RMBG-3.0: Bria's Four-Step Background Removal Based on Generative Models

Bria's third-generation background removal model, uploaded on 11 10, is the first built on a generative model: starting from the editing weights of the 8B-parameter Fibo 2.0 (Flow Matching DiT, Qwen3-VL-4B encoder, FLUX.2 VAE), it is fine-tuned to "input a photo and draw an alpha mask according to a fixed instruction," then distilled to 4 steps without CFG. The output is a single-channel 8-bit mask: white represents the subject, black represents the background, and gray represents semi-transparent areas such as hair and glass. The weights are licensed under CC BY-NC 4.0, for non-commercial use only.

Original title: briaai/RMBG-3.0

Bria's third-generation background removal model, uploaded on 11 10, is the first built on a generative model: starting from the editing weights of the 8B-parameter Fibo 2.0 (Flow Matching DiT, Qwen3-VL-4B encoder, FLUX.2 VAE), it is fine-tuned to "input a photo and draw an alpha mask according to a fixed instruction," then distilled to 4 steps without CFG. The output is a single-channel 8-bit mask: white represents the subject, black represents the background, and gray represents semi-transparent areas such as hair and glass. The weights are licensed under CC BY-NC 4.0, for non-commercial use only.

Why it matters Officially, it identifies the subject based on image content, handling group photos (each person), motion-blurred limbs, and cutout grids more cleanly, with a focus on e-commerce product images

Upvotes 60 · Likes 60 · Downloads 0

Reasons

  • Officially, blind evaluations and visual model comparisons were conducted on 234 difficult images against Photoroom, remove.bg, Ideogram, BiRefNet, and RMBG-2.0: it ranked first with an average rank of 1.51, achieving the best or tied-best result on 185 images; pairwise preference rates ranged from 55.1% (against Photoroom) to 75.9% (against BiRefNet)
  • Officially, it ranked first on 7 categories of difficult images including motion blur, multiple subjects, cutout grids, shadows/reflections, and hair; on 48 online store product images, it achieved an average rank of 1.15, ensuring that text, logos, and badges on products are not removed
  • The model card notes limitations: glass, sheer fabric, and smoke are easily rendered as opaque; cluttered backgrounds and low-contrast scenes perform worse than remove.bg and RMBG-2.0, with 5 of the 234 images ranking last

Usage

  • Diffusers >= 0.42.0: Load the transformer subdirectory from this repository, while using briaai/Fibo-2.0-edit-turbo for the remaining components (text encoder, VAE, font) to compose the BriaFibo2Pipeline
  • You must use the fixed English instruction specified in the model card (requiring the generation of a grayscale alpha mask), with num_inference_steps=4 and guidance_scale=1.0; modifying the instruction will result in a color image instead of a mask
  • Convert the output to grayscale, scale it back to the original image size, and apply it as the alpha channel to the original image; if a binary mask is needed, apply thresholding. ComfyUI has official Bria nodes, but they use a paid API

Risks

  • Weights are for non-commercial use only (CC BY-NC 4.0); commercial use requires a separate agreement with Bria or use of the paid API; downloading requires logging into Hugging Face and filling out a form, and agreeing to the terms for Fibo-2.0-edit-turbo is also required
  • With 8B parameters, official tests on L40S showed a peak VRAM usage of 28.5 GiB for the entire pipeline and approximately 6 seconds per image, recommending GPUs with 32 GB or more; for smaller GPUs or batch processing, the official recommendation is to continue using RMBG-2.0
  • Rankings come from blind evaluations and visual model comparisons organized by the official team; this entry has not been independently verified

Open

Hugging Face Trending

Explore the latest AI Radar updates