youcutai
Bria AI via fal.airemoval

Bria RMBG 2.0: Portrait-Grade Background Removal

Bria RMBG 2.0 is a background removal model from Bria AI, trained exclusively on licensed imagery. It is particularly strong on people and portraits, producing smooth alpha edges around skin and hair. youcut.ai uses it as the primary engine for Portrait mode and as the automatic fallback whenever BiRefNet returns a low-confidence matte. Bria also publishes a video background removal model that powers the Video Background Remover.

Strengths

  • • Trained on fully licensed data – safe for enterprise use
  • • Smooth, natural edges on skin and hair
  • • Robust to low-contrast portraits
  • • Companion video model for frame-consistent cutouts

Limitations

  • • Slightly slower and costlier than BiRefNet
  • • Less precise on very thin non-human structures (wires, jewelry)
  • • Proprietary weights

Best for

  • • Portraits and headshots
  • • ID and passport photos
  • • Enterprise workloads requiring licensed training data
  • • Video background removal
Specs
TaskBackground removal (image and video variants)
Max input2048 px image; 1080p video
Latency~1–2 s per image
LicenseCommercial, licensed training data
Provider keysfal-ai/bria/background/remove, fal-ai/bria/video/background-removal

Try Bria RMBG 2.0 in

Compare

Frequently asked questions

Why does youcut.ai use two background removal models?+

BiRefNet v2 is fastest and cheapest for products and graphics; Bria RMBG excels on people. Running the best model per mode, with automatic fallback, gives better results than any single model.

What does "licensed training data" mean for me?+

Bria trains only on images it has rights to, which matters for companies with strict compliance requirements around AI provenance.

Does Bria power the video tool?+

Yes. The Video Background Remover uses Bria's video background removal model, which keeps masks consistent across frames.

Can I choose which model runs?+

Indirectly: Portrait mode uses Bria first; other modes use BiRefNet first. API users can pass an explicit engine parameter.

Is the output resolution limited?+

Mattes are computed at up to 2048 px and upsampled with guided filtering to your original resolution, up to 4K.