lhk229 c7a2819d33 Run torch models in fp16 under inference_mode ...

- ViTMatte and BiRefNet load with torch_dtype=float16 (the BiRefNet
  checkpoint ships fp16 anyway; the old .float() upcast is gone) and
  floating-point inputs are cast to half to match.
- torch.no_grad() -> torch.inference_mode() in both predict paths.
- Outputs already downcast via .float() before .numpy(), so downstream
  stays float32. anime-seg is ONNX and unaffected.

Effect verified by the user beforehand.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-07-06 02:50:42 +08:00

..

__init__.py

Remove dead colour-model/despill code; simplify despill; refresh stale docs

2026-07-02 17:49:20 +08:00

alpha_post.py

Add cross-model veto of hair-gap background residue (default on)

2026-07-04 19:52:24 +08:00

chroma.py

Remove dead code: unused detect_background_color and ForegroundEstimate extras

2026-07-04 22:41:11 +08:00

cli.py

Add cross-model veto of hair-gap background residue (default on)

2026-07-04 19:52:24 +08:00

config.py

Add cross-model veto of hair-gap background residue (default on)

2026-07-04 19:52:24 +08:00

deps.py

Implement green screen matting pipeline

2026-06-30 14:14:15 +08:00

despill.py

Remove dead colour-model/despill code; simplify despill; refresh stale docs

2026-07-02 17:49:20 +08:00

foreground.py

Remove dead code: unused detect_background_color and ForegroundEstimate extras

2026-07-04 22:41:11 +08:00

io.py

Implement green screen matting pipeline

2026-06-30 14:14:15 +08:00

pipeline.py

Remove dead code: unused detect_background_color and ForegroundEstimate extras

2026-07-04 22:41:11 +08:00

qa.py

Implement green screen matting pipeline

2026-06-30 14:14:15 +08:00

quality_cli.py

Improve despill and add quality checks

2026-06-30 14:29:19 +08:00

quality.py

Improve despill and add quality checks

2026-06-30 14:29:19 +08:00

segmentation.py

Run torch models in fp16 under inference_mode

2026-07-06 02:50:42 +08:00

service.py

Add per-request cross_check toggle to the HTTP service (lazy-loaded)

2026-07-04 22:34:08 +08:00

settings.py

Add cross-model veto of hair-gap background residue (default on)

2026-07-04 19:52:24 +08:00

trimap.py

Add cross-model veto of hair-gap background residue (default on)

2026-07-04 19:52:24 +08:00

vitmatte_infer.py

Run torch models in fp16 under inference_mode

2026-07-06 02:50:42 +08:00

weights.py

Support loading model weights from local project folders

2026-07-03 02:50:19 +00:00