040b30733d
* feat(matte): first-party mask-guided one-step video matting nodes
Adds matte/ — an independent implementation of the published VideoMaMa
one-step SVD method (matte/pipeline.py) plus Koolook's temporal windowing,
subject-focused crop/stitch, resolution control (/64, shortest-side) and
auto OOM resolution-backoff. Nodes: KoolookMatteLoader, KoolookMatteSampler,
KoolookMatteFocusCrop, KoolookMatteFocusStitch.
Registered via the existing guarded _merge_node_group("Koolook Matte",
".matte"), so a missing diffusers disables only these nodes. diffusers is
already pulled in by the radiance/whatdreamscost forks, so no new dependency.
Loads (never bundles) the VideoMaMa UNet (CC-BY-NC) + SVD VAE at runtime;
matte/NOTICE records the model licenses and user responsibility.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore(dev-sync): include matte/ in RUNTIME_PATHS
So `sync_to_dev` copies the new matte node package into the live ComfyUI
install alongside the rest of the koolook pack.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(matte): raw cond latents; fp16 default; preserve checkpoints on dev-sync
pipeline.py: _encode_latents must NOT divide by vae scaling_factor. The
VideoMaMa UNet is trained on raw VAE latents (matching okdalto's
VideoInferencePipeline, which multiplies by scaling_factor then divides it
back out -> net x1). Dividing over-scaled the conditioning ~5.5x and produced
a systematically blurry matte on every frame. Verified 5/5 vs the desktop /
vanilla reference on identical single-frame inputs.
nodes.py: harden the loader precision fallback so fp16 is the true default -
only an explicit "bf16" opts out; a stale "auto" widget now resolves to fp16
instead of bf16.
sync_to_dev.py: the rmtree+copytree mirror deleted matte/checkpoints/ (the
gitignored ~10GB SVD+VideoMaMa weights) on every matte sync, forcing a
re-download. Add PRESERVE_IN_DEST + selective _clean_dest so runtime-
downloaded model data survives a sync.
docs: add docs/maintainers/matte-pipeline.md (latent-scaling invariant,
black-output causes, fidelity levers, single-frame + vanilla-oracle debug
loop, checkpoint preservation).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(matte): address PR #269 review
- license: align matte to the repo's GPL-3.0 (was self-contradictory: claimed
"MIT licensed - see LICENSE" while LICENSE is GPL-3.0). pipeline.py header,
NOTICE, and the maintainer doc updated. NOTICE still credits VideoMaMa
(CC BY-NC) + SVD (Stability Community License).
- nodes.py _ensure_2d_mask: squeeze first, then handle [H,W,C]/[C,H,W]
explicitly and RAISE on ambiguous shapes -- previously [H,W,3] silently
became [W,3] (wrong geometry).
- nodes.py OOM backoff: catch RuntimeError (OutOfMemoryError subclasses it, but
cuBLAS/cuDNN/older paths raise plain "RuntimeError: CUDA out of memory"); the
existing message guard still filters non-OOM errors.
- nodes.py: warn when the guide mask is empty instead of silently falling back
to full-frame.
- sync_to_dev.py: PRESERVE_IN_DEST is now an exact relative path
("matte/checkpoints"), not a bare name matched at any depth; handle symlinks
before is_dir() in _clean_dest so a symlinked dir isn't recursed/rmtree'd.
- CHANGELOG: add the Koolook Matte nodes under [Unreleased] with the
non-commercial model-license caveat.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(matte): address PR #269 re-review
- sync_to_dev.py [BLOCKING]: a top-level symlinked dest (e.g. target/matte -> X)
was passed straight into _clean_dest()/copytree(), which follow the link and
can mutate files outside the target. Now unlink a symlinked dest and own the
path as a real dir (checked before exists(), which follows symlinks). Verified
with a repro: the symlink target's file survives.
- nodes.py: force float32 on a CPU fallback (CUDA unavailable). Half precision
is unreliable on CPU; fp16 kernels would error. Warn that CPU is very slow.
- docs: correct the preserved-path guidance to the target-relative
"matte/checkpoints" (was a bare "checkpoints" name); reword "near-verbatim
reimplementation" -> "independent implementation targeting behavioural parity"
to match the NOTICE/pipeline framing.
- nodes.py: also warn on a degenerate (<8px) guide-mask region before the
full-frame fallback (was silent).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(matte): address PR #269 re-review (CI red + FocusCrop frame check)
- nodes.py: KoolookMatteFocusCrop now validates images vs coarse_mask frame
counts (raises like KoolookMatteSampler.run_inference). A mismatch previously
made cropped_images/cropped_mask diverge and silently drop/misalign frames at
FocusStitch.
- pipeline.py: split "vae.to('cpu'); _clear_cache(...)" onto two lines (Ruff
E702 was failing Lint & Smoke).
- constraints-test.txt: bump pillow 12.2.0 -> 12.3.0 (pip-audit CVE, fixed in
12.3.0). Pillow is a leaf dep so the closure is unaffected; CI re-audits.
- README: note the Koolook Matte nodes need diffusers (already via the LTX /
Radiance forks) and huggingface_hub for runtime weight downloads.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: AI Machine <ai.machine@x.baconcph.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
36 lines
1.5 KiB
Plaintext
36 lines
1.5 KiB
Plaintext
Koolook Matte — NOTICE
|
|
======================
|
|
|
|
The CODE in this pack is licensed under GPL-3.0 (see the repository LICENSE) and
|
|
is an ORIGINAL, independent implementation — it does not incorporate third-party
|
|
wrapper code.
|
|
|
|
Method credit
|
|
-------------
|
|
This pack implements the method described in:
|
|
"VideoMaMa: Mask-Guided Video Matting via Generative Prior" (CVPR 2026),
|
|
cvlab-kaist — https://github.com/cvlab-kaist/VideoMaMa
|
|
The method, the original reference code, and the trained weights are the work of
|
|
those authors. This project re-implements the (published) inference method using
|
|
the standard `diffusers` API and adds its own windowing / focus-crop tooling.
|
|
|
|
Model weights (NOT bundled or redistributed — loaded at runtime)
|
|
----------------------------------------------------------------
|
|
* VideoMaMa fine-tuned UNet — **CC BY-NC 4.0 (NON-COMMERCIAL)**.
|
|
Source: `SammyLim/VideoMaMa` (Hugging Face) / cvlab-kaist/VideoMaMa.
|
|
==> Do NOT use these weights for commercial work without a separate license
|
|
from the model authors.
|
|
* Stable Video Diffusion img2vid-xt (its temporal VAE is used) —
|
|
**Stability AI Community License**.
|
|
Source: `stabilityai/stable-video-diffusion-img2vid-xt`.
|
|
|
|
You (the user) are solely responsible for complying with the above model
|
|
licenses. This pack downloading/loading a model does not grant you any rights
|
|
to it.
|
|
|
|
Technique note
|
|
--------------
|
|
The temporal sliding-window batching and the subject bounding-box crop/paste
|
|
implemented here are standard, well-known techniques, written independently for
|
|
this pack.
|