeds"]["shape"]
+            if self.text_embeds is None:
+                self.text_embeds = torch.zeros(
+                    len(chunked_x), *text_embeds_shape[1:]
+                ).to(self.x.device)
+            chunked_text_embeds = chunk_batch(self.text_embeds, text_embeds_shape)
+
+        return [
+            CoreMLInputs(
+                x,
+                t,
+                context,
+                control,
+                timestep_cond=ts_cond,
+                time_ids=time_ids,
+                text_embeds=text_embeds,
+            )
+            for x, t, context, control, ts_cond, time_ids, text_embeds in zip(
+                chunked_x,
+                ts,
+                chunked_context,
+                chunked_control,
+                chunked_ts_cond,
+                chunked_time_ids,
+                chunked_text_embeds,
+            )
+        ]

coreml_suite/core/latents.py

@@ -0,0 +1,42 @@
+"""Pure batch-chunking helpers for Core ML's fixed-shape UNet inputs.
+
+Re-exported by coreml_suite.latents. Characterization tests cover the
+contract (padding-zero regions, truncation in merge_chunks,
+identity-passthrough when shape already matches).
+"""
+import torch
+
+
+def chunk_batch(input_tensor, target_shape):
+    if input_tensor.shape == target_shape:
+        return [input_tensor]
+
+    batch_size = input_tensor.shape[0]
+    target_batch_size = target_shape[0]
+
+    num_chunks = batch_size // target_batch_size
+    if num_chunks == 0:
+        padding = torch.zeros(target_batch_size - batch_size, *target_shape[1:]).to(
+            input_tensor.device
+        )
+        return [torch.cat((input_tensor, padding), dim=0)]
+
+    mod = batch_size % target_batch_size
+    if mod != 0:
+        chunks = list(torch.chunk(input_tensor[:-mod], num_chunks))
+        padding = torch.zeros(target_batch_size - mod, *target_shape[1:]).to(
+            input_tensor.device
+        )
+        padded = torch.cat((input_tensor[-mod:], padding), dim=0)
+        chunks.append(padded)
+        return chunks
+
+    chunks = list(torch.chunk(input_tensor, num_chunks))
+    return chunks
+
+
+def merge_chunks(chunks, orig_shape):
+    merged = torch.cat(chunks, dim=0)
+    if merged.shape == orig_shape:
+        return merged
+    return merged[: orig_shape[0]]

coreml_suite/core/sdxl.py

@@ -0,0 +1,91 @@
+"""Pure SDXL detection + time_ids/text_embeds assembly.
+
+The framework-coupled adapter `add_sdxl_model_options` lives in models.py
+and delegates the math here. Characterization tests cover base (len 6) vs
+refiner (len 5) and the closure free-vars produced by
+`sdxl_model_function_wrapper`.
+"""
+import torch
+
+
+def is_sdxl(coreml_model):
+    return (
+        "time_ids" in coreml_model.expected_inputs
+        and "text_embeds" in coreml_model.expected_inputs
+    )
+
+
+def is_sdxl_base(coreml_model):
+    return (
+        is_sdxl(coreml_model)
+        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 6
+    )
+
+
+def is_sdxl_refiner(coreml_model):
+    return (
+        is_sdxl(coreml_model)
+        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 5
+    )
+
+
+def build_sdxl_time_ids(pos_dict, neg_dict, *, is_base: bool, is_refiner: bool):
+    """Compose the (2, N) time_ids tensor for the SDXL Core ML UNet.
+
+    - base: N=6  -> [h, w, crop_h, crop_w, target_h, target_w]
+    - refiner: N=5 -> [h, w, crop_h, crop_w, aesthetic_score]
+    - neither: N=4 -> [h, w, crop_h, crop_w] (edge case kept for parity)
+    """
+    pos_time_ids = [
+        pos_dict.get("height", 768),
+        pos_dict.get("width", 768),
+        pos_dict.get("crop_h", 0),
+        pos_dict.get("crop_w", 0),
+    ]
+    neg_time_ids = [
+        neg_dict.get("height", 768),
+        neg_dict.get("width", 768),
+        neg_dict.get("crop_h", 0),
+        neg_dict.get("crop_w", 0),
+    ]
+
+    if is_base:
+        pos_time_ids += [
+            pos_dict.get("target_height", 768),
+            pos_dict.get("target_width", 768),
+        ]
+        neg_time_ids += [
+            neg_dict.get("target_height", 768),
+            neg_dict.get("target_width", 768),
+        ]
+
+    if is_refiner:
+        pos_time_ids += [pos_dict.get("aesthetic_score", 6)]
+        neg_time_ids += [neg_dict.get("aesthetic_score", 2.5)]
+
+    return torch.tensor([pos_time_ids, neg_time_ids])
+
+
+def build_sdxl_text_embeds(pos_pooled, neg_pooled):
+    """Concat pos then neg along the batch dim. Locked contract."""
+    return torch.cat((pos_pooled, neg_pooled))
+
+
+def sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False):
+    def wrapper(model_function, params):
+        x = params["input"]
+        t = params["timestep"]
+        c = params["c"]
+
+        context = c.get("c_crossattn")
+
+        if context is None:
+            return torch.zeros_like(x)
+
+        if refiner and context is not None:
+            # converted refiner accepts only g clip
+            c["c_crossattn"] = context[:, :, 768:]
+
+        return model_function(x, t, **c, time_ids=time_ids, text_embeds=text_embeds)
+
+    return wrapper

coreml_suite/coreml_model.py

@@ -0,0 +1,42 @@
+import time
+
+import coremltools as ct
+
+from coreml_suite.logger import logger
+
+
+class CoreMLModel:
+    """Small runtime wrapper around coremltools.models.MLModel.
+
+    This keeps the inference path independent from apple/ml-stable-diffusion's
+    CoreMLModel wrapper while preserving the contract used by the sampler code:
+    ``expected_inputs`` and callable prediction.
+    """
+
+    def __init__(self, model_path, compute_unit):
+        self.model_path = model_path
+        self.compute_unit = self._compute_unit(compute_unit)
+
+        logger.info(f"Loading {model_path} to {self.compute_unit.name}")
+        start = time.time()
+        self.model = ct.models.MLModel(model_path, compute_units=self.compute_unit)
+        logger.info(f"Loading {model_path} took {time.time() - start:.1f} seconds")
+
+        self.expected_inputs = self._expected_inputs()
+
+    def __call__(self, **kwargs):
+        return self.model.predict(kwargs)
+
+    @staticmethod
+    def _compute_unit(compute_unit):
+        if isinstance(compute_unit, ct.ComputeUnit):
+            return compute_unit
+        return ct.ComputeUnit[compute_unit]
+
+    def _expected_inputs(self):
+        return {
+            feature.name: {
+                "shape": tuple(feature.type.multiArrayType.shape),
+            }
+            for feature in self.model.get_spec().description.input
+        }

coreml_suite/latents.py

@@ -0,0 +1,4 @@
+"""Compatibility shim — re-exports from coreml_suite.core.latents."""
+from coreml_suite.core.latents import chunk_batch, merge_chunks
+
+__all__ = ["chunk_batch", "merge_chunks"]

coreml_suite/lcm/__init__.py

@@ -0,0 +1,8 @@
+"""LCM runtime support (sampler-side).
+
+The dedicated LCM converter node was removed once the standard ``CoreMLConverter``
+gained model-version auto-detection (full-distill LCM is detected from the
+checkpoint). What remains here is runtime sampling support — ``utils`` patches the
+model sampling and supplies the guidance embedding when a converted UNet exposes
+``timestep_cond``.
+"""

coreml_suite/lcm/utils.py

@@ -0,0 +1,73 @@
+import torch
+
+from comfy.model_management import get_torch_device
+from comfy_extras.nodes_model_advanced import ModelSamplingDiscreteDistilled, LCM
+
+
+def is_lcm(coreml_model):
+    return "timestep_cond" in coreml_model.expected_inputs
+
+
+def get_w_embedding(w, embedding_dim=512, dtype=torch.float32):
+    assert len(w.shape) == 1
+    w = w * 1000.0
+
+    half_dim = embedding_dim // 2
+    emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+    emb = w.to(dtype)[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0, 1))
+    assert emb.shape == (w.shape[0], embedding_dim)
+    return emb
+
+
+def model_function_wrapper(w_embedding):
+    def wrapper(model_function, params):
+        x = params["input"]
+        t = params["timestep"]
+        c = params["c"]
+
+        context = c.get("c_crossattn")
+
+        if context is None:
+            return torch.zeros_like(x)
+
+        return model_function(x, t, **c, timestep_cond=w_embedding)
+
+    return wrapper
+
+
+def lcm_patch(model):
+    m = model.clone()
+    sampling_type = LCM
+    sampling_base = ModelSamplingDiscreteDistilled
+
+    class ModelSamplingAdvanced(sampling_base, sampling_type):
+        pass
+
+    model_sampling = ModelSamplingAdvanced()
+    m.add_object_patch("model_sampling", model_sampling)
+
+    return m
+
+
+def add_lcm_model_options(model_patcher, cfg, latent_image):
+    mp = model_patcher.clone()
+
+    latent = latent_image["samples"].to(get_torch_device())
+    batch_size = latent.shape[0]
+    dtype = latent.dtype
+    device = get_torch_device()
+
+    w = torch.tensor(cfg).repeat(batch_size)
+    w_embedding = get_w_embedding(w, embedding_dim=256).to(device=device, dtype=dtype)
+
+    model_options = {
+        "model_function_wrapper": model_function_wrapper(w_embedding),
+        "sampler_cfg_function": lambda x: x["cond"].to(device),
+    }
+    mp.model_options |= model_options
+
+    return mp

coreml_suite/loaders.py

@@ -1,84 +0,0 @@
-import os.path
-
-from coremltools import ComputeUnit
-from python_coreml_stable_diffusion.coreml_model import CoreMLModel
-
-import folder_paths
-
-from coreml_suite.logger import logger
-
-
-class CoreMLLoader:
-    PACKAGE_DIRNAME = ""
-
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "coreml_name": (list(s.coreml_filenames().keys()),),
-                "compute_unit": (
-                    [
-                        ComputeUnit.CPU_AND_NE.name,
-                        ComputeUnit.CPU_AND_GPU.name,
-                        ComputeUnit.ALL.name,
-                        ComputeUnit.CPU_ONLY.name,
-                    ],
-                ),
-            }
-        }
-
-    FUNCTION = "load"
-    CATEGORY = "Core ML Suite"
-
-    @classmethod
-    def coreml_filenames(cls):
-        extensions = (".mlmodelc", ".mlpackage")
-        all_paths = folder_paths.get_filename_list_(cls.PACKAGE_DIRNAME)[1]
-        coreml_paths = folder_paths.filter_files_extensions(all_paths, extensions)
-
-        return {os.path.split(p)[-1]: p for p in coreml_paths}
-
-    def load(self, coreml_name, compute_unit):
-        logger.info(f"Loading {coreml_name} to {compute_unit}")
-
-        coreml_path = self.coreml_filenames()[coreml_name]
-
-        sources = "compiled" if coreml_name.endswith(".mlmodelc") else "packages"
-
-        return self._load(coreml_path, compute_unit, sources)
-
-    def _load(self, coreml_path, compute_unit, sources):
-        return (CoreMLModel(coreml_path, compute_unit, sources),)
-
-
-class CoreMLLoaderCkpt(CoreMLLoader):
-    PACKAGE_DIRNAME = "checkpoints"
-    RETURN_TYPES = ("MODEL", "CLIP", "VAE")
-
-    def load(self, coreml_name, compute_unit):
-        # TODO: Implement this
-        pass
-
-
-class CoreMLLoaderTextEncoder(CoreMLLoader):
-    PACKAGE_DIRNAME = "clip"
-    RETURN_TYPES = ("CLIP",)
-
-    def load(self, coreml_name, compute_unit):
-        # TODO: Implement this
-        pass
-
-
-class CoreMLLoaderUNet(CoreMLLoader):
-    PACKAGE_DIRNAME = "unet"
-    RETURN_TYPES = ("COREML_UNET",)
-    RETURN_NAMES = ("coreml_model",)
-
-
-class CoreMLLoaderVAE(CoreMLLoader):
-    PACKAGE_DIRNAME = "vae"
-    RETURN_TYPES = ("VAE",)
-
-    def load(self, coreml_name, compute_unit):
-        # TODO: Implement this
-        pass

coreml_suite/models.py

@@ -1,63 +1,147 @@
-import numpy as np
+"""Framework-coupled glue between Core ML UNets and ComfyUI's sampler stack.
+
+Pure math (CoreMLInputs, SDXL detection, time_ids/text_embeds assembly,
+sdxl_model_function_wrapper) lives in coreml_suite.core.*.
+This module is what touches comfy.*: model_base, ModelPatcher, the
+diffusion_model wrapper, and the maintainer-facing add_sdxl_model_options
+adapter.
+"""
 import torch
 
+from comfy import model_base
+from comfy.model_management import get_torch_device
+from comfy.model_patcher import ModelPatcher
 
-from comfy import supported_models_base
-from comfy.latent_formats import SD15
-from comfy.model_base import BaseModel
+from coreml_suite.config import get_model_config, ModelVersion
+from coreml_suite.core.inputs import CoreMLInputs
+from coreml_suite.core.latents import merge_chunks
+from coreml_suite.core.sdxl import (
+    build_sdxl_text_embeds,
+    build_sdxl_time_ids,
+    is_sdxl,
+    is_sdxl_base,
+    is_sdxl_refiner,
+    sdxl_model_function_wrapper,
+)
+from coreml_suite.lcm.utils import is_lcm
+from coreml_suite.logger import logger
 
-from coreml_suite.utils import expand_inputs, extract_residual_kwargs
+__all__ = [
+    "CoreMLInputs",
+    "CoreMLModelWrapper",
+    "CoreMLModelWrapperLCM",
+    "add_sdxl_model_options",
+    "get_latent_image",
+    "get_model_patcher",
+    "is_sdxl",
+    "is_sdxl_base",
+    "is_sdxl_refiner",
+    "sdxl_model_function_wrapper",
+]
 
 
-def get_model_config():
-    # TODO: This is a dummy model config, but it should be enough to
-    #  get the model to load - implement a proper model config
-    model_config = supported_models_base.BASE({})
-    model_config.latent_format = SD15()
-    model_config.unet_config = {
-        "disable_unet_model_creation": True,
-        "num_res_blocks": 2,
-        "attention_resolutions": [1, 2, 4],
-        "channel_mult": [1, 2, 4, 4],
-        "transformer_depth": [1, 1, 1, 0],
+class CoreMLModelWrapper:
+    def __init__(self, coreml_model):
+        self.coreml_model = coreml_model
+        self.dtype = torch.float16
+
+    def __call__(self, x, t, context, control, transformer_options=None, **kwargs):
+        inputs = CoreMLInputs(x, t, context, control, **kwargs)
+        input_list = inputs.chunks(self.expected_inputs)
+
+        chunked_out = [
+            self.get_torch_outputs(
+                self.coreml_model(**input_kwargs.coreml_kwargs(self.expected_inputs)),
+                x.device,
+            )
+            for input_kwargs in input_list
+        ]
+        merged_out = merge_chunks(chunked_out, x.shape)
+
+        return merged_out
+
+    @staticmethod
+    def get_torch_outputs(model_output, device):
+        return torch.from_numpy(model_output["noise_pred"]).to(device)
+
+    @property
+    def expected_inputs(self):
+        return self.coreml_model.expected_inputs
+
+    @property
+    def is_lcm(self):
+        return is_lcm(self.coreml_model)
+
+    @property
+    def is_sdxl_base(self):
+        return is_sdxl_base(self.coreml_model)
+
+    @property
+    def is_sdxl_refiner(self):
+        return is_sdxl_refiner(self.coreml_model)
+
+    @property
+    def config(self):
+        if self.is_sdxl_base:
+            return get_model_config(ModelVersion.SDXL)
+
+        if self.is_sdxl_refiner:
+            return get_model_config(ModelVersion.SDXL_REFINER)
+
+        return get_model_config(ModelVersion.SD15)
+
+
+class CoreMLModelWrapperLCM(CoreMLModelWrapper):
+    def __init__(self, coreml_model):
+        super().__init__(coreml_model)
+        self.config = None
+
+
+def add_sdxl_model_options(model_patcher, positive, negative):
+    mp = model_patcher.clone()
+
+    pos_dict = positive[0][1]
+    neg_dict = negative[0][1]
+
+    is_base = model_patcher.model.diffusion_model.is_sdxl_base
+    is_refiner = model_patcher.model.diffusion_model.is_sdxl_refiner
+
+    time_ids = build_sdxl_time_ids(
+        pos_dict, neg_dict, is_base=is_base, is_refiner=is_refiner
+    )
+    text_embeds = build_sdxl_text_embeds(
+        pos_dict["pooled_output"], neg_dict["pooled_output"]
+    )
+
+    mp.model_options |= {
+        "model_function_wrapper": sdxl_model_function_wrapper(
+            time_ids, text_embeds, is_refiner
+        ),
     }
-    return model_config
+    return mp
 
 
-class CoreMLModelWrapper(BaseModel):
-    def __init__(self, model_config, coreml_model):
-        super().__init__(model_config)
-        self.diffusion_model = coreml_model
+def get_latent_image(coreml_model, latent_image):
+    if latent_image is not None:
+        return latent_image
 
-    def apply_model(
-        self,
-        x,
-        t,
-        c_concat=None,
-        c_crossattn=None,
-        c_adm=None,
-        control=None,
-        transformer_options={},
-    ):
-        sample = x.cpu().numpy().astype(np.float16)
+    logger.warning("No latent image provided, using empty tensor.")
+    expected = coreml_model.expected_inputs["sample"]["shape"]
+    batch_size = max(expected[0] // 2, 1)
+    latent_image = {"samples": torch.zeros(batch_size, *expected[1:])}
+    return latent_image
 
-        context = c_crossattn.cpu().numpy().astype(np.float16)
-        context = context.transpose(0, 2, 1)[:, :, None, :]
 
-        t = t.cpu().numpy().astype(np.float16)
+def get_model_patcher(coreml_model):
+    wrapped_model = CoreMLModelWrapper(coreml_model)
 
-        model_input_kwargs = {
-            "sample": sample,
-            "encoder_hidden_states": context,
-            "timestep": t,
-        }
-        residual_kwargs = extract_residual_kwargs(self.diffusion_model, control)
-        model_input_kwargs |= residual_kwargs
-        model_input_kwargs = expand_inputs(model_input_kwargs)
+    if wrapped_model.is_sdxl_base:
+        model = model_base.SDXL(wrapped_model.config, device=get_torch_device())
+    elif wrapped_model.is_sdxl_refiner:
+        model = model_base.SDXLRefiner(wrapped_model.config, device=get_torch_device())
+    else:
+        model = model_base.BaseModel(wrapped_model.config, device=get_torch_device())
 
-        np_out = self.diffusion_model(**model_input_kwargs)["noise_pred"]
-        return torch.from_numpy(np_out).to(x.device)
-
-    def get_dtype(self):
-        # Hardcoding torch-compatible dtype (used for memory allocation)
-        return torch.float16
+    model.diffusion_model = wrapped_model
+    model_patcher = ModelPatcher(model, get_torch_device(), None)
+    return model_patcher

coreml_suite/nodes.py

@@ -0,0 +1,390 @@
+import os
+
+from coremltools import ComputeUnit
+
+import folder_paths
+from coreml_suite import COREML_NODE
+from coreml_suite.coreml_model import CoreMLModel
+from coreml_suite.lcm.utils import add_lcm_model_options, lcm_patch, is_lcm
+from coreml_suite.logger import logger
+from nodes import KSampler, LoraLoader, KSamplerAdvanced
+
+from coreml_suite.models import (
+    add_sdxl_model_options,
+    is_sdxl,
+    get_model_patcher,
+    get_latent_image,
+)
+
+
+def _discover(fn_name, fallback):
+    """Populate a converter dropdown from coreml_diffusion's discovery API.
+
+    Fails soft: if the package is missing, too old to expose ``fn_name``, or
+    errors, the node still registers with the fallback list instead of vanishing
+    from the menu. Evaluated on every INPUT_TYPES call, so installing a newer
+    coreml_diffusion surfaces new conversion types with no Suite change.
+    """
+    try:
+        import coreml_diffusion
+
+        return getattr(coreml_diffusion, fn_name)()
+    except Exception as exc:  # missing/old package, import error, etc.
+        logger.warning(
+            f"coreml_diffusion.{fn_name} unavailable ({exc}); "
+            f"using fallback {fallback}"
+        )
+        return fallback
+
+
+class CoreMLSampler(COREML_NODE, KSampler):
+    @classmethod
+    def INPUT_TYPES(s):
+        old_required = KSampler.INPUT_TYPES()["required"].copy()
+        old_required.pop("model")
+        old_required.pop("negative")
+        old_required.pop("latent_image")
+        new_required = {"coreml_model": ("COREML_UNET",)}
+        return {
+            "required": new_required | old_required,
+            "optional": {"negative": ("CONDITIONING",), "latent_image": ("LATENT",)},
+        }
+
+    def sample(
+        self,
+        coreml_model,
+        seed,
+        steps,
+        cfg,
+        sampler_name,
+        scheduler,
+        positive,
+        negative=None,
+        latent_image=None,
+        denoise=1.0,
+    ):
+        model_patcher = get_model_patcher(coreml_model)
+        latent_image = get_latent_image(coreml_model, latent_image)
+
+        if is_lcm(coreml_model):
+            negative = [[None, {}]]
+            positive[0][1]["control_apply_to_uncond"] = False
+            model_patcher = add_lcm_model_options(model_patcher, cfg, latent_image)
+            model_patcher = lcm_patch(model_patcher)
+        else:
+            assert (
+                negative is not None
+            ), "Negative conditioning is optional only for LCM models."
+
+        if is_sdxl(coreml_model):
+            model_patcher = add_sdxl_model_options(model_patcher, positive, negative)
+
+        return super().sample(
+            model_patcher,
+            seed,
+            steps,
+            cfg,
+            sampler_name,
+            scheduler,
+            positive,
+            negative,
+            latent_image,
+            denoise,
+        )
+
+
+class CoreMLSamplerAdvanced(COREML_NODE, KSamplerAdvanced):
+    @classmethod
+    def INPUT_TYPES(s):
+        old_required = KSamplerAdvanced.INPUT_TYPES()["required"].copy()
+        old_required.pop("model")
+        old_required.pop("negative")
+        old_required.pop("latent_image")
+        new_required = {"coreml_model": ("COREML_UNET",)}
+        return {
+            "required": new_required | old_required,
+            "optional": {"negative": ("CONDITIONING",), "latent_image": ("LATENT",)},
+        }
+
+    def sample(
+        self,
+        coreml_model,
+        add_noise,
+        noise_seed,
+        steps,
+        cfg,
+        sampler_name,
+        scheduler,
+        positive,
+        start_at_step,
+        end_at_step,
+        return_with_leftover_noise,
+        negative=None,
+        latent_image=None,
+        denoise=1.0,
+    ):
+        model_patcher = get_model_patcher(coreml_model)
+        latent_image = get_latent_image(coreml_model, latent_image)
+
+        if is_lcm(coreml_model):
+            negative = [[None, {}]]
+            positive[0][1]["control_apply_to_uncond"] = False
+            model_patcher = add_lcm_model_options(model_patcher, cfg, latent_image)
+            model_patcher = lcm_patch(model_patcher)
+        else:
+            assert (
+                negative is not None
+            ), "Negative conditioning is optional only for LCM models."
+
+        if is_sdxl(coreml_model):
+            model_patcher = add_sdxl_model_options(model_patcher, positive, negative)
+
+        return super().sample(
+            model_patcher,
+            add_noise,
+            noise_seed,
+            steps,
+            cfg,
+            sampler_name,
+            scheduler,
+            positive,
+            negative,
+            latent_image,
+            start_at_step,
+            end_at_step,
+            return_with_leftover_noise,
+            denoise,
+        )
+
+
+class CoreMLLoader(COREML_NODE):
+    PACKAGE_DIRNAME = ""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "coreml_name": (list(s.coreml_filenames().keys()),),
+                "compute_unit": (
+                    [
+                        ComputeUnit.CPU_AND_NE.name,
+                        ComputeUnit.CPU_AND_GPU.name,
+                        ComputeUnit.ALL.name,
+                        ComputeUnit.CPU_ONLY.name,
+                    ],
+                ),
+            }
+        }
+
+    FUNCTION = "load"
+
+    @classmethod
+    def coreml_filenames(cls):
+        extensions = (".mlpackage",)
+        all_paths = folder_paths.get_filename_list_(cls.PACKAGE_DIRNAME)[1]
+        coreml_paths = folder_paths.filter_files_extensions(all_paths, extensions)
+
+        return {os.path.split(p)[-1]: p for p in coreml_paths}
+
+    def load(self, coreml_name, compute_unit):
+        logger.info(f"Loading {coreml_name} to {compute_unit}")
+
+        coreml_path = self.coreml_filenames()[coreml_name]
+
+        return (CoreMLModel(coreml_path, compute_unit),)
+
+
+class CoreMLLoaderUNet(CoreMLLoader):
+    PACKAGE_DIRNAME = "unet"
+    RETURN_TYPES = ("COREML_UNET",)
+    RETURN_NAMES = ("coreml_model",)
+
+
+class CoreMLModelAdapter(COREML_NODE):
+    """
+    Adapter Node to use CoreML models as Comfy models. This is an experimental
+    feature and may not work as expected.
+    """
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "coreml_model": ("COREML_UNET",),
+            }
+        }
+
+    RETURN_TYPES = ("MODEL",)
+
+    FUNCTION = "wrap"
+    CATEGORY = "Core ML Suite"
+
+    def wrap(self, coreml_model):
+        model_patcher = get_model_patcher(coreml_model)
+        return (model_patcher,)
+
+
+class CoreMLConverter(COREML_NODE):
+    """Converts a Stable Diffusion checkpoint (UNet) to Core ML.
+
+    The model version (SD15 / SDXL / SDXL refiner / LCM) is auto-detected from
+    the checkpoint's architecture, so there is no version dropdown — one node
+    converts every supported family, including full-distill LCM.
+    """
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "ckpt_name": (folder_paths.get_filename_list("checkpoints"),),
+                "height": ("INT", {"default": 512, "min": 8, "step": 8}),
+                "width": ("INT", {"default": 512, "min": 8, "step": 8}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
+                "attention_implementation": (
+                    _discover(
+                        "list_attention_impls",
+                        ["SPLIT_EINSUM", "SPLIT_EINSUM_V2", "ORIGINAL"],
+                    ),
+                ),
+                "compute_unit": (
+                    [
+                        ComputeUnit.CPU_AND_NE.name,
+                        ComputeUnit.CPU_AND_GPU.name,
+                        ComputeUnit.ALL.name,
+                        ComputeUnit.CPU_ONLY.name,
+                    ],
+                ),
+                "controlnet_support": ("BOOLEAN", {"default": False}),
+            },
+            "optional": {
+                # k-means weight palettization. Kept optional so workflows
+                # that omit it still validate — ComfyUI rejects a prompt that
+                # omits any `required` input. When omitted it defaults to
+                # "none", identical to unquantized behavior and filename, so
+                # existing cached .mlpackages still resolve.
+                "quantize_nbits": (
+                    _discover("list_quant_modes", ["none", "8", "6", "4"]),
+                    {"default": "none"},
+                ),
+                "lora_params": ("LORA_PARAMS",),
+            },
+        }
+
+    RETURN_TYPES = ("COREML_UNET",)
+    RETURN_NAMES = ("coreml_model",)
+    FUNCTION = "convert"
+
+    def convert(
+        self,
+        ckpt_name,
+        height,
+        width,
+        batch_size,
+        attention_implementation,
+        compute_unit,
+        controlnet_support,
+        quantize_nbits="none",
+        lora_params=None,
+    ):
+        """Converts a checkpoint's UNet to Core ML.
+
+        Args:
+            ckpt_name (str): Checkpoint to convert; its model version is
+                auto-detected from the weights.
+            height (int): Height of the target image.
+            width (int): Width of the target image.
+            batch_size (int): Batch size.
+            compute_unit (str): Compute unit to use when loading the model.
+
+        Returns:
+            coreml_model: The converted Core ML model.
+
+        The converted model is also saved to "models/unet" directory and
+        can be loaded with the "Load Core ML UNet" node.
+        """
+        lora_params = lora_params or {}
+        lora_params = [(k, v[0]) for k, v in lora_params.items()]
+        lora_params = sorted(lora_params, key=lambda lora: lora[0])
+        lora_weights = [(self.lora_path(lora[0]), lora[1]) for lora in lora_params]
+
+        h = height
+        w = width
+        sample_size = (h // 8, w // 8)
+        import coreml_diffusion
+
+        out_name = coreml_diffusion.compose_out_name(
+            ckpt_name=ckpt_name,
+            batch_size=batch_size,
+            width=w,
+            height=h,
+            controlnet_support=controlnet_support,
+            attention_implementation=attention_implementation,
+            lora_names=coreml_diffusion.lora_names_from_params(lora_params),
+            quantize_nbits=quantize_nbits,
+        )
+
+        logger.info(f"Converting {ckpt_name} to {out_name}")
+        logger.info(f"Batch size: {batch_size}")
+        logger.info(f"Width: {w}, Height: {h}")
+        logger.info(f"ControlNet support: {controlnet_support}")
+        logger.info(f"Attention implementation: {attention_implementation}")
+
+        if lora_params:
+            logger.info("LoRAs used:")
+            for lora_param in lora_params:
+                logger.info(f"  {lora_param[0]} - strength: {lora_param[1]}")
+
+        # Resolve the ComfyUI models/unet path here (a node concern); the package
+        # takes the output path as an injected argument.
+        unet_path = folder_paths.get_folder_paths("unet")[0]
+        unet_out_path = os.path.join(unet_path, f"{out_name}_unet.mlpackage")
+        ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
+
+        config_filename = ckpt_name.split(".")[0] + ".yaml"
+        config_path = folder_paths.get_full_path("configs", config_filename)
+        if config_path:
+            logger.info(f"Using config file {config_path}")
+
+        coreml_diffusion.convert(
+            ckpt_path,
+            None,  # model_version auto-detected from the checkpoint
+            unet_out_path,
+            sample_size=sample_size,
+            batch_size=batch_size,
+            controlnet_support=controlnet_support,
+            lora_weights=lora_weights,
+            attn_impl=attention_implementation,
+            config_path=config_path,
+            quantize_nbits=quantize_nbits,
+        )
+        return (CoreMLModel(unet_out_path, compute_unit),)
+
+    @staticmethod
+    def lora_path(lora_name):
+        return folder_paths.get_full_path("loras", lora_name)
+
+
+class COREML_LOAD_LORA(COREML_NODE, LoraLoader):
+    @classmethod
+    def INPUT_TYPES(s):
+        required = LoraLoader.INPUT_TYPES()["required"].copy()
+        required.pop("model")
+        return {
+            "required": required,
+            "optional": {"lora_params": ("LORA_PARAMS",)},
+        }
+
+    RETURN_TYPES = ("CLIP", "LORA_PARAMS")
+    RETURN_NAMES = ("CLIP", "lora_params")
+
+    def load_lora(
+        self, clip, lora_name, strength_model, strength_clip, lora_params=None
+    ):
+        _, lora_clip = super().load_lora(
+            None, clip, lora_name, strength_model, strength_clip
+        )
+
+        lora_params = lora_params or {}
+        lora_params[lora_name] = (strength_model, strength_clip)
+
+        return lora_clip, lora_params

coreml_suite/samplers.py

@@ -1,74 +0,0 @@
-import torch
-from torchvision.transforms.functional import resize
-
-from comfy.model_management import get_torch_device
-from comfy.model_patcher import ModelPatcher
-from coreml_suite.logger import logger
-from nodes import KSampler
-
-from coreml_suite.models import CoreMLModelWrapper, get_model_config
-
-
-def reshape_latent_image(latent_image, target_shape):
-    if latent_image is None:
-        logger.warning("No latent image provided, using zeros.")
-        return {"samples": torch.zeros(target_shape)}
-
-    if latent_image["samples"].shape == target_shape:
-        return latent_image
-
-    logger.warning(
-        "Latent image shape does not match model input shape,"
-        " resizing to match models expected input shape."
-    )
-    resized = resize(latent_image["samples"], target_shape[-2:])
-    return {"samples": resized}
-
-
-class CoreMLSampler(KSampler):
-    @classmethod
-    def INPUT_TYPES(s):
-        old_required = KSampler.INPUT_TYPES()["required"].copy()
-        old_required.pop("model")
-        old_required.pop("latent_image")
-        new_required = {"coreml_model": ("COREML_UNET",)}
-        return {
-            "required": new_required | old_required,
-            "optional": {"latent_image": ("LATENT",)},
-        }
-
-    CATEGORY = "Core ML Suite"
-
-    def sample(
-        self,
-        coreml_model,
-        seed,
-        steps,
-        cfg,
-        sampler_name,
-        scheduler,
-        positive,
-        negative,
-        latent_image=None,
-        denoise=1.0,
-    ):
-        sample_shape = coreml_model.expected_inputs["sample"]["shape"]
-        latent_image = reshape_latent_image(latent_image, sample_shape)
-        latent_image["samples"] = latent_image["samples"][0:1]
-
-        model_config = get_model_config()
-        wrapped_model = CoreMLModelWrapper(model_config, coreml_model)
-        model = ModelPatcher(wrapped_model, get_torch_device(), None)
-
-        return super().sample(
-            model,
-            seed,
-            steps,
-            cfg,
-            sampler_name,
-            scheduler,
-            positive,
-            negative,
-            latent_image,
-            denoise,
-        )

coreml_suite/utils.py

@@ -1,62 +0,0 @@
-from itertools import chain
-
-import numpy as np
-import torch
-
-from coreml_suite.logger import logger
-
-
-def expand_inputs(inputs):
-    expanded = inputs.copy()
-    for k, v in inputs.items():
-        if isinstance(v, np.ndarray):
-            expanded[k] = np.concatenate([v] * 2) if v.shape[0] == 1 else v
-        elif isinstance(v, torch.Tensor):
-            expanded[k] = torch.cat([v] * 2) if v.shape[0] == 1 else v
-        elif isinstance(v, list):
-            expanded[k] = v * 2 if len(v) == 1 else v
-        elif isinstance(v, dict):
-            expand_inputs(v)
-    return expanded
-
-
-def extract_residual_kwargs(model, control):
-    if "additional_residual_0" not in model.expected_inputs.keys():
-        return {}
-    if control is None:
-        return no_control(model)
-
-    residual_kwargs = {
-        "additional_residual_{}".format(i): r.cpu().numpy().astype(np.float16)
-        for i, r in enumerate(chain(control["output"], control["middle"]))
-    }
-    return residual_kwargs
-
-
-def no_control(model):
-    # Dirty hack to get the expected input shape when doing partial ControlNet
-    # 0.18215 is the latent scale factor (IDK, it kinda works)
-    # TODO: Find a better way to do this or tweak the values
-
-    logger.warning(
-        "No ControlNet input, despite the model supports it. "
-        "Using random noise as ControlNet residuals. "
-        "For better results, please use a ControlNet or a model "
-        "that does not support ControlNet."
-    )
-    residuals_names = [
-        name
-        for name in model.expected_inputs.keys()
-        if name.startswith("additional_residual")
-    ]
-    residual_kwargs = {
-        "additional_residual_{}".format(i): 0.18215
-        * torch.randn(
-            *model.expected_inputs["additional_residual_{}".format(i)]["shape"]
-        )
-        .cpu()
-        .numpy()
-        .astype(dtype=np.float16)
-        for i in range(len(residuals_names))
-    }
-    return residual_kwargs

pyproject.toml

@@ -0,0 +1,63 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "comfyui-coremlsuite"
+description = "This extension contains a set of custom nodes for ComfyUI that allow you to use Core ML models in your ComfyUI workflows."
+version = "2.1.2"
+license = "MIT"
+requires-python = ">=3.12"
+dependencies = [
+    # torch is provided by the host (ComfyUI) and intentionally left unpinned
+    # here: a hard torch cap would downgrade the host's torch and break its
+    # torchvision/torchaudio ABI. coreml-diffusion pulls torch>=2.7 transitively.
+    # >=0.1.6: model-version auto-detection (convert(model_version=None)) and the
+    # dropped <3.13 Python cap (kept in sync with this package's requires-python).
+    "coreml-diffusion>=0.1.6,<0.2",
+    "coremltools>=9,<10",
+    "numpy>=2,<3",
+]
+
+[project.urls]
+Repository = "https://github.com/aszc-dev/ComfyUI-CoreMLSuite"
+
+[tool.hatch.build.targets.wheel]
+packages = ["coreml_suite"]
+
+[tool.comfy]
+PublisherId = "aszc-dev"
+DisplayName = "ComfyUI-CoreMLSuite"
+Icon = "https://raw.githubusercontent.com/aszc-dev/ComfyUI-CoreMLSuite/main/assets/snake.png"
+requires-comfyui = ">=0.3.27"
+
+[dependency-groups]
+dev = [
+    "pillow>=12.2.0",
+    "psutil>=7.2.2",
+    "pytest>=9.0.3",
+]
+comfy = [
+    "comfyui-frontend-package==1.14.6",
+    "torchvision",
+    "torchaudio",
+    "torchsde",
+    "einops",
+    "tokenizers>=0.13.3",
+    "safetensors>=0.4.2",
+    "aiohttp>=3.11.8",
+    "yarl>=1.18.0",
+    "kornia>=0.7.1",
+    "spandrel",
+    "soundfile",
+    "sentencepiece",
+]
+
+[tool.pytest.ini_options]
+markers = [
+    "unit: framework-free unit test (Tier 0)",
+    "smoke: macOS-ARM smoke test on a synthetic micro-model (Tier 1)",
+    "m2: requires Apple Silicon + Neural Engine (Tier 2)",
+]
+testpaths = ["tests"]
+addopts = ["--import-mode=importlib", "--confcutdir=tests"]

requirements.txt

@@ -1,2 +1,4 @@
-git+https://github.com/apple/ml-stable-diffusion.git
-coremltools
+coreml-diffusion>=0.1.4,<0.2
+coremltools>=9,<10
+numpy>=2,<3
+diffusers>=0.30

seam.md

@@ -0,0 +1,208 @@
+# Conversion Extraction — Seam Inventory (`docs/extraction/seam.md`)
+
+> **Gate E0 deliverable.** Symbol-by-symbol cut line between the future `coreml_diffusion`
+> package (CONVERSION) and what stays in `coreml_suite` (the ComfyUI side).
+>
+> **Confidence legend:**
+> - ✅ **verified** — read directly from the current source in this repo.
+> - 🔍 **confirm** — inferred / partially seen; Claude Code must `grep`-verify before acting.
+>
+> **Cut rule:** a symbol goes to `coreml_diffusion` iff it participates in producing the `.mlpackage`
+> artifact AND can be made free of `comfy` / `folder_paths` / `comfy_extras`. The runtime
+> *loader* that **runs** a compiled model stays in the suite.
+
+---
+
+## 1. File-level map
+
+| File | Side | Status | Note |
+|---|---|---|---|
+| `coreml_suite/model_version.py` | **coreml_diffusion** | ✅ | Already `Enum`-only, zero comfy. Becomes pkg source of truth. |
+| `coreml_suite/attention.py` | **coreml_diffusion** | ✅ | `ATTENTION_IMPLEMENTATIONS` tuple; pure constant. |
+| `coreml_suite/core/naming.py` | **coreml_diffusion** | ✅ | `compose_out_name` = cache-key contract. Move (not copy). |
+| `coreml_suite/converter.py` | **coreml_diffusion** (mostly) | ✅ | Main conversion. One symbol stays-adjacent: `get_out_path` (folder_paths) is replaced by injected `out_path`. |
+| `coreml_suite/conversion/attention.py` | **coreml_diffusion** | ✅ | `apply_attention_implementation`. Imports `logging`,`torch` only — no comfy. |
+| `coreml_suite/conversion/shapes.py` | **coreml_diffusion** | ✅ | `conv2d_output_shape`. Pure math, no imports. |
+| `coreml_suite/conversion/trace.py` | **coreml_diffusion** | ✅ | Imports `types.MethodType`, `diffusers...Transformer2DModel` only — torch/diffusers. |
+| `coreml_suite/conversion/unet.py` | **coreml_diffusion** | ✅ | `CoreMLUNetWrapper`. Imports `torch` only — no comfy. |
+| `coreml_suite/lcm/converter.py` | **coreml_diffusion** (after dedup) | ✅ | Dup helpers deleted; `MODEL_VERSION` HF-hardcode (L22) → E-LCM. `folder_paths` (L111) + `comfy.model_management` (L54) confirmed present → CUT. |
+| `coreml_suite/lcm/unet.py` | **coreml_diffusion** | ✅ | `UNet2DConditionModelLCM(UNet2DConditionModel)`. diffusers-only, no comfy. |
+| `coreml_suite/config.py` | **STAYS** | ✅ | Imports `comfy.supported_models_base`/`latent_formats`/`model_detection`. **Inference-side** (`get_model_config`), NOT conversion. |
+| `coreml_suite/coreml_model.py` | **STAYS** | ✅ | `CoreMLModel` = runtime loader (runs `.mlpackage`). Desktop/Python inference; not used on iOS. |
+| `coreml_suite/nodes.py` | **STAYS** | ✅ | Nodes; will call `coreml_diffusion` + own `folder_paths` path resolution + discovery dropdowns. |
+| `coreml_suite/lcm/nodes.py` | **STAYS** | ✅ | `COREML_CONVERT_LCM` node. |
+| `coreml_suite/models.py` | **STAYS** | ✅ | Inference: `add_sdxl_model_options`, `is_sdxl`, `get_model_patcher`, `get_latent_image`. |
+| `coreml_suite/latents.py` | **STAYS** | ✅ | Inference chunking (MODERNIZATION Phase 3 target, not this spec). |
+| `coreml_suite/controlnet.py` | **STAYS** | ✅ | Inference-side controlnet. Distinct from converter `add_cnet_support`. |
+| `coreml_suite/lcm/utils.py` | **STAYS** | ✅ | `add_lcm_model_options`, `lcm_patch`, `is_lcm`; imports `comfy_extras`. Inference. |
+| `coreml_suite/logger.py` | **both / copy** | ✅ | Trivial. Package gets its own logger; suite keeps its. |
+
+---
+
+## 2. Symbol-level: `coreml_suite/converter.py` (main conversion)
+
+| Symbol | Side | Status | Cut action |
+|---|---|---|---|
+| `DEFAULT_TRACE_TIMESTEP`, `TEXT_TOKEN_SEQUENCE_LENGTH` | coreml_diffusion | ✅ | Move as-is (module constants). |
+| `get_unet(model_version, ref_unet, attention_implementation)` | coreml_diffusion | ✅ | Move. Uses `conversion.{trace,attention,unet}`. No comfy. |
+| `get_encoder_hidden_states_shape(ref_unet, batch_size)` | coreml_diffusion | ✅ | Move. Reads `ref_unet.config.cross_attention_dim`. Pure. |
+| `get_coreml_inputs(sample_inputs)` | coreml_diffusion | ✅ | Move. `ct.TensorType` build. |
+| `load_coreml_model(out_path)` | coreml_diffusion | ✅ | Move. `ct.models.MLModel(out_path)`. (Dedup target vs LCM copy.) |
+| `convert_to_coreml(submodule, ts_module, inputs, names, out_path)` | coreml_diffusion | ✅ | Move. `ct.convert(...)`. (Dedup target vs LCM copy.) |
+| `get_sample_input(batch, ehs_shape, sample_shape)` | coreml_diffusion | ✅ | Move. **Merge** with LCM variant (LCM passes extra `scheduler` → optional param). |
+| `lcm_inputs(sample_unet_inputs)` | coreml_diffusion | ✅ | Move. Adds `timestep_cond`. |
+| `sdxl_inputs(sample_unet_inputs, ref_unet, model_version)` | coreml_diffusion | ✅ | Move. `time_ids`/`text_embeds`/`add_embeds`. |
+| `add_cnet_support(sample_shape, ref_unet)` | coreml_diffusion | ✅ | Move. Builds `additional_residual_*` inputs from unet block channels. |
+| `convert_unet(ref_unet, model_version, unet_out_path, ...)` | coreml_diffusion | ✅ | Move. Orchestrates trace→convert→**quant (palettize)**→save. Quant travels here (E6). |
+| `convert(ckpt_path, model_version, unet_out_path, ...)` | coreml_diffusion | ✅ | Move. **Make kw-only past `ckpt_path,model_version,out_path`** (contract). Validates `attn_impl`. |
+| `load_unet(ckpt_path, config_path)` | coreml_diffusion | ✅ | Move. `UNet2DConditionModel.from_single_file`. |
+| `get_out_path(submodule_name, model_name)` | **STAYS (node)** | ✅ | Uses `folder_paths.get_folder_paths`. **Delete from converter; node resolves path and passes `out_path` in.** |
+
+**Apple `python_coreml_stable_diffusion` footprint on this path:** ✅ **none.** Verified by grep:
+zero imports in `converter.py` / `conversion/*`. Main path uses `diffusers` +
+local `CoreMLUNetWrapper`. (And the runtime `CoreMLModel` is now a local coremltools wrapper too —
+see §6 stale-spec note.)
+
+---
+
+## 3. Symbol-level: `coreml_suite/lcm/converter.py` (LCM — dedup + defer)
+
+| Symbol | Side | Status | Cut action |
+|---|---|---|---|
+| `load_coreml_model` (LCM copy) | DELETE | ✅ | Duplicate of main. Remove; use `coreml_diffusion.load_coreml_model`. |
+| `convert_to_coreml` (LCM copy) | DELETE | ✅ | Duplicate of main. Remove. |
+| `get_out_path` (LCM copy, folder_paths) | DELETE | ✅ | Duplicate + comfy. Remove; node injects `out_path`. |
+| `get_sample_input(..., scheduler)` (LCM copy) | MERGE → coreml_diffusion | ✅ | Fold `scheduler` into shared `get_sample_input` as optional param. |
+| `MODEL_NAME` (= LCM_Dreamshaper) | **E-LCM** | ✅ | HF hardcode. Removing it is the behavior change → E-LCM, not E2. |
+| `convert(out_path, sample_size, batch_size, controlnet_support)` (LCM, L190) | coreml_diffusion (via unified) | ✅ | Route through `coreml_diffusion.convert(model_version=LCM, ...)` in E-LCM. |
+| `from comfy.model_management import get_torch_device` (L54, in `get_scheduler`) | **CUT** | ✅ | Confirmed present. Inject `device`. |
+| module-global attention set at import | n/a | ✅ | **No module global.** Attention already per-call: `get_unets` (L36) calls `apply_attention_implementation(ref_unet, "SPLIT_EINSUM")`. No `ATTENTION_IMPLEMENTATION_IN_EFFECT` anywhere in repo. (Note: LCM hardcodes `"SPLIT_EINSUM"` — pass `attn_impl` through in dedup.) |
+
+---
+
+## 4. Symbol-level: `coreml_suite/core/naming.py` → `coreml_diffusion/naming.py`
+
+| Symbol | Side | Status | Cut action |
+|---|---|---|---|
+| `compose_out_name(...)` | coreml_diffusion | ✅ | **Move** (cache-key contract). Node imports from pkg. |
+| `lora_names_from_params(...)` | coreml_diffusion | ✅ | Move. |
+| `ATTN_SUFFIX` dict | coreml_diffusion | ✅ | Move. |
+| `QUANT_NBITS_VALUES` | coreml_diffusion | ✅ | Move; backs `list_quant_modes()`. |
+| `tests/unit/test_characterization_out_name.py` | re-point | ✅ | Change import to `coreml_diffusion.naming`. Assertions/values **unchanged**. |
+
+---
+
+## 5. Discovery API + status registry (new in `coreml_diffusion/__init__.py`)
+
+```python
+from enum import Enum
+
+class Status(Enum):
+    VERIFIED = "verified"          # has a golden anchor + passing [M2-ANE] check
+    EXPERIMENTAL = "experimental"  # convertible, not yet anchored/verified
+
+# Single source of truth. Suite gates on this, NOT on a hardcoded node list.
+# KEY by ModelVersion enum MEMBER (not a bare string) so list_model_versions can
+# emit .name — see the .name decision below. Keying by the lowercase .value string
+# (as an earlier draft of this block did) returns ["sd15",...], which the node then
+# reverses via ModelVersion[...] → KeyError. Do NOT key by .value.
+_MODEL_STATUS = {
+    ModelVersion.SD15:         Status.VERIFIED,
+    ModelVersion.SDXL:         Status.VERIFIED,
+    ModelVersion.SDXL_REFINER: Status.EXPERIMENTAL,   # → VERIFIED after a refiner golden anchor
+    ModelVersion.LCM:          Status.EXPERIMENTAL,   # → VERIFIED after E-LCM golden anchor
+}
+
+def list_model_versions(include_experimental: bool = False) -> list[str]:
+    return [v.name for v, s in _MODEL_STATUS.items()   # .name → "SD15","SDXL" (see decision)
+            if s is Status.VERIFIED or (include_experimental and s is Status.EXPERIMENTAL)]
+
+def list_attention_impls() -> list[str]:   # from attention.ATTENTION_IMPLEMENTATIONS
+    ...
+def list_quant_modes() -> list[str]:       # from naming.QUANT_NBITS_VALUES
+    ...
+
+CONTRACT_VERSION = "1.0"
+# Additive-only: adding an id or promoting EXPERIMENTAL→VERIFIED = minor bump (Suite unaffected).
+# Removing/renaming an id, or demoting VERIFIED→EXPERIMENTAL = MAJOR bump + migration note.
+```
+
+**Decision check (`.name` vs `.value`): RESOLVED → `.name`.** ✅
+Verified in current source:
+- Node renders `ModelVersion.SD15.name` / `ModelVersion.SDXL.name` → `"SD15"`, `"SDXL"`
+  (`nodes.py:224-225`).
+- Node reverses the dropdown string with `model_version = ModelVersion[model_version]`
+  (`nodes.py:286`) — i.e. **lookup by NAME**. Feeding it a `.value` (`"sd15"`) raises `KeyError`.
+- Enum values are lowercase (`model_version.py`: `SD15="sd15"`, `SDXL="sdxl"`,
+  `SDXL_REFINER="sdxl_refiner"`, `LCM="lcm"`).
+- `compose_out_name` does NOT consume the model_version string (grep of `core/naming.py` empty) —
+  no coupling there, so no constraint from that side.
+
+**Decision:** `list_model_versions()` returns `.name` (uppercase). Saved workflows store `"SD15"`,
+node already validates them via `ModelVersion[...]`. The `_MODEL_STATUS` block above was corrected
+to key by enum member and emit `.name`. **The earlier `v.value` form was a latent bug.**
+
+---
+
+## 6. `python_coreml_stable_diffusion` split (Gate E0 line to fill by grep)
+
+| Use | Side | Status |
+|---|---|---|
+| `coreml_model.CoreMLModel` (runs compiled model) | **STAYS** (suite runtime) | ✅ — **local class**, not Apple's |
+| `unet.UNet2DConditionModel*` internals | **gone** — `converter.py:319` uses `diffusers.UNet2DConditionModel.from_single_file` | ✅ |
+| `AttentionImplementations` enum | gone — local `apply_attention_implementation` + `attention.py` tuple | ✅ |
+| `calculate_conv2d_output_shape` | gone — replaced by `conversion/shapes.conv2d_output_shape` | ✅ |
+
+> ### ⚠️ SPEC IS STALE: `ml-stable-diffusion` is already fully removed
+> Commit #58 ("replace apple/ml-stable-diffusion with native diffusers conversion") already did
+> the de-Apple work. Verified now:
+> - **Zero** `python_coreml_stable_diffusion` runtime imports anywhere in `coreml_suite` (only a
+>   docstring mention at `core/__init__.py:4`).
+> - `coreml_suite/coreml_model.py:8` `CoreMLModel` is a **local** wrapper over
+>   `coremltools.models.MLModel` (`coreml_model.py:22`) — it does **not** import Apple's class.
+> - `ml-stable-diffusion` / `python_coreml_stable_diffusion` appears in **neither** `pyproject.toml`
+>   **nor** `requirements.txt`. It is not a dependency at all.
+>
+> **Consequences for the spec (correct these in CONVERTER_EXTRACTION_SPEC.md):**
+> - §0.3 premise ("runtime loader = `python_coreml_stable_diffusion.coreml_model.CoreMLModel`,
+>   stays in suite") is **wrong**: the loader is already the local `coreml_model.CoreMLModel`. The
+>   "stays in suite" conclusion still holds; the identity does not.
+> - **Gate E0 item "ml-stable-diffusion pinned SHA — BLOCKER if unpinned" is MOOT** — there is no
+>   such dep to pin. Mark it N/A, not BLOCKER.
+> - **E4/E5 dependency lists must drop `git+...ml-stable-diffusion@<sha>`.** Package runtime deps
+>   are: `coremltools`, `diffusers`, `peft` (LoRA), `omegaconf` (config), `numpy`, `torch`. Confirm
+>   `peft`/`omegaconf` actually used before listing (grep at E4).
+> - The "keep `python_coreml_stable_diffusion` as a suite dep for the loader" instruction in E5 is
+>   **void** — coremltools backs the loader.
+
+---
+
+## 7. Pre-flight checklist before E1 (run these greps)
+
+```
+grep -rn "import comfy"            coreml_suite/conversion coreml_suite/converter.py coreml_suite/lcm/converter.py coreml_suite/lcm/unet.py
+grep -rn "folder_paths"            coreml_suite/converter.py coreml_suite/lcm/converter.py
+grep -rn "model_management"        coreml_suite/lcm
+grep -rn "python_coreml_stable_diffusion" coreml_suite
+grep -rn "ATTENTION_IMPLEMENTATION_IN_EFFECT" coreml_suite
+grep -rn "SimianLuo\|LCM_Dreamshaper" coreml_suite/lcm
+```
+Every 🔍 above resolves to ✅ or a correction once these run. Do not start moving code (E2)
+with any 🔍 unresolved on the CONVERSION side.
+
+**STATUS (run 2026-05-26): all 🔍 resolved.** Summary of what the greps found:
+- `conversion/*`, `lcm/unet.py`: comfy-free (torch/diffusers only). ✅
+- `converter.py`: only comfy reach-in is `folder_paths` in `get_out_path` (L91-94) → inject `out_path`.
+- `lcm/converter.py`: `folder_paths` (L111-114) + `comfy.model_management.get_torch_device` (L54)
+  → cut both. Dup helpers (`load_coreml_model`,`convert_to_coreml`,`get_out_path`,`get_sample_input`)
+  confirmed → dedup E2. `MODEL_VERSION="SimianLuo/LCM_Dreamshaper_v7"` (L22) → E-LCM.
+- No attention module-global anywhere (`ATTENTION_IMPLEMENTATION_IN_EFFECT` absent); already per-call.
+  LCM hardcodes `"SPLIT_EINSUM"` in `get_unets` — thread `attn_impl` through during dedup.
+- `.name` vs `.value`: **decided `.name`** (node reverses via `ModelVersion[...]`). §5 corrected.
+- `ml-stable-diffusion`: **already gone** (#58). §6 stale-spec note added — fix the spec's E0/E4/E5
+  dep + pinning items.
+
+Two grep blind-spots to note (the checklist above doesn't cover them, but cheap to add): the
+`folder_paths` grep only scans the two converter files — also grep `coreml_suite/lcm/utils.py`
+(it imports `comfy.model_management` at L3, but it's inference/STAYS, so fine) and confirm no other
+`conversion/` file grew a comfy import since.

tests/conftest.py

@@ -0,0 +1,61 @@
+"""Pytest bootstrap for ComfyUI-CoreMLSuite tests.
+
+- Adds the ComfyUI checkout to sys.path so the framework-coupled modules
+  that transitively import `comfy.*` resolve when pytest is invoked from
+  this package's root.
+- Auto-applies tier markers based on the directory a test lives in, so
+  individual files don't have to repeat @pytest.mark.unit / .smoke.
+"""
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+COMFY_DIR = REPO_ROOT.parents[1]
+
+for p in (str(COMFY_DIR), str(REPO_ROOT)):
+    if p not in sys.path:
+        sys.path.insert(0, p)
+
+
+_TIER_BY_DIR = {
+    "tests/unit": "unit",
+    "tests/m2": "m2",
+    "tests/integration": "m2",
+    "tests/smoke": "smoke",
+}
+
+# When the user asks for a single tier (-m unit / -m smoke), skip the other
+# directories at collection time. Tier-0 cannot afford to import tests/smoke
+# files because they pull in coremltools which Linux CI won't have.
+_TIER_DIRS = {
+    "unit": ("/tests/unit/",),
+    "m2": ("/tests/m2/", "/tests/integration/"),
+    "smoke": ("/tests/smoke/",),
+}
+
+
+def pytest_ignore_collect(collection_path, config):
+    expr = config.option.markexpr
+    if expr not in _TIER_DIRS:
+        return None
+    allowed = _TIER_DIRS[expr]
+    rel = str(collection_path).replace("\\", "/")
+    if "/tests/" not in rel:
+        return None
+    # Always allow tests/ root + the tier's own dirs.
+    if rel.endswith("/tests"):
+        return None
+    if any(frag in rel + "/" for frag in allowed):
+        return None
+    return True
+
+
+def pytest_collection_modifyitems(config, items):
+    for item in items:
+        path = str(item.fspath).replace("\\", "/")
+        for fragment, marker in _TIER_BY_DIR.items():
+            if f"/{fragment}/" in path:
+                item.add_marker(getattr(pytest.mark, marker))
+                break

tests/integration/workflows/e2e-1.5-basic-conversion.json

@@ -0,0 +1,181 @@
+{
+  "3": {
+    "inputs": {
+      "seed": 0,
+      "steps": 20,
+      "cfg": 8,
+      "sampler_name": "dpmpp_2m",
+      "scheduler": "karras",
+      "denoise": 1,
+      "model": [
+        "4",
+        0
+      ],
+      "positive": [
+        "6",
+        0
+      ],
+      "negative": [
+        "7",
+        0
+      ],
+      "latent_image": [
+        "5",
+        0
+      ]
+    },
+    "class_type": "KSampler",
+    "_meta": {
+      "title": "KSampler"
+    }
+  },
+  "4": {
+    "inputs": {
+      "ckpt_name": "dreamshaper_8.safetensors"
+    },
+    "class_type": "CheckpointLoaderSimple",
+    "_meta": {
+      "title": "Load Checkpoint"
+    }
+  },
+  "5": {
+    "inputs": {
+      "width": 512,
+      "height": 512,
+      "batch_size": 1
+    },
+    "class_type": "EmptyLatentImage",
+    "_meta": {
+      "title": "Empty Latent Image"
+    }
+  },
+  "6": {
+    "inputs": {
+      "text": "beautiful scenery nature glass bottle landscape, purple galaxy bottle",
+      "clip": [
+        "4",
+        1
+      ]
+    },
+    "class_type": "CLIPTextEncode",
+    "_meta": {
+      "title": "CLIP Text Encode (Prompt)"
+    }
+  },
+  "7": {
+    "inputs": {
+      "text": "text, watermark",
+      "clip": [
+        "4",
+        1
+      ]
+    },
+    "class_type": "CLIPTextEncode",
+    "_meta": {
+      "title": "CLIP Text Encode (Prompt)"
+    }
+  },
+  "8": {
+    "inputs": {
+      "samples": [
+        "3",
+        0
+      ],
+      "vae": [
+        "4",
+        2
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAE Decode"
+    }
+  },
+  "9": {
+    "inputs": {
+      "filename_prefix": "E2E-1.5-MPS",
+      "images": [
+        "8",
+        0
+      ]
+    },
+    "class_type": "SaveImage",
+    "_meta": {
+      "title": "Save Image"
+    }
+  },
+  "10": {
+    "inputs": {
+      "ckpt_name": "dreamshaper_8.safetensors",
+      "height": 512,
+      "width": 512,
+      "batch_size": 1,
+      "attention_implementation": "SPLIT_EINSUM",
+      "compute_unit": "CPU_AND_NE",
+      "controlnet_support": false
+    },
+    "class_type": "Core ML Converter",
+    "_meta": {
+      "title": "Convert Checkpoint to Core ML"
+    }
+  },
+  "11": {
+    "inputs": {
+      "seed": 0,
+      "steps": 20,
+      "cfg": 8,
+      "sampler_name": "dpmpp_2m",
+      "scheduler": "karras",
+      "denoise": 1,
+      "coreml_model": [
+        "10",
+        0
+      ],
+      "positive": [
+        "6",
+        0
+      ],
+      "negative": [
+        "7",
+        0
+      ],
+      "latent_image": [
+        "5",
+        0
+      ]
+    },
+    "class_type": "CoreMLSampler",
+    "_meta": {
+      "title": "Core ML Sampler"
+    }
+  },
+  "13": {
+    "inputs": {
+      "samples": [
+        "11",
+        0
+      ],
+      "vae": [
+        "4",
+        2
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAE Decode"
+    }
+  },
+  "14": {
+    "inputs": {
+      "filename_prefix": "E2E-1.5-CoreML",
+      "images": [
+        "13",
+        0
+      ]
+    },
+    "class_type": "SaveImage",
+    "_meta": {
+      "title": "Save Image"
+    }
+  }
+}

tests/m2/goldens/sd15_seed42.sha256

@@ -0,0 +1 @@
+e89344e544d4edfbd3ebe9a1c78dadb2729f53549666052b74ac7308f326f4fc

tests/m2/test_golden_image.py

@@ -0,0 +1,170 @@
+"""[M2-ANE] golden-image anchor.
+
+Runs the e2e SD1.5 + CoreML workflow against a local ComfyUI server, fetches
+the generated PNG, and asserts both:
+  - byte-identical SHA256 against the stored golden, OR
+  - PSNR >= GOLDEN_PSNR_MIN_DB against the stored golden PNG.
+
+The hash is the strict gate (a refactor that doesn't touch the math
+should hit it). PSNR is the soft gate that tolerates the drift a
+toolchain bump injects through different MIL graphs / kernel selection
+/ fp accumulation order — anything below the threshold is treated as a
+regression.
+
+The 20 dB default absorbs Apple Neural Engine run-to-run nondeterminism:
+the same model and seed can drift several dB between runs as the 20
+sampling steps amplify tiny per-step UNet differences (kernel selection /
+fp accumulation order). Same-scene ANE outputs have been observed at
+~23 dB, so 20 leaves margin while still catching gross regressions — a
+broken image lands far lower. Bump it up for pure-refactor PRs that must
+not change math; down for toolchain bumps.
+
+Skips entirely on non-Apple-Silicon hosts or when the server / converted
+model is missing, so the unit lane on Linux still passes.
+
+The first run with no golden writes one and fails so it's reviewed before
+being committed.
+"""
+import hashlib
+import json
+import os
+import platform
+import shutil
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+import numpy as np
+import pytest
+from PIL import Image
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+COMFY_DIR = Path(os.environ.get("COMFY_DIR", REPO_ROOT.parents[1])).resolve()
+COMFY_HOST = os.environ.get("COMFY_HOST", "localhost")
+COMFY_PORT = int(os.environ.get("COMFY_PORT", "8188"))
+COMFY_URL = f"http://{COMFY_HOST}:{COMFY_PORT}"
+
+CKPT_NAME = os.environ.get("CKPT_NAME", "v1-5-pruned-emaonly.safetensors")
+WORKFLOW_PATH = (
+    REPO_ROOT / "tests" / "integration" / "workflows" / "e2e-1.5-basic-conversion.json"
+)
+GOLDEN_DIR = Path(__file__).parent / "goldens"
+GOLDEN_HASH_PATH = GOLDEN_DIR / "sd15_seed42.sha256"
+GOLDEN_PNG_PATH = GOLDEN_DIR / "sd15_seed42.png"
+GOLDEN_PSNR_MIN_DB = float(os.environ.get("GOLDEN_PSNR_MIN_DB", "20"))
+SEED = 42
+
+
+def _server_reachable() -> bool:
+    try:
+        with urllib.request.urlopen(f"{COMFY_URL}/prompt", timeout=3) as r:
+            return r.status == 200
+    except (urllib.error.URLError, urllib.error.HTTPError, ConnectionError):
+        return False
+
+
+@pytest.fixture(scope="module")
+def comfy_server():
+    if platform.machine() != "arm64":
+        pytest.skip("requires Apple Silicon")
+    if not _server_reachable():
+        pytest.skip(f"ComfyUI server not reachable at {COMFY_URL}")
+    return COMFY_URL
+
+
+def _http_post_json(path: str, payload: dict) -> dict:
+    data = json.dumps(payload).encode("utf-8")
+    req = urllib.request.Request(
+        f"{COMFY_URL}{path}", data=data,
+        headers={"Content-Type": "application/json"}, method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=300) as r:
+        return json.loads(r.read().decode())
+
+
+def _http_get_json(path: str, timeout: int = 300) -> dict:
+    """ComfyUI runs UNet inference on its single asyncio loop, so GET /prompt
+    blocks while the queued prompt is executing. Use a generous timeout."""
+    with urllib.request.urlopen(f"{COMFY_URL}{path}", timeout=timeout) as r:
+        return json.loads(r.read().decode())
+
+
+def _drain_queue(timeout_s: int = 600) -> None:
+    deadline = time.time() + timeout_s
+    while time.time() < deadline:
+        try:
+            q = _http_get_json("/prompt")
+        except (urllib.error.URLError, TimeoutError):
+            # Transient block while server executes; retry until our overall
+            # deadline expires.
+            continue
+        if q.get("exec_info", {}).get("queue_remaining", -1) == 0:
+            return
+        time.sleep(2)
+    raise TimeoutError(f"queue did not drain within {timeout_s}s")
+
+
+def _post_workflow_and_collect_png() -> bytes:
+    workflow = json.loads(WORKFLOW_PATH.read_text())
+    for nid in ("4", "10"):
+        if nid in workflow:
+            workflow[nid]["inputs"]["ckpt_name"] = CKPT_NAME
+    for nid in ("3", "11"):
+        if nid in workflow and "seed" in workflow[nid].get("inputs", {}):
+            workflow[nid]["inputs"]["seed"] = SEED
+    # Drop the MPS reference branch — only the Core ML pipeline is needed here.
+    for nid in ("3", "8", "9"):
+        workflow.pop(nid, None)
+
+    _http_post_json("/prompt", {"prompt": workflow})
+    _drain_queue()
+
+    comfy_out = COMFY_DIR / "output"
+    matches = sorted(comfy_out.glob("E2E-1.5-CoreML_*.png"), reverse=True)
+    if not matches:
+        raise FileNotFoundError(f"no Core ML image under {comfy_out}")
+    return matches[0].read_bytes()
+
+
+def _psnr(a: np.ndarray, b: np.ndarray) -> float:
+    mse = float(np.mean((a.astype(np.float64) - b.astype(np.float64)) ** 2))
+    if mse == 0:
+        return 100.0
+    return 20.0 * float(np.log10(255.0 / np.sqrt(mse)))
+
+
+def test_sd15_seed42_image_matches_golden(comfy_server):
+    GOLDEN_DIR.mkdir(parents=True, exist_ok=True)
+    png_bytes = _post_workflow_and_collect_png()
+    sha = hashlib.sha256(png_bytes).hexdigest()
+
+    if not GOLDEN_HASH_PATH.exists() or not GOLDEN_PNG_PATH.exists():
+        GOLDEN_HASH_PATH.write_text(sha + "\n")
+        # Persist the PNG too for visual diffing + PSNR.
+        tmp_path = Path(__file__).parent / "_latest_generated.png"
+        tmp_path.write_bytes(png_bytes)
+        shutil.copy2(tmp_path, GOLDEN_PNG_PATH)
+        pytest.fail(
+            f"No golden present; wrote {GOLDEN_HASH_PATH.name} and "
+            f"{GOLDEN_PNG_PATH.name}. Review the image and re-run."
+        )
+
+    expected_hash = GOLDEN_HASH_PATH.read_text().strip()
+    if sha == expected_hash:
+        return
+
+    # Hash drift: fall back to PSNR to distinguish a refactor-safe rounding
+    # change from a real regression.
+    a = np.array(Image.open(GOLDEN_PNG_PATH).convert("RGB"))
+    b_path = Path(__file__).parent / "_latest_generated.png"
+    b_path.write_bytes(png_bytes)
+    b = np.array(Image.open(b_path).convert("RGB"))
+    if a.shape != b.shape:
+        pytest.fail(f"shape mismatch: golden={a.shape} actual={b.shape}")
+    psnr_db = _psnr(a, b)
+    assert psnr_db >= GOLDEN_PSNR_MIN_DB, (
+        f"hash drifted (got {sha[:12]}.., expected {expected_hash[:12]}..) and "
+        f"PSNR {psnr_db:.2f} dB < {GOLDEN_PSNR_MIN_DB} dB threshold; "
+        f"diff PNG at {b_path}"
+    )

tests/test_latents.py

@@ -1,19 +0,0 @@
-import pytest
-
-import torch
-
-from coreml_suite.samplers import reshape_latent_image
-
-
-def test_fix_latents_no_latent_image():
-    reshaped = reshape_latent_image(None, (2, 4, 64, 64))
-    assert reshaped["samples"].shape == (2, 4, 64, 64)
-
-
-@pytest.mark.parametrize(
-    "latent_shape", [(2, 4, 64, 64), (2, 4, 128, 128), (2, 4, 32, 32), (2, 4, 128, 64)]
-)
-def test_reshape_latents(latent_shape):
-    latent_image = {"samples": torch.zeros(latent_shape)}
-    reshaped = reshape_latent_image(latent_image, (2, 4, 64, 64))
-    assert reshaped["samples"].shape == (2, 4, 64, 64)

tests/unit/test_characterization_controlnet.py

@@ -0,0 +1,186 @@
+"""Characterization tests for coreml_suite.controlnet.
+
+Locks shapes + dtypes + zero-fill behavior of expand_inputs / no_control /
+extract_residual_kwargs / chunk_control. These pure helpers feed the Core ML
+UNet's additional_residual_N inputs; any drift here silently breaks
+ControlNet-based workflows.
+"""
+import numpy as np
+import pytest
+import torch
+
+from coreml_suite.core.controlnet import (
+    chunk_control,
+    expand_inputs,
+    extract_residual_kwargs,
+    no_control,
+)
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+    np.random.seed(0)
+
+
+SD15_RESIDUAL_SPEC = {
+    "additional_residual_0": {"shape": (2, 320, 64, 64)},
+    "additional_residual_1": {"shape": (2, 640, 32, 32)},
+    "additional_residual_2": {"shape": (2, 1280, 8, 8)},
+}
+NON_RESIDUAL_SPEC = {
+    "sample": {"shape": (2, 4, 64, 64)},
+    "encoder_hidden_states": {"shape": (2, 77, 768)},
+}
+
+
+# ---------- expand_inputs ----------------------------------------------------
+
+
+def test_expand_inputs_doubles_singleton_numpy():
+    inputs = {"a": np.ones((1, 4), dtype=np.float32)}
+    out = expand_inputs(inputs)
+    assert out["a"].shape == (2, 4)
+    assert np.array_equal(out["a"], np.ones((2, 4)))
+
+
+def test_expand_inputs_doubles_singleton_torch():
+    inputs = {"a": torch.ones(1, 4)}
+    out = expand_inputs(inputs)
+    assert out["a"].shape == (2, 4)
+    assert torch.equal(out["a"], torch.ones(2, 4))
+
+
+def test_expand_inputs_doubles_singleton_list():
+    inputs = {"a": [42]}
+    out = expand_inputs(inputs)
+    assert out["a"] == [42, 42]
+
+
+def test_expand_inputs_skips_already_batched():
+    """batch > 1 inputs are returned unchanged (same object identity)."""
+    arr = np.ones((2, 4), dtype=np.float32)
+    tensor = torch.ones(3, 4)
+    lst = [1, 2]
+    out = expand_inputs({"a": arr, "b": tensor, "c": lst})
+    assert out["a"] is arr
+    assert out["b"] is tensor
+    assert out["c"] is lst
+
+
+def test_expand_inputs_preserves_unknown_value_types():
+    # Strings/None pass through untouched — locks current permissive contract.
+    inputs = {"s": "hello", "none": None, "int": 7}
+    out = expand_inputs(inputs)
+    assert out == {"s": "hello", "none": None, "int": 7}
+
+
+# ---------- no_control -------------------------------------------------------
+
+
+def test_no_control_returns_zero_fp16_for_residuals():
+    out = no_control({**SD15_RESIDUAL_SPEC, **NON_RESIDUAL_SPEC})
+    # Only additional_residual_* keys are produced.
+    assert set(out.keys()) == set(SD15_RESIDUAL_SPEC.keys())
+    for key, spec in SD15_RESIDUAL_SPEC.items():
+        arr = out[key]
+        assert arr.shape == spec["shape"]
+        assert arr.dtype == np.float16
+        assert np.all(arr == 0)
+
+
+def test_no_control_returns_empty_when_no_residuals():
+    out = no_control(NON_RESIDUAL_SPEC)
+    assert out == {}
+
+
+# ---------- extract_residual_kwargs -----------------------------------------
+
+
+def test_extract_residual_kwargs_empty_when_model_has_no_residual_inputs():
+    out = extract_residual_kwargs(NON_RESIDUAL_SPEC, control={"output": [], "middle": []})
+    assert out == {}
+
+
+def test_extract_residual_kwargs_none_control_returns_no_control_shapes():
+    out = extract_residual_kwargs(SD15_RESIDUAL_SPEC, control=None)
+    assert set(out.keys()) == set(SD15_RESIDUAL_SPEC.keys())
+    for key, spec in SD15_RESIDUAL_SPEC.items():
+        assert out[key].shape == spec["shape"]
+        assert out[key].dtype == np.float16
+        assert np.all(out[key] == 0)
+
+
+def test_extract_residual_kwargs_flattens_output_then_middle_and_casts_fp16():
+    """output residuals come first (indexed 0..N-1), then middle residuals
+    (indexed N..M-1). Values come out of CPU as fp16 numpy arrays."""
+    control = {
+        "output": [torch.ones(2, 320, 64, 64) * 0.5, torch.ones(2, 640, 32, 32) * 2.0],
+        "middle": [torch.ones(2, 1280, 8, 8) * -1.0],
+    }
+    out = extract_residual_kwargs(SD15_RESIDUAL_SPEC, control)
+    assert set(out.keys()) == {"additional_residual_0", "additional_residual_1", "additional_residual_2"}
+    assert out["additional_residual_0"].shape == (2, 320, 64, 64)
+    assert out["additional_residual_1"].shape == (2, 640, 32, 32)
+    assert out["additional_residual_2"].shape == (2, 1280, 8, 8)
+    for arr in out.values():
+        assert arr.dtype == np.float16
+    # Locked order: index 0 == first output residual (0.5), index 2 == middle (-1.0).
+    assert np.allclose(out["additional_residual_0"], 0.5)
+    assert np.allclose(out["additional_residual_1"], 2.0)
+    assert np.allclose(out["additional_residual_2"], -1.0)
+
+
+# ---------- chunk_control ----------------------------------------------------
+
+
+def test_chunk_control_none_returns_list_of_nones_with_length_target():
+    """`no_control` path: when there's no control, you get [None] * target_size
+    (NOT [None, None] regardless of target — this is the contract today)."""
+    assert chunk_control(None, 1) == [None]
+    assert chunk_control(None, 2) == [None, None]
+    assert chunk_control(None, 4) == [None, None, None, None]
+
+
+@pytest.mark.parametrize(
+    "batch,target,expected_chunks",
+    [(1, 2, 1), (2, 2, 1), (3, 2, 2), (4, 2, 2), (5, 3, 2), (9, 4, 3)],
+)
+def test_chunk_control_shapes_after_chunking(batch, target, expected_chunks):
+    cn = {
+        "output": [
+            torch.randn(batch, 320, 64, 64),
+            torch.randn(batch, 640, 32, 32),
+        ],
+        "middle": [torch.randn(batch, 1280, 8, 8)],
+    }
+    chunks = chunk_control(cn, target)
+    assert len(chunks) == expected_chunks
+    for c in chunks:
+        assert c["output"][0].shape == (target, 320, 64, 64)
+        assert c["output"][1].shape == (target, 640, 32, 32)
+        assert c["middle"][0].shape == (target, 1280, 8, 8)
+
+
+def test_chunk_control_preserves_keys_order():
+    """Output dicts contain exactly {"output", "middle"} in that order."""
+    cn = {
+        "output": [torch.zeros(2, 4, 4, 4)],
+        "middle": [torch.zeros(2, 4, 4, 4)],
+    }
+    chunks = chunk_control(cn, 2)
+    assert list(chunks[0].keys()) == ["output", "middle"]
+
+
+def test_chunk_control_zero_pads_remainder():
+    """A batch=3, target=2 split puts the third row alongside a zero row."""
+    cn = {
+        "output": [torch.arange(3 * 4).reshape(3, 1, 2, 2).float()],
+        "middle": [torch.arange(3 * 4).reshape(3, 1, 2, 2).float()],
+    }
+    chunks = chunk_control(cn, 2)
+    assert len(chunks) == 2
+    last_out = chunks[-1]["output"][0]
+    # First row is the original third row; second row is padding zeros.
+    assert torch.equal(last_out[0], cn["output"][0][2])
+    assert torch.equal(last_out[1], torch.zeros(1, 2, 2))

tests/unit/test_characterization_inputs.py

@@ -0,0 +1,228 @@
+"""Characterization tests for coreml_suite.models.CoreMLInputs.
+
+Locks the shape transforms applied by chunks() and coreml_kwargs() for the
+four model variants the suite supports: SD1.5, LCM (SD1.5 + timestep_cond),
+SDXL base (time_ids len 6), and SDXL refiner (time_ids len 5).
+
+These contracts feed the Core ML UNet at runtime; if a refactor silently
+re-shapes them, generation breaks.
+"""
+import numpy as np
+import pytest
+import torch
+
+from coreml_suite.core.inputs import CoreMLInputs
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+    np.random.seed(0)
+
+
+# ---------- expected_inputs fixtures (mirror real model expectations) -------
+
+SD15_EXPECTED = {
+    "sample": {"shape": (2, 4, 64, 64)},
+    "timestep": {"shape": (2,)},
+    "encoder_hidden_states": {"shape": (2, 77, 768)},
+}
+
+SD15_WITH_CN = {
+    **SD15_EXPECTED,
+    "additional_residual_0": {"shape": (2, 320, 64, 64)},
+    "additional_residual_1": {"shape": (2, 640, 32, 32)},
+}
+
+LCM_EXPECTED = {
+    **SD15_EXPECTED,
+    "timestep_cond": {"shape": (2, 256)},
+}
+
+SDXL_BASE_EXPECTED = {
+    "sample": {"shape": (2, 4, 128, 128)},
+    "timestep": {"shape": (2,)},
+    "encoder_hidden_states": {"shape": (2, 77, 2048)},
+    "time_ids": {"shape": (2, 6)},
+    "text_embeds": {"shape": (2, 1280)},
+}
+
+SDXL_REFINER_EXPECTED = {
+    "sample": {"shape": (2, 4, 128, 128)},
+    "timestep": {"shape": (2,)},
+    "encoder_hidden_states": {"shape": (2, 77, 1280)},
+    "time_ids": {"shape": (2, 5)},
+    "text_embeds": {"shape": (2, 1280)},
+}
+
+
+def _sd15_inputs(batch=1, with_control=False, with_ts_cond=False):
+    x = torch.randn(batch, 4, 64, 64)
+    t = torch.full((batch,), 999.0)
+    context = torch.randn(batch, 77, 768)
+    control = None
+    if with_control:
+        control = {
+            "output": [torch.randn(batch, 320, 64, 64), torch.randn(batch, 640, 32, 32)],
+            "middle": [],
+        }
+    kwargs = {}
+    if with_ts_cond:
+        kwargs["timestep_cond"] = torch.randn(batch, 256)
+    return CoreMLInputs(x, t, context, control, **kwargs)
+
+
+def _sdxl_inputs(batch=1, refiner=False):
+    x = torch.randn(batch, 4, 128, 128)
+    t = torch.full((batch,), 999.0)
+    ctx_dim = 1280 if refiner else 2048
+    context = torch.randn(batch, 77, ctx_dim)
+    time_ids_dim = 5 if refiner else 6
+    time_ids = torch.randn(batch, time_ids_dim)
+    text_embeds = torch.randn(batch, 1280)
+    return CoreMLInputs(
+        x, t, context, control=None, time_ids=time_ids, text_embeds=text_embeds
+    )
+
+
+# ---------- coreml_kwargs ---------------------------------------------------
+
+
+def test_coreml_kwargs_sd15_shapes_and_fp16():
+    out = _sd15_inputs(batch=1).coreml_kwargs(SD15_EXPECTED)
+    assert set(out.keys()) == {"sample", "encoder_hidden_states", "timestep"}
+    assert out["sample"].shape == (1, 4, 64, 64)
+    assert out["sample"].dtype == np.float16
+    # encoder_hidden_states keeps Comfy's native (b, seq, dim) layout.
+    assert out["encoder_hidden_states"].shape == (1, 77, 768)
+    assert out["encoder_hidden_states"].dtype == np.float16
+    assert out["timestep"].shape == (1,)
+    assert out["timestep"].dtype == np.float16
+
+
+def test_coreml_kwargs_sd15_with_controlnet_emits_residuals():
+    inputs = _sd15_inputs(batch=1, with_control=True)
+    out = inputs.coreml_kwargs(SD15_WITH_CN)
+    assert "additional_residual_0" in out
+    assert "additional_residual_1" in out
+    assert out["additional_residual_0"].shape == (1, 320, 64, 64)
+    assert out["additional_residual_1"].shape == (1, 640, 32, 32)
+
+
+def test_coreml_kwargs_sd15_without_controlnet_zero_fills_residuals():
+    inputs = _sd15_inputs(batch=1, with_control=False)
+    out = inputs.coreml_kwargs(SD15_WITH_CN)
+    assert np.all(out["additional_residual_0"] == 0)
+    assert np.all(out["additional_residual_1"] == 0)
+
+
+def test_coreml_kwargs_lcm_adds_timestep_cond():
+    inputs = _sd15_inputs(batch=1, with_ts_cond=True)
+    out = inputs.coreml_kwargs(LCM_EXPECTED)
+    assert "timestep_cond" in out
+    assert out["timestep_cond"].shape == (1, 256)
+    assert out["timestep_cond"].dtype == np.float16
+
+
+def test_coreml_kwargs_lcm_skips_timestep_cond_when_not_provided():
+    """timestep_cond is only forwarded when the input supplied one — even if
+    the model's expected_inputs lists it."""
+    inputs = _sd15_inputs(batch=1, with_ts_cond=False)
+    out = inputs.coreml_kwargs(LCM_EXPECTED)
+    assert "timestep_cond" not in out
+
+
+def test_coreml_kwargs_sdxl_base_emits_time_ids_and_text_embeds():
+    out = _sdxl_inputs(batch=1, refiner=False).coreml_kwargs(SDXL_BASE_EXPECTED)
+    assert out["time_ids"].shape == (1, 6)
+    assert out["text_embeds"].shape == (1, 1280)
+    assert out["time_ids"].dtype == np.float16
+    assert out["text_embeds"].dtype == np.float16
+
+
+def test_coreml_kwargs_sdxl_refiner_uses_len5_time_ids():
+    out = _sdxl_inputs(batch=1, refiner=True).coreml_kwargs(SDXL_REFINER_EXPECTED)
+    assert out["time_ids"].shape == (1, 5)
+
+
+# ---------- chunks ----------------------------------------------------------
+
+
+def test_chunks_sd15_pad_to_batch2_returns_one_chunk():
+    chunked = _sd15_inputs(batch=1).chunks(SD15_EXPECTED)
+    assert len(chunked) == 1
+    c = chunked[0]
+    assert c.x.shape == (2, 4, 64, 64)
+    assert c.t.shape == (2,)
+    # context shape: (b, seq, dim) padded along batch dim.
+    assert c.context.shape == (2, 77, 768)
+    assert c.control is None
+    assert c.ts_cond is None
+    assert c.time_ids is None
+    assert c.text_embeds is None
+
+
+def test_chunks_sd15_with_controlnet_chunks_residuals_too():
+    chunked = _sd15_inputs(batch=1, with_control=True).chunks(SD15_EXPECTED)
+    assert len(chunked) == 1
+    cn = chunked[0].control
+    assert cn is not None
+    assert cn["output"][0].shape == (2, 320, 64, 64)
+    assert cn["output"][1].shape == (2, 640, 32, 32)
+
+
+def test_chunks_lcm_carries_timestep_cond_per_chunk():
+    chunked = _sd15_inputs(batch=1, with_ts_cond=True).chunks(LCM_EXPECTED)
+    assert len(chunked) == 1
+    assert chunked[0].ts_cond is not None
+    assert chunked[0].ts_cond.shape == (2, 256)
+
+
+def test_chunks_sdxl_base_propagates_time_ids_and_text_embeds():
+    chunked = _sdxl_inputs(batch=1, refiner=False).chunks(SDXL_BASE_EXPECTED)
+    assert len(chunked) == 1
+    c = chunked[0]
+    assert c.time_ids is not None and c.time_ids.shape == (2, 6)
+    assert c.text_embeds is not None and c.text_embeds.shape == (2, 1280)
+
+
+def test_chunks_sdxl_refiner_uses_len5_time_ids():
+    chunked = _sdxl_inputs(batch=1, refiner=True).chunks(SDXL_REFINER_EXPECTED)
+    assert chunked[0].time_ids.shape == (2, 5)
+
+
+def test_chunks_sdxl_synthesizes_zero_time_ids_when_caller_omits():
+    """If the model expects time_ids but caller passed nothing, the suite
+    fabricates a zero-filled tensor. Lock that fallback."""
+    x = torch.randn(1, 4, 128, 128)
+    t = torch.full((1,), 999.0)
+    context = torch.randn(1, 77, 2048)
+    inputs = CoreMLInputs(x, t, context, control=None)
+    chunked = inputs.chunks(SDXL_BASE_EXPECTED)
+    assert chunked[0].time_ids.shape == (2, 6)
+    assert torch.equal(chunked[0].time_ids, torch.zeros(2, 6))
+    assert chunked[0].text_embeds.shape == (2, 1280)
+    assert torch.equal(chunked[0].text_embeds, torch.zeros(2, 1280))
+
+
+def test_chunks_splits_batch_into_multiple_target2_chunks():
+    """batch=5 with target_batch=2 -> 3 chunks (last padded)."""
+    chunked = _sd15_inputs(batch=5).chunks(SD15_EXPECTED)
+    assert len(chunked) == 3
+    for c in chunked:
+        assert c.x.shape == (2, 4, 64, 64)
+        assert c.context.shape == (2, 77, 768)
+    # Last chunk's second batch row is the zero-pad.
+    assert torch.equal(chunked[-1].x[1], torch.zeros(4, 64, 64))
+
+
+def test_chunks_timestep_is_broadcast_from_first_value():
+    """t is rebuilt from t[0] across all chunks: locks current behavior that
+    discards any per-row timestep variation."""
+    x = torch.randn(2, 4, 64, 64)
+    t = torch.tensor([42.0, 99.0])  # the second value will be lost
+    context = torch.randn(2, 77, 768)
+    inputs = CoreMLInputs(x, t, context, control=None)
+    chunked = inputs.chunks(SD15_EXPECTED)
+    assert chunked[0].t.shape == (2,)
+    assert torch.equal(chunked[0].t, torch.full((2,), 42.0))

tests/unit/test_characterization_latents.py

@@ -0,0 +1,118 @@
+"""Characterization tests for coreml_suite.latents.
+
+Locks the *current* behavior of chunk_batch / merge_chunks — including the
+zero-pad regions and the truncation in merge — so a refactor
+cannot silently shift either contract.
+"""
+import pytest
+import torch
+
+from coreml_suite.core.latents import chunk_batch, merge_chunks
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+
+
+def _const_tensor(batch, *rest):
+    return torch.arange(batch * 4 * 8 * 8, dtype=torch.float32).reshape(batch, 4, 8, 8)
+
+
+# ---------- chunk_batch ------------------------------------------------------
+
+
+def test_chunk_batch_passthrough_when_shape_matches():
+    x = _const_tensor(2)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 1
+    # passthrough: the same object identity is returned (no copy).
+    assert out[0] is x
+
+
+def test_chunk_batch_pads_single_chunk_when_input_smaller():
+    """batch=1, target=2 -> one padded chunk; the second row is exact zero."""
+    x = _const_tensor(1)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 1
+    assert out[0].shape == (2, 4, 8, 8)
+    assert torch.equal(out[0][0], x[0])
+    assert torch.equal(out[0][1], torch.zeros(4, 8, 8))
+
+
+def test_chunk_batch_splits_exact_multiple():
+    """batch=4, target=2 -> two chunks, no padding."""
+    x = _const_tensor(4)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 2
+    assert out[0].shape == (2, 4, 8, 8)
+    assert out[1].shape == (2, 4, 8, 8)
+    assert torch.equal(out[0], x[:2])
+    assert torch.equal(out[1], x[2:])
+
+
+def test_chunk_batch_pads_remainder_chunk():
+    """batch=5, target=2 -> chunks=[x[0:2], x[2:4]] then [x[4], 0]."""
+    x = _const_tensor(5)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 3
+    assert torch.equal(out[0], x[0:2])
+    assert torch.equal(out[1], x[2:4])
+    last = out[-1]
+    assert last.shape == (2, 4, 8, 8)
+    assert torch.equal(last[0], x[4])
+    # The remainder row is zero-padded; lock that exact contract.
+    assert torch.equal(last[1], torch.zeros(4, 8, 8))
+    assert last[1].sum() == 0
+
+
+@pytest.mark.parametrize(
+    "batch_size,target,expected_chunks",
+    [
+        (1, 4, 1),
+        (3, 2, 2),
+        (5, 3, 2),
+        (9, 4, 3),
+    ],
+)
+def test_chunk_batch_pad_region_is_zero(batch_size, target, expected_chunks):
+    x = _const_tensor(batch_size)
+    out = chunk_batch(x, (target, 4, 8, 8))
+    assert len(out) == expected_chunks
+    mod = batch_size % target
+    if mod == 0 and batch_size >= target:
+        return
+    last = out[-1]
+    pad_rows = target - (mod if (mod != 0 and batch_size >= target) else batch_size)
+    pad_region = last[-pad_rows:]
+    assert torch.equal(pad_region, torch.zeros_like(pad_region))
+
+
+# ---------- merge_chunks -----------------------------------------------------
+
+
+def test_merge_chunks_exact_concat():
+    x = _const_tensor(4)
+    chunks = chunk_batch(x, (2, 4, 8, 8))
+    merged = merge_chunks(chunks, x.shape)
+    assert merged.shape == x.shape
+    assert torch.equal(merged, x)
+
+
+def test_merge_chunks_truncates_padding():
+    """Round-trip with a padded last chunk drops the pad rows."""
+    x = _const_tensor(5)
+    chunks = chunk_batch(x, (2, 4, 8, 8))
+    merged = merge_chunks(chunks, x.shape)
+    assert merged.shape == x.shape
+    assert torch.equal(merged, x)
+
+
+def test_merge_chunks_singleton_returns_equal_copy_when_shape_matches():
+    """A singleton chunk list still goes through torch.cat, so we get a new
+    tensor equal to the input — locked here because a refactor might be tempted
+    to short-circuit and accidentally return the same object."""
+    x = _const_tensor(2)
+    out = merge_chunks([x], x.shape)
+    assert torch.equal(out, x)
+    assert out is not x

tests/unit/test_characterization_sdxl_options.py

@@ -0,0 +1,127 @@
+"""Characterization tests for the SDXL options math.
+
+The SDXL time_ids / text_embeds math lives in
+coreml_suite.core.sdxl as pure builders. The framework adapter
+add_sdxl_model_options lives in models.py; here we just lock the pure
+math.
+"""
+import inspect
+
+import pytest
+import torch
+
+from coreml_suite.core.sdxl import (
+    build_sdxl_text_embeds,
+    build_sdxl_time_ids,
+    sdxl_model_function_wrapper,
+)
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+
+
+# ---------- build_sdxl_time_ids: base (len 6) -------------------------------
+
+
+def test_build_time_ids_base_defaults():
+    out = build_sdxl_time_ids({}, {}, is_base=True, is_refiner=False)
+    expected = torch.tensor([[768, 768, 0, 0, 768, 768], [768, 768, 0, 0, 768, 768]])
+    assert out.shape == (2, 6)
+    assert torch.equal(out, expected)
+
+
+def test_build_time_ids_base_respects_overrides():
+    pos = {"height": 1024, "width": 512, "crop_h": 8, "crop_w": 4,
+           "target_height": 1024, "target_width": 1024}
+    neg = {"height": 256, "width": 256, "crop_h": 0, "crop_w": 0,
+           "target_height": 256, "target_width": 256}
+    out = build_sdxl_time_ids(pos, neg, is_base=True, is_refiner=False)
+    expected = torch.tensor([[1024, 512, 8, 4, 1024, 1024], [256, 256, 0, 0, 256, 256]])
+    assert torch.equal(out, expected)
+
+
+# ---------- build_sdxl_time_ids: refiner (len 5) ----------------------------
+
+
+def test_build_time_ids_refiner_defaults():
+    out = build_sdxl_time_ids({}, {}, is_base=False, is_refiner=True)
+    expected = torch.tensor([[768, 768, 0, 0, 6.0], [768, 768, 0, 0, 2.5]])
+    assert out.shape == (2, 5)
+    assert torch.equal(out, expected)
+
+
+def test_build_time_ids_refiner_respects_aesthetic_score():
+    pos = {"aesthetic_score": 8.5}
+    neg = {"aesthetic_score": 1.5}
+    out = build_sdxl_time_ids(pos, neg, is_base=False, is_refiner=True)
+    expected = torch.tensor([[768, 768, 0, 0, 8.5], [768, 768, 0, 0, 1.5]])
+    assert torch.equal(out, expected)
+
+
+# ---------- build_sdxl_time_ids: edge case ----------------------------------
+
+
+def test_build_time_ids_neither_base_nor_refiner_returns_len4():
+    out = build_sdxl_time_ids({}, {}, is_base=False, is_refiner=False)
+    assert out.shape == (2, 4)
+
+
+# ---------- build_sdxl_text_embeds ------------------------------------------
+
+
+def test_text_embeds_concat_pos_then_neg():
+    pos = torch.full((1, 1280), 1.0)
+    neg = torch.full((1, 1280), -1.0)
+    out = build_sdxl_text_embeds(pos, neg)
+    assert out.shape == (2, 1280)
+    assert torch.equal(out[0], pos[0])
+    assert torch.equal(out[1], neg[0])
+
+
+# ---------- sdxl_model_function_wrapper closure -----------------------------
+
+
+def test_wrapper_captures_time_ids_text_embeds_refiner_via_closure():
+    time_ids = torch.zeros(2, 6)
+    text_embeds = torch.zeros(2, 1280)
+    wrapper = sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False)
+    closure = inspect.getclosurevars(wrapper).nonlocals
+    assert closure["time_ids"] is time_ids
+    assert closure["text_embeds"] is text_embeds
+    assert closure["refiner"] is False
+
+
+def test_wrapper_returns_zero_when_context_missing():
+    """When c_crossattn is None the wrapper short-circuits to zeros_like(x).
+    Locked here because the refactor mustn't change this default."""
+    wrapper = sdxl_model_function_wrapper(torch.zeros(2, 6), torch.zeros(2, 1280))
+    x = torch.randn(2, 4, 16, 16)
+    out = wrapper(
+        model_function=lambda *a, **kw: pytest.fail("model_function must not run"),
+        params={"input": x, "timestep": torch.zeros(2), "c": {}},
+    )
+    assert torch.equal(out, torch.zeros_like(x))
+
+
+def test_wrapper_refiner_truncates_context_to_g_clip():
+    """refiner=True slices c_crossattn[:, :, 768:] before forwarding."""
+    captured = {}
+
+    def fake_model(x, t, **c):
+        captured["context_shape"] = c["c_crossattn"].shape
+        captured["time_ids_shape"] = c["time_ids"].shape
+        return x
+
+    wrapper = sdxl_model_function_wrapper(
+        torch.zeros(2, 5), torch.zeros(2, 1280), refiner=True
+    )
+    x = torch.randn(2, 4, 16, 16)
+    context = torch.randn(2, 77, 2048)  # 768 + 1280 dims
+    wrapper(
+        model_function=fake_model,
+        params={"input": x, "timestep": torch.zeros(2), "c": {"c_crossattn": context}},
+    )
+    assert captured["context_shape"] == (2, 77, 1280)
+    assert captured["time_ids_shape"] == (2, 5)

tests/unit/test_chunks.py

@@ -0,0 +1,122 @@
+"""Smoke tests for the pure batch-chunking helpers in coreml_suite.core.
+
+Uses torch.device('cpu') instead of comfy.model_management.get_torch_device
+so Tier 0 runs without ComfyUI.
+"""
+import pytest
+import torch
+
+from coreml_suite.core.controlnet import chunk_control
+from coreml_suite.core.inputs import CoreMLInputs
+from coreml_suite.core.latents import chunk_batch, merge_chunks
+
+
+CPU = torch.device("cpu")
+
+
+@pytest.fixture
+def expected_inputs():
+    return {
+        "sample": {"shape": (2, 4, 64, 64)},
+        "timestep": {"shape": (2,)},
+        "timestep_cond": {"shape": (2, 256)},
+        "encoder_hidden_states": {"shape": (2, 77, 768)},
+        "additional_residual_0": {"shape": (2, 320, 64, 64)},
+        "additional_residual_1": {"shape": (2, 640, 32, 32)},
+    }
+
+
+@pytest.mark.parametrize("batch_size", [1, 2, 4, 5, 9])
+def test_batch_chunking(batch_size):
+    latent_image = torch.randn(batch_size, 4, 64, 64).to(CPU)
+    target_shape = (4, 4, 64, 64)
+
+    chunked = chunk_batch(latent_image, target_shape)
+
+    for chunk in chunked:
+        assert chunk.shape == target_shape
+
+    if batch_size % target_shape[0] != 0:
+        assert chunked[-1][batch_size % target_shape[0] :].sum() == 0
+
+
+@pytest.mark.parametrize("batch_size", [1, 2, 4, 5, 9])
+def test_merge_chunks(batch_size):
+    input_tensor = torch.randn(batch_size, 4, 64, 64).to(CPU)
+    target_shape = (4, 4, 64, 64)
+    chunked = chunk_batch(input_tensor, target_shape)
+
+    merged = merge_chunks(chunked, input_tensor.shape)
+
+    assert merged.shape == input_tensor.shape
+    assert torch.equal(input_tensor, merged)
+
+
+@pytest.fixture
+def inputs():
+    x = torch.randn(1, 4, 64, 64).to(CPU)
+    t = torch.randn([1]).to(CPU)
+    c_crossattn = torch.randn(1, 77, 768).to(CPU)
+    control = {
+        "output": [
+            torch.randn(1, 320, 64, 64).to(CPU),
+            torch.randn(1, 640, 32, 32).to(CPU),
+        ],
+    }
+    timestep_cond = torch.randn(1, 256).to(CPU)
+
+    return CoreMLInputs(x, t, c_crossattn, control, timestep_cond=timestep_cond)
+
+
+@pytest.mark.parametrize(
+    "b, target_size, num_chunks",
+    [
+        (1, 2, 1),
+        (1, 1, 1),
+        (2, 2, 1),
+        (3, 2, 2),
+        (4, 2, 2),
+        (5, 3, 2),
+        (9, 4, 3),
+    ],
+)
+def test_chunking_controlnet(b, target_size, num_chunks):
+    cn = {
+        "output": [
+            torch.randn(b, 320, 64, 64).to(CPU),
+            torch.randn(b, 640, 32, 32).to(CPU),
+        ],
+        "middle": [
+            torch.randn(b, 1280, 8, 8).to(CPU),
+        ],
+    }
+
+    chunked = chunk_control(cn, target_size)
+
+    assert len(chunked) == num_chunks
+    for chunk in chunked:
+        assert chunk["output"][0].shape == (target_size, 320, 64, 64)
+        assert chunk["output"][1].shape == (target_size, 640, 32, 32)
+        assert chunk["middle"][0].shape == (target_size, 1280, 8, 8)
+
+
+def test_chunking_no_control():
+    cn = None
+    target_size = 2
+
+    chunked = chunk_control(cn, target_size)
+
+    assert chunked == [None, None]
+
+
+def test_chunking_inputs(expected_inputs, inputs):
+    chunked = inputs.chunks(expected_inputs)
+
+    assert len(chunked) == 1
+
+    assert chunked[0].x.shape == (2, 4, 64, 64)
+    assert chunked[0].t.shape == (2,)
+    assert chunked[0].context.shape == (2, 77, 768)
+    assert chunked[0].control["output"][0].shape == (2, 320, 64, 64)
+    assert chunked[0].control["output"][1].shape == (2, 640, 32, 32)
+    assert chunked[0].ts_cond.shape == (2, 256)

tests/unit/test_controlnet.py

@@ -0,0 +1,16 @@
+from coreml_suite.controlnet import no_control
+
+
+def test_no_control():
+    expected_inputs = {
+        "additional_residual_0": {"shape": (2, 2, 2)},
+        "additional_residual_1": {"shape": (2, 4, 4)},
+        "additional_residual_2": {"shape": (2, 8, 8)},
+    }
+
+    residual_kwargs = no_control(expected_inputs)
+
+    assert len(residual_kwargs) == 3
+    assert residual_kwargs["additional_residual_0"].shape == (2, 2, 2)
+    assert residual_kwargs["additional_residual_1"].shape == (2, 4, 4)
+    assert residual_kwargs["additional_residual_2"].shape == (2, 8, 8)

tests/unit/test_tier0_purity.py

@@ -0,0 +1,43 @@
+"""Gate: prove the Tier-0 lane is framework-free.
+
+In a pure `pytest -m unit` run, none of the banned runtime modules
+(comfy, coremltools, python_coreml_stable_diffusion, folder_paths,
+nodes, comfy_extras, diffusers, diffusionkit) may be in sys.modules
+after collection. If they are, a tests/unit/ file is transitively
+pulling them in and the Tier-0 promise — "runs on Linux with no Mac
+stack" — is broken.
+
+When other tiers are also collected, framework modules may be imported
+deliberately (e.g. smoke pulls in coremltools), so the check is skipped
+unless the run is purely `-m unit` — Tier-0 purity is only meaningful
+when nothing else is loaded.
+"""
+import sys
+
+import pytest
+
+BANNED_ROOTS = {
+    "comfy",
+    "comfy_extras",
+    "coremltools",
+    "python_coreml_stable_diffusion",
+    "folder_paths",
+    "nodes",
+    "diffusers",
+    "diffusionkit",
+}
+
+
+def test_no_framework_modules_loaded_by_unit_tier(request):
+    markexpr = request.config.option.markexpr
+    if markexpr != "unit":
+        pytest.skip(
+            "purity gate only meaningful in a pure `-m unit` run "
+            f"(got markexpr={markexpr!r}); other tiers are expected to "
+            "import comfy/coremltools."
+        )
+    loaded = {name for name in sys.modules if name.split(".")[0] in BANNED_ROOTS}
+    assert not loaded, (
+        f"Tier-0 leakage: these framework modules are in sys.modules after "
+        f"collecting tests/unit/: {sorted(loaded)}. Pure-core promise broken."
+    )