How the FeyNoBg model removes difficult backgrounds

FeyNoBg is an automatic background removal model created by Feyn Labs. It predicts an alpha value for every pixel so the background becomes transparent while the subject and partially transparent boundary pixels remain.

Recognition and boundary precision

The model must solve two related problems: recognize the whole foreground subject and reconstruct its edges. The second problem becomes especially difficult around hair, fur, thin wires, motion blur, and low-contrast boundaries.

BiRefNet foundation

FeyNoBg builds on BiRefNet and expands the third feature-extraction stage from 18 to 24 blocks, growing from 222 million to 263 million parameters. Compatible pretrained weights were retained while the six new blocks learned additional capacity.

Diverse training

The reported final training mix contained 26,100 images from ten sources covering crowded scenes, camouflage, high-resolution subjects, portraits, and anime. Each source was capped to reduce domination by the largest datasets.

Reported benchmark results

Feyn Labs reports that FeyNoBg led four of eight evaluated benchmarks by S-measure and remained within two percent of the published leader on the other four. Benchmark performance is useful evidence, but your own image types remain the best practical test.

Provenance

This independent guide summarizes public model information. For research details and source code, consult the original Feyn Labs publication, the FeyNoBg model card, and the NoBg repository.