Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use
Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.
Bria AI has released RMBG-3.0, a background removal model that generates a soft alpha matte in 4 inference steps. It is the third generation of Bria's background removal line and the first built on a generative foundation: Fibo 2.0, fine-tuned and distilled for this single task. Weights are on Hugging Face for non-commercial use only, under the provided license. Access requires accepting a form.
Architecture and speed
According to Bria, RMBG-3.0 starts from an image-understanding model rather than an edge-focused one, so it recognizes what the subject is. The distilled model needs no classifier-free guidance and takes about 6 seconds per image on a single L40S. Like every FIBO model, it is trained exclusively on licensed data, per Bria. The output is an 8-bit alpha matte with anti-aliased edges rather than a hard binary mask.
The repository contains only the fine-tuned transformer. The text encoder, VAE and fonts come from Fibo 2.0 Edit Turbo (briaai/Fibo-2.0-edit-turbo), which has its own access form. Running it with Diffusers requires diffusers 0.42.0 or later and uses a fixed edit instruction prompt, with 4 inference steps and guidance scale 1.0.
Benchmark results (Bria's own evaluation)
Bria compared RMBG-3.0 with Photoroom, remove.bg (Canva), Ideogram (via fal), BiRefNet (via fal) and its previous RMBG-2.0 on 234 deliberately hard images, including 20 real customer images. All systems ran at default settings on original full-resolution images. Results combine blind voting by professional and casual users with a VLM side-by-side comparison. Images were ranked 1 to 6, with ties allowed.
| Model | Avg. rank (lower is better) | Best or tied-best |
|---|---|---|
| RMBG-3.0 | 1.51 | 185 / 234 |
| Photoroom | 1.87 | 162 / 234 |
| RMBG-2.0 | 2.16 | 140 / 234 |
| remove.bg (Canva) | 2.22 | 151 / 234 |
| Ideogram (via fal) | 2.76 | 113 / 234 |
| BiRefNet (via fal) | 3.56 | 78 / 234 |
In head-to-head preference (ties count as half a win), Bria reports RMBG-3.0 at 55.1% against Photoroom, 59.4% against remove.bg, 60.9% against RMBG-2.0, 65.6% against Ideogram and 75.9% against BiRefNet.
Bria says the model leads 7 of 12 hard categories and is in the top 2 in 10. It reports first place on motion blur, multiple subjects, holes and mesh, shadows and reflections, hair and fur, transparent objects and reflective surfaces. On 48 real online-store product shots, RMBG-3.0 posts an average rank of 1.15, against 1.83 for Photoroom and 1.94 for remove.bg.
The category table shows where it does not lead. Photoroom has the better average rank on thin structures (1.80 vs 2.10), smoke, fire and splashes (1.80 vs 2.10) and semi-transparent items (1.89 vs 2.67). remove.bg leads on cluttered backgrounds (1.90 vs 2.70), and RMBG-2.0 leads on low contrast (1.50 vs 2.30). On transparent objects, RMBG-3.0 has the best average rank, but Photoroom wins 55% of those images head to head because it renders glass as see-through. Each category holds only 7 to 10 images.
Access
Beyond the open weights, RMBG-3.0 is available through the Bria API (commercially licensed, fully managed, model value "RMBG3"), Bria Create, fal.ai (bria/rmbg-3), Bria ComfyUI nodes and a Bria agent skill. API pricing is not yet disclosed.
What this means
RMBG-3.0 reframes background removal as an image-editing task on a generative model, and distillation to 4 steps keeps the runtime at about 6 seconds on one L40S. The benchmark was designed and run by Bria, with a hard-image set it selected, so the rankings are vendor-reported. The category breakdown shows a mixed picture on thin structures, low contrast and semi-transparent subjects. The license split matters for builders: the open weights are non-commercial, so production use goes through Bria's commercially licensed API. The weights also depend on a second gated download from Fibo 2.0 Edit Turbo.
Related Articles
Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license
Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Apple releases FastFrames-L, a 14B two-step video model distilled from Wan2.1-14B for interactive 480p generation
Apple has released FastFrames-L (14B) and FastFrames-S (1.3B), autoregressive two-step video diffusion models that generate 480p video at interactive latency. FastFrames-L is distilled from Wan2.1-14B, and the repository lists 35.4 ms total latency for FastFrames-S on one GPU.
SenseNova-Loop-9B released: 9B parameters, 56 layer executions, 512K context
SenseNova has released SenseNova-Loop-9B, a roughly 9B-parameter multimodal model that runs 56 decoder-layer executions per forward pass from 32 distinct layers. It supports up to 512K tokens of context and, per the model card, needs about 15 GiB of GPU memory at 256K context with 4-bit weights and an 8-bit KV cache.
Comments
Loading...