Removing a background sounds simple until your subject has hair, fur, or glass in it. Then it stops being simple very fast.
What we tested
We benchmarked five matting models against a fixed set of tricky inputs — backlit hair, motion-blurred limbs, transparent objects — and scored them on edge accuracy, alpha softness, and time-to-result.
Edge cases that broke most models
- Wispy hair against a similarly-colored background
- Glass and translucent fabric
- Tight crops where the subject runs off the frame
What we shipped
BiRefNet won on quality-per-second by a clear margin. It's now the model behind the Background Remover, and it's the reason that tool only costs 1 credit — fast enough to keep a tight margin.
