Skip to main content
RMBG 3.0 is Bria’s third background removal model and the first built on a generative model instead of a segmentation network. It renders a soft grayscale alpha matte rather than a hard binary mask, so hair, fur, mesh, veils, and motion blur keep partial transparency. The base is Fibo 2.0 with an 8B diffusion transformer and a Qwen3-VL-4B prompt and image encoder, distilled to 4 sampling steps with no classifier-free guidance. ComfyUI support landed in ComfyUI #17240, and the template needs ComfyUI 0.46.0 or later. The workflow takes one image and returns three outputs: an RGBA cutout, a background plate, and the processed mask.
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you can’t find them in the template, your ComfyUI may be outdated.If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

Remove Background: RMBG 3.0

RMBG 3.0 remove background workflow with before and after comparison

Download Workflow

Download JSON or search “Remove Background: RMBG 3.0” in Template Library
Models

Diffusion Model

rmbg3_int8_convrot.safetensors → ComfyUI/models/diffusion_models/

Text Encoder

qwen3vl_4b_int8_convrot.safetensors → ComfyUI/models/text_encoders/

VAE

flux2-vae.safetensors → ComfyUI/models/vae/
The same repository holds BF16 versions of the diffusion model and the text encoder if you would rather trade memory for precision. Input material

withered_hydrangeas_still_life_vintage_lit.png

Load into the LoadImage node
  1. Drop your photo into the LoadImage node.
  2. Turn on auto_prompt, which is off in the shipped template. With it off, the node sends the fixed instruction and subject_note is unused.
  3. Optional: write one short sentence about the material in subject_note, for example Clear glass bottle, paper label, metal cap.
  4. Run the workflow. SaveImage writes both results under ComfyUI/output/rmbg3/.

Workflow controls

  • auto_prompt: when on, a vision model writes the matting instruction from the image and your subject_note. When off, the fixed instruction the model was trained on is sent as is and subject_note is ignored.
  • cfg: keep it at 1.0. The model is distilled without guidance, so a higher value trades fidelity to the source for transparency.
  • alpha_brightness, alpha_contrast, mask_threshold, hard_mask: matte post-processing. Leave hard_mask off to keep soft edges.
  • mask_expand, mask_tapered_corners: grow and round the matte before it is applied.

When the cutout is wrong

Two limits come from the model itself: very busy or low-contrast scenes can pull in grass, props, or a watermark next to the subject, and fine structures such as insect wings can lose pixels.

License

The weights are released under CC BY-NC 4.0. Commercial use needs a separate agreement with Bria, either a self-hosted license or the Bria API. The Bria partner node runs the same model through that API.