'/><circle cx='400' cy='230' r='120' fill='%23e8ecf1'/><text x='400' y='240' font-size='28' text-anchor='middle' fill='%234f7cff' font-family='sans-serif'>your cutout</text><text x='400' y='420' font-size='16' text-anchor='middle' fill='%2398a3b8' font-family='sans-serif'>drop a white-background photo in the tool below</text></svg>)
Pro editor
Drag & drop → Remove → preview on checker / white / mask → download. Original resolution preserved.
JPG / PNG / WEBP · stays on your machine
History & gallery
Last cutouts in this browser (localStorage). Re-download anytime. Server files also in outputs/.
Why precise?
Tuned for white-background studio shots with difficult foregrounds.
BiRefNet Heavy, not U²-Net
Dichotomous segmentation with gradient supervision for hair-thin structures. No 320px blocky edges.
Full-resolution soft alpha
Mask bilinear-upsampled to your original size. Threshold 0 keeps flyaways semi-transparent.
White despill
Suppresses white halo on petals/hair from white studios without touching opaque pixels.
HR 2048 + Matting modes
Switch to HR for branches/petals, Matting for portraits — same tool, no reinstall.
Commercial-safe MIT
No BRIA non-commercial trap. BiRefNet weights are MIT — use in client work.
Private by default
FastAPI on localhost. Images never leave your PC. History in your browser only.
How it works
Drop a photo
White-background JPG/PNG. Kept at full size; model sees a 1024–2048 normalized copy.
Heavy inference
Swin-Large BiRefNet predicts a soft alpha matte. GPU fp16 if available, CPU fallback.
Refine + export
Despill whites, composite RGBA, save transparent PNG + mask to outputs/.
FAQ
Which model for hair vs trees vs petals?
Default General handles most photos. Portraits → Portrait Matting. HD BiRefNet gives the finest trees/petals edges but needs WebGPU + lots of free RAM — on weaker machines it fails, so the site auto-uses General.
Why is my first run slow?
Weights (~885MB) download from HuggingFace once to ~/.cache/huggingface, then cached. GPU does 1024px in ~1–3s; CPU takes 10–60s.
Can I use results commercially?
The in-browser default model (RMBG-1.4) is free for non-commercial use — check Bria's license before client work. For commercial jobs, run the Python backend with BiRefNet (MIT) or pick HD BiRefNet in the tool on a strong machine.
How do I get the sharpest edges?
Keep threshold 0, despill 0.5–0.7, inference 1536–2048, and inspect at 100% zoom. Avoid recompressing the PNG.
Where are files saved?
Browser history (thumbnails) + outputs/*_rgba.png on disk. GET /api/outputs lists server files.