Cut out a clear subject and download a transparent PNG. First use downloads ~19 MB (model 4.6 MB + engine 14 MB), then cached. The lightweight u2netp model works best with clear subjects; fine hair, soft edges and busy scenes may need manual cleanup. Your image stays on your device.
Choose an image to get started.
FAQ
What result can I expect?
A PNG with transparency at the original pixel size, up to 25 MP. The mask is predicted at 320 × 320 and enlarged, so small details and complex backgrounds may be imperfect.
Is my image private?
Yes. Inference runs on your device. Only the self-hosted model and engine are downloaded and cached; your image is never uploaded. Downloading sends only the tool name to our counter.
Which licences cover the model and engine?
U²-Net code and released weights use Apache-2.0 (/vendor/models/LICENSE-U2Net); the ONNX export comes from rembg under MIT (/vendor/models/LICENSE-rembg). ONNX Runtime Web uses MIT (/vendor/ort/LICENSE).