youcutai
ZhengPeng7 (open source) via fal.airemoval

BiRefNet v2: Background Removal Model

BiRefNet (Bilateral Reference Network) v2 is an open-source high-resolution dichotomous image segmentation model. It predicts a soft alpha matte that separates the main subject from the background, handling hair, fur and thin structures at up to 2K resolution in well under a second on a modern GPU. It is the default engine for General, Product and Logo modes on youcut.ai.

Strengths

  • • Excellent edge quality on hair, fur and thin objects
  • • Very low cost per image (roughly $0.001–0.003)
  • • Handles 2K inputs natively with sub-second latency
  • • Open weights, no vendor lock-in

Limitations

  • • Can struggle when subject and background share similar colors
  • • No semantic understanding of which object is the "subject" in cluttered scenes
  • • Glass and semi-transparent objects produce hard edges

Best for

  • • Product photos on any background
  • • Portraits with loose hair
  • • Logos and graphics
  • • High-volume API workloads
Specs
TaskDichotomous image segmentation (alpha matte)
Max input2048 px (tiled for larger)
Latency~0.5–1 s
LicenseMIT (weights), commercial use allowed
Provider keyfal-ai/birefnet/v2

Try BiRefNet v2 in

Compare

Frequently asked questions

Is BiRefNet v2 better than remove.bg?+

Edge quality is comparable on most photos and often better on hair. The bigger difference is cost: BiRefNet runs for a fraction of a cent, which is why youcut.ai can offer HD cutouts from $0.03 instead of $0.22.

Which youcut.ai modes use BiRefNet v2?+

General, Product, Logo and Anime background removal, batch and GIF removal, and the cutout step inside passport photos, background generation and product tools.

Can it remove backgrounds from anime images?+

Yes. The v2 weights generalize well to illustrations and flat-shaded art, and youcut.ai applies edge post-processing tuned for line art in Anime mode.

What happens when BiRefNet fails?+

youcut.ai runs a quality check on the matte; low-confidence results are automatically retried with Bria RMBG as a fallback.

Is it open source?+

Yes, the code and weights are published under the MIT license. youcut.ai hosts it via fal.ai so you get GPU speed without running anything yourself.