Adopts master as the standard for all overlapping work. Master's landed
optimizations supersede server-edition's own fp16 experiment:
- Unified --precision fp32|bf16 knob (bgfilter/precision.py) driving all
three models: ViTMatte weight cast + both BiRefNets via autocast, with a
hardware gate (falls back to fp32 off native-bf16 CPUs) and an
AutocastCPU fp32 shim for torchvision deform_conv2d.
- Query-chunked ViTMatte global attention (bgfilter/attn_chunk.py), exact
and bitwise-identical, caps the N^2 spike (~19 -> ~4 GB at 2048).
- Cross-check HR-matting forward reused as the seg mask (birefnet backend
only), skipping the primary seg model; MIMALLOC_PURGE_DELAY=0.
- inference_mode and the detect_background_color removal converge with
server-edition's earlier equivalents.
Conflict resolution (favoring master, preserving server-only features):
- vitmatte_infer/segmentation: dropped server's device-derived fp16 for
master's precision path, kept resolve_model_source (local weights).
- service.py: cross-check SegmentationSettings now passes precision so the
HTTP service honors bf16 like the CLI's _second_opinion does.
Verified on CPU: default fp32 pipeline loads 2 models (seg-reuse active)
and bf16 path runs (autocast + deform_conv2d shim) — both exit 0.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three optimizations from profiling the cross-check-dominated pipeline
(9700X CPU, all pilot-validated on TestImage3/FixImage1):
- Reuse the cross-check HR-matting@2048 forward as the segmentation mask
(cross_check.reuse_as_seg, default ON; --no-cross-check-as-seg to opt
out). Skips the BiRefNet@1024 load+forward entirely: ~66s -> ~45s,
one less 0.9GB model. Trimap 99.8% identical, no structural change.
- --precision bf16 now fans out to all three models: ViTMatte keeps its
weight cast; both BiRefNets run their forward under autocast with a
dispatcher-level AutocastCPU fp32 shim for torchvision::deform_conv2d
(no bf16 CPU kernel, no autocast wrapper upstream). Shared hardware
gate in bgfilter/precision.py falls back to fp32 off native-bf16
hardware. TestImage3: 51.9s -> 33.1s; alpha diff max 0.15, none >0.25.
- MIMALLOC_PURGE_DELAY=0 (bgfilter/__init__.py, before torch loads):
Windows torch's bundled mimalloc lazily retains ~10GB of freed
BiRefNet activations, stacking under ViTMatte's attention peak.
2048x2048 bf16: peak 25.2 -> 21.4GB and slightly faster (73 -> 63s).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Both call sites convert outputs to numpy immediately, so the stricter
inference-mode tensors are safe; saves autograd view/version tracking
overhead. Verified end-to-end on CUDA (alpha finite, no NaN).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Benchmarking c7a2819 on CPU showed unconditional fp16 is pathological
there: ~1870 s/image vs ~14.4 s fp32 (~130x) -- x86 has no native fp16
conv/gemm kernels so PyTorch falls back to a slow reference path. fp16
only pays off on CUDA, so ViTMatteRunner and BiRefNetSegmenter now pick
float16 iff the resolved device is cuda, float32 otherwise; inputs cast
to the same dtype. The dtype is intentionally not a setting.
Also replaces the deprecated torch_dtype= kwarg with dtype=.
Verified on CPU: single-image CLI run back to normal speed (~31 s wall
including model load) and output byte-identical (SHA256) to the
fp32-era baseline, confirming inference_mode alone changes nothing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- 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>
Add bgfilter/weights.py: resolve_model_source() maps a HuggingFace repo id
to a local folder under the weights dir (models/ by default, override with
BGFILTER_WEIGHTS_DIR) when one named after the repo basename exists; otherwise
the repo id is returned unchanged. Opt-in and backward compatible.
- vitmatte_infer.py / segmentation.py resolve model_name through it; anime-seg
reads <dir>/isnetis.onnx directly instead of hf_hub_download when local.
- .gitignore: models/, model-cache/
- README: document bundling weights as plain project folders.
Verified: anime-seg loads from a local models/anime-seg/ folder offline.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Strategic pivot: the dropout-fill rescue, the anime-seg fill backend, the
BiRefNet+anime-seg intersection, and the recolor pass all existed to repair
source images whose hair was already green-contaminated or broken at generation
time. Fixing the source instead (clean pastel-background generation) makes the
matte high-contrast, so that whole compensation layer is unnecessary. Remove it.
- Two pipelines via segmentation.enabled:
false -> chroma-only (Chroma + ViTMatte)
true -> single segmenter (default anime-seg, switchable birefnet) -> trimap
then ViTMatte refines, pymatting estimates foreground, despill cleans spill.
- Generalise the green-hardcoded colour logic to a screen_color prior (hex, e.g.
"#CFEFFF"; default null = green auto-detect, byte-identical). chroma keys off
Lab/RGB distance to the colour; despill removes chroma along its Lab direction.
Added --screen-color CLI flag.
- Removed dead code: fill_seg_dropouts + seg_fill*/intersection params,
SegmentationSettings.fill_backend/fill_model_name/sharpen, unsharp_mask,
recolor.py + RecolorSettings.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add SegmentationSettings.fill_backend / fill_model_name so the dropout-fill can
use a different (higher-recall) segmenter than the trimap. Default is now the
combination the investigation converged on: BiRefNet builds the trimap (reliable
topology -- finger gap removed, strands kept as sure_fg), anime-seg drives the
fill (more, more-coherent anime wisps). Because the fill only adds (max), a
higher-recall-but-less-reliable mask there is harmless.
On TestImage2 this eliminates the strand breakage a full anime-seg pipeline
caused (lost 699 -> 0) while keeping ~91% of its extra wisp recall (gained
1133 -> 1031). Strands stay continuous; the recovered wisps are anime-seg's
coherent curves rather than BiRefNet's grainy fill.
Also adds an optional unsharp-mask knob (segmentation.sharpen, default off).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add AnimeSegSegmenter (skytnt/anime-seg ISNet ONNX via onnxruntime, no remote
code) and a make_segmenter factory selected by SegmentationSettings.backend
("birefnet" | "anime-seg"). The ONNX output is already 0..1, so it slots into the
same soft-mask interface BiRefNetSegmenter uses.
On the anime samples anime-seg recovers more and more-coherent hair wisps than
BiRefNet (TestImage2 shoulder rescue: added px 1886 -> 3278, largest connected
component 147 -> 454), as expected from an anime-trained model. Default backend
stays birefnet.
Adds onnxruntime to requirements.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds bgfilter/segmentation.py: BiRefNetSegmenter produces a soft subject mask
that defines topology -- it includes colour-contaminated foreground (green-tinted
hair) and excludes see-through holes (finger gaps), the distinction chroma/colour
cues cannot make. Validated on both samples: green hair -> foreground, finger gap
-> background, both correct.
Foundation of the semantic-guided redesign; the mask will drive trimap
construction next, with ViTMatte still refining soft edges.
- SegmentationSettings (ZhengPeng7/BiRefNet, 1024 input, fp32).
- Deps: torchvision, timm, einops, kornia. Weights load from HuggingFace via
trust_remote_code; behind a firewall set HF_ENDPOINT=https://hf-mirror.com and
bypass the proxy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
lhk229