FeyNoBg: A SOTA Model For Background Removal
Summary
Feyn presents FeyNoBg, a 263-million-parameter background-removal model that leads four of eight reported benchmarks and comes within two percent of the leader on the rest.
Treat recognition and boundary precision as coupled skills.
Good cutouts require both identifying the whole subject and tracing translucent or fine edges, so optimizing either skill alone can degrade the other.
Across eight benchmarks, it posts the best published S-measure on four and comes within 2% of the leader on the rest.
Add capacity without discarding prior learning.
FeyNoBg expands BiRefNet’s most information-rich feature stage from 18 to 24 blocks while retaining every compatible pretrained weight.
We preserved every compatible pre-trained weight during the expansion. Only the six new blocks started untrained.
Balance the data mixture deliberately.
Per-source caps prevent large datasets from dominating, while varied scenes teach the model to handle camouflage, crowds, portraits, high resolution, and fine structures.
Normalize unlike annotations to one target.
Converting masks and alpha mattes into binary foreground masks creates consistent supervision, though it gives up soft-opacity training signals.
A poor training mix can produce unbalanced models where improvements in one skill come at the expense of the other.
The matting datasets therefore contributed precisely outlined subjects, not soft-opacity supervision.
Package the experimental substrate.
NoBg exposes a common interface for running, comparing, and fine-tuning background-removal models so follow-on work needs fewer bespoke adapters.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.