Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)

BRIA releases RMBG-3.0: background removal in 6 seconds, beats BiRefNet in 75.9% of head-to-head comparisons

BRIA AI has released RMBG-3.0, a background removal model that understands what is in the frame: a person, a logo on packaging, or a shadow on the floor. As a result, it clears background out of mesh fences and bicycle spokes, while small lettering on products stays intact. A full-resolution image takes about 6 seconds on an NVIDIA L40S GPU. For a product catalog, it makes sense to run your hardest images through it.

AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
BRIA releases RMBG-3.0: background removal in 6 seconds, beats BiRefNet in 75.9% of head-to-head comparisons
Source: TG: @cgevent (Метаверсище и ИИще). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @cgevent (Метаверсище и ИИще) ↗
Editor-in-Chief
Blind test on 234 hard photosfirst place, average rank 1.51
Key metric #2Beats BiRefNet in 75.9% of comparisons and remove.bg in 59.4%
Key metric #3Runs in 4 steps without CFG; weights are open on Hugging Face

Automatic background removal looks like a solved problem until a hard image comes along: a chain-link fence, a bicycle with spokes, a group photo, or a jar with small lettering on the label. Ordinary cutout tools leave patches of background in the gaps or erase fine text. BRIA AI has released RMBG-3.0, the third generation of its background removal model, which first "understands" what is in the frame and only then cuts. It takes about 6 seconds per full-resolution image on a single NVIDIA L40S GPU. In a blind test on 234 hard photos, it took first place with an average rank of 1.51 (the closer to one, the more often the model came out best).

What changed under the hood

Earlier models were segmenters: they looked for the boundary between object and background and cut along the outline. RMBG-3.0 is the first to be built on a generative base, Fibo 2.0, a model that can create images and therefore has a better grasp of how a scene is put together. BRIA then distilled it, compressing the large model into a lighter version that needs only 4 steps instead of dozens. Inference also runs without CFG, a mode that in generative models usually requires extra computation at every step.

Because the model understands scene semantics, BRIA says it keeps every person in group shots, does not crop motion-blurred limbs, leaves small logos and product labels in place, cleans out the gaps in mesh fences and between bicycle spokes, and discards cast shadows.

What the blind test showed

In the blind comparison on 234 hard photos, RMBG-3.0 took first place. The new model's win rates in pairwise comparisons are as follows:

  • against Photoroom: 55.1%;
  • against remove.bg from Canva: 59.4%;
  • against the previous RMBG-2.0: 60.9%;
  • against Ideogram: 65.6%;
  • against the open BiRefNet: 75.9%.

The lead over commercial services is more modest than over open models. Against Photoroom, the advantage is just over half of cases, which is a slight edge rather than a rout. The new model beats its predecessor RMBG-2.0 by a wider margin, though, which is an argument for those still on the previous generation.

What to do in practice

The weights are published on Hugging Face (briaai/RMBG-3.0), there is an online demo at remove-bg.bria.ai, and you can run the model without your own server through fal.ai. That lets you test it on your own problem photos first and only then decide whether you need your own infrastructure.

A few caveats. The original post does not say who exactly ran the test, and 234 photos is a small sample, so results on your data may differ. Read the license terms for commercial use on the model page before deploying. And a note on cost: at high volume, 6 seconds per image on an L40S adds up to a noticeable bill, so it makes sense to keep simple photos with a plain background on a faster, cheaper tool.

A practical approach for an online store team: collect 50–100 of the nastiest images from your own pipeline (mesh, transparent details, labels, group shots), run them through both RMBG-3.0 and your current service, and compare the results by eye. If the improvement is visible, it makes sense to send at least the hard cases to the new model.

Story

Key Questions Answered

What changed under the hood?

Earlier models were segmenters: they looked for the boundary between object and background and cut along the outline. RMBG-3.0 is the first to be built on a generative base, Fibo 2.0, a model that can create images and therefore has a better grasp of how a scene is put together. BRIA then distilled it, compressing the large model into a lighter version that needs only 4 steps instead of dozens. Inference also runs without CFG, a mode that in generative models usually requires extra computation at every step.

What the blind test showed?

In the blind comparison on 234 hard photos, RMBG-3.0 took first place. The new model's win rates in pairwise comparisons are as follows: against Photoroom: 55.1%; against remove.bg from Canva: 59.4%; against the previous RMBG-2.0: 60.9%; against Ideogram: 65.6%; against the open BiRefNet: 75.9%. The lead over commercial services is more modest than over open models. Against Photoroom, the advantage is just over half of cases, which is a slight edge rather than a rout.

What to do in practice?

The weights are published on Hugging Face (briaai/RMBG-3.0), there is an online demo at remove-bg.bria.ai, and you can run the model without your own server through fal.ai. That lets you test it on your own problem photos first and only then decide whether you need your own infrastructure. A few caveats. The original post does not say who exactly ran the test, and 234 photos is a small sample, so results on your data may differ. Read the license terms for commercial use on the model page before deploying.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…