BRIA AI just released RMBG-3.0, a new generative AI background removal model built on Fibo 2.0.
And according to BRIA's own benchmark, it outperforms Photoroom, remove.bg, Ideogram, BiRefNet, and the previous RMBG-2.0 on challenging images.
But the interesting part isn't just the benchmark.
It's the architecture.
Traditional background removal models learn to identify objects and separate their boundaries from the surrounding image.
RMBG-3.0 takes a different approach.
Instead of relying on a conventional segmentation architecture, it uses a generative image-editing foundation model to produce an alpha matte.
The goal is to understand what belongs to the subject, not just where its visible edges appear.
WHAT'S NEW IN RMBG-3.0?
• Fibo 2.0 foundation: an 8-billion-parameter generative model adapted specifically for background removal.
• Four-step generation: distilled to produce alpha mattes in just four inference steps.
• No classifier-free guidance: optimized for a simpler, faster inference process.
• Better group photos: aims to preserve every person instead of accidentally removing secondary subjects.
• Motion blur handling: improved treatment of blurred hands, arms, and moving objects.
• Fine details: better preservation of hair, fur, small logos, product labels, and complex edges.
• Open structures: handles gaps between bicycle spokes, fences, meshes, and other intricate objects.
• Shadow separation: distinguishes the main subject from unwanted cast shadows.
• Alpha transparency: produces soft masks with partially transparent edges rather than only hard cutouts.
• Licensed training data: BRIA says Fibo models are trained exclusively on licensed content.
THE BENCHMARK RESULTS
BRIA evaluated RMBG-3.0 against five competing background removal systems on 234 difficult photographs.
Reported results:
• Average rank: 1.51
• Best or tied-best result: 185 of 234 images
• Preferred over Photoroom: 55.1%
• Preferred over remove.bg: 59.4%
• Preferred over RMBG-2.0: 60.9%
• Preferred over Ideogram: 65.6%
• Preferred over BiRefNet: 75.9%
These figures come from BRIA's internal evaluation, not an independent public benchmark.
HOW FAST IS IT?
RMBG-3.0 runs in approximately six seconds per image on a single NVIDIA L40S at around one megapixel.
It uses four diffusion-style generation steps and requires substantial GPU memory, approximately 32 GB or more for the full pipeline.
That's an important trade-off.
RMBG-3.0 prioritizes difficult, high-quality cutouts over the lightweight, low-latency inference offered by older segmentation models.
RMBG-2.0 remains a more practical option for smaller GPUs and high-volume processing where speed matters most.
WHAT ABOUT THE LIMITATIONS?
RMBG-3.0 isn't perfect.
BRIA reports that it can still struggle with genuinely transparent materials such as glass, smoke, veils, and fine mesh.
Busy backgrounds and low-contrast objects can also confuse the model.
And in five benchmark examples, RMBG-3.0 ranked last.
So this isn't a universal replacement for every background removal tool.
But for product photography, advertising, ecommerce catalogs, and demanding image editing workflows, the improved semantic understanding could be genuinely useful.
OPEN WEIGHTS AND COMMERCIAL USE
The model weights are available on Hugging Face for noncommercial use under restricted licensing terms.
Commercial self-hosting requires a separate agreement with BRIA, while commercially licensed inference is available through the BRIA API and fal.ai.
TRY RMBG-3.0
Free browser demo:
https://remove-bg.bria.ai/
Model weights on Hugging Face:
https://huggingface.co/briaai/RMBG-3.0
Run through fal.ai:
https://fal.ai/models/bria/rmbg-3
Official BRIA website:
https://bria.ai/rmbg
Would you choose a slower generative background remover for better edge quality, or stick with a faster traditional segmentation model?
#RMBG3 #BRIAAI #BackgroundRemoval
YouTube / ABVX
