BRIA AI launched RMBG-3.0 background remover based on Fibo 2.0.
Unlike classic contour segments, the third generation is built for the first time on the generative foundation of Fibo 2.0 and refined to work in 4 steps.
That's not really what the news is about. I fixed the effect of black photos and videos in the cart on the iPhone.
- If it's normal now.
- Everything is black.
- I have an Android, everything is fine, hehe.
Let's get back to the topic:
Because the model understands the semantics of the scene, it correctly preserves all people in group shots, blurred limbs, small logos and labels on goods, while simultaneously cleaning gaps in meshes, bicycle spokes and casting drop shadows. Inference works without CFG and takes about 6 seconds for a full resolution image on one NVIDIA L40S.
In a blind test on 234 challenging images, RMBG-3.0 took first place with an average rating of 1.51, beating Photoroom (55.1% win rate), Kanba's Remove.bg (59.4%), RMBG-2.0 (60.9%), Ideograms (65.6%) and BiRefNet (75.9%).
Sources:
Weights on a hugging face (healthy/RMBG-3.0).
demo remove-bg. Bria.ai.
Launch at fal.ai.