LibreFeyNobgl-matte-fp8

FeyNobg background removal, repackaged for LibreYOLO's matte task. Predicts a soft alpha matte at a fixed native 1024x1024.

This repo hosts the fp8 (E4M3 weights, calibrated static scales, fp16 remainder; on Ada/Hopper/Blackwell GPUs the Linear layers execute natively on the fp8 tensor cores via torch._scaled_mm, matching the fp16 checkpoint's speed at half its size - pass cuda_graph=True to predict for the full effect) post-training-quantized variant. The default-precision weights auto-download; quantized variants are opt-in: download the .pt and pass its path as the weights argument (the checkpoint's quant manifest rebuilds the quantized structure at load time).

from libreyolo import LibreYOLO

m = LibreYOLO("LibreFeyNobgl-matte-fp8.pt")
res = m.predict("product.jpg")
res[0].matte            # (H, W) float alpha in [0, 1]
res[0].save("cut.png")  # transparent-background PNG

Source

Derived from feyninc/FeyNobg (nobg library), Apache-2.0, Copyright (c) 2026 Feyn Inc. FeyNobg builds on ZhengPeng7/BiRefNet (MIT, Copyright (c) 2024 ZhengPeng).

Backbone: Swin Transformer v1, Swin-L tier with stage 3 deepened from 18 to 24 blocks (263M parameters). Training data provenance (upstream): not disclosed by Feyn Inc.; this repo redistributes the author's released weights under their Apache-2.0 grant and does not redistribute training data.

Modifications

State-dict metadata-wrap into the LibreYOLO v1.0 checkpoint schema, then post-training quantization with LibreYOLO's quantize API (fp8 (E4M3 weights, calibrated static scales, fp16 remainder; on Ada/Hopper/Blackwell GPUs the Linear layers execute natively on the fp8 tensor cores via torch._scaled_mm, matching the fp16 checkpoint's speed at half its size - pass cuda_graph=True to predict for the full effect)), stored in the packed finalized format documented in docs/quantization.md and docs/checkpoint_schema.md of the LibreYOLO source repository.

License

Apache License 2.0. See the LICENSE and NOTICE files.