postprocess at model res, defer resize+write to CPU (saves ~35s GPU)

compact UI: resolution+screen color in row, despeckle in row

fix: remove tree paths from credit links

pin upstream commit IDs in credits

fix: blue CPU is available (1024 only), not green-only

update all references: green → green/blue

status shows CPU + GPU time breakdown

add blue 1024 ONNX (FP16 on disk, FP32 at runtime), rename models

blue screen support: GPU-only, auto-detection, no ONNX/LFS

safetensors loading, Phase 0 4x faster (uint8), total time in status

quality: lower clean_matte threshold 0.25→0.02, always keep largest component

cleanup: stale comments, dead import, redundant makedirs, fix batch size in UI

simplify: merge write functions, fix missing Processed output, bulk transfer

remove dead code: AOTI export, inductor/triton cache, shared_results, deferred write

disable torch.compile on ZeroGPU — net negative for GreenFormer

fix: reduce-overhead instead of max-autotune (118s→~30s), dedicated export endpoint

triple-fallback artifact export + triton cache

export .pt2 after compile + download button

always compile with triton (max-autotune) + probe AOTI

simplify pipeline timing in README

enable triton+CUDA graphs via torch.compile(max-autotune)

fix README: accurate torch.compile description, no triton/AOTI claim

update README: fix batch size, add GPU optimization docs

fix randperm overflow at 4K: per-frame connected components

GPU postprocessing pipeline + TF32 + conditional torch.compile

add ZeroGPU GPU inference (FP16, flash-attn, batch=32@1024/16@2048)

professional output package (Comp+FG+Matte+Processed)

download button for all output modes

fix gradio cache_examples None output crash

fix example cache crash

initial commit

initial commit