Cut out backgrounds with on-device AI segmentation. Drag a photo, click Remove, download a transparent PNG. Works batch. Zero uploads — your images never leave your browser.
100% on-device AI. Your photos are never uploaded. The segmentation model runs locally via ONNX Runtime — first load downloads ~17 MB, then everything stays in your browser cache.
Runs entirely in your browser. Your data never leaves your device.
Drag & drop images here
or click to select files · JPG, PNG, WebP supported
Unlike older chroma-key or threshold methods, modern segmentation models understand what an object "is" — edges, textures, even wispy hair. Here's the pipeline:
| Runtime | MacBook M2 Air | iPhone 15 Safari | Windows Chrome (RTX 3060) |
|---|---|---|---|
| WebGPU | 2.1 s | — | 0.9 s |
| WASM (multi-thread) | 3.8 s | 11.4 s | 3.2 s |
WebGPU is available in Chromium 113+, Safari 17.4+, and Edge 113+. All users get WASM as a reliable fallback.
This tool accepts JPG, PNG, WebP, and BMP as input — any format your browser can decode. The output is always a PNG with full alpha transparency so you can drop it directly onto designs, presentations, or product listings.
There is no hard limit, but images larger than 20–25 MP may cause noticeable memory pressure on lower-end devices. The model internally resizes very large images to 1024 pixels on the longest side for inference, then rescales the mask back to full resolution for pixel-accurate output.
Yes. BiRefNet-lite is trained on a diverse set of real photographs, digital illustrations, and graphic designs. Photos with clear foreground subjects (people, products, animals) produce the cleanest cutouts. Illustrations work well too when the subject has distinct edges from the background.
The model produces per-pixel alpha masks with soft edges, which usually handles hair, fur, and thin objects well. However, extremely fine or transparent subjects (glass, lace, smoke, semi-transparent fabric) remain challenging for any single-model segmentation. For professional results we recommend combining the output with manual touch-up in your editor.
No. The model runs 100% inside your browser via Transformers.js + ONNX Runtime Web. No image bytes leave your device — not even for the AI model itself after first load. Pull the ethernet cable and the tool still works on already-downloaded models.
We ran this tool on a curated set of 30 real-world photos to measure processing time and visual quality:
| Test Sample | Resolution | Subject | Runtime | Quality |
|---|---|---|---|---|
| Portrait photo | 4032 × 3024 | Person, hair details | 3.1 s | ⭐⭐⭐⭐⭐ |
| Product shot | 5472 × 3648 | Shoes on white | 2.6 s | ⭐⭐⭐⭐⭐ |
| Pet photo | 3024 × 4032 | Cat, fur details | 2.9 s | ⭐⭐⭐⭐ |
| Travel photo | 6000 × 4000 | Person at landmark | 3.4 s | ⭐⭐⭐⭐⭐ |
| Flat lay | 4000 × 3000 | Small objects | 2.4 s | ⭐⭐⭐⭐ |
We tested 30 photos — avg processing 2.8s on M2 Air. Quality ratings from blind side-by-side reviews. Last verified: September 2026.