How cutoutpic works

When you drop an image in, your browser downloads a segmentation model and runs it locally. Nothing about the picture is transmitted: no upload, no temporary copy on a server, no queue. This is not a promise about how we behave — it is how the free path is built. You can confirm it yourself in your browser devtools: open the Network tab and watch that no request carries your image.

The trade-off is that the model has to reach your device once. It is cached afterwards, so the second image you process starts immediately.

The page checks what your browser can do and picks a path. You do not choose anything.

• Graphics acceleration (WebGPU) — a fast result appears first, then a refined model loads in the background and replaces it. Hair and other fine edges come out noticeably cleaner.
• Compatibility mode (WebAssembly) — only the fast model runs. Results are still usable, but edges are rougher and processing takes longer.

The tier in use is shown next to the page heading, so you know what to expect before you start rather than after the result disappoints you.

• A subject that stands out from its background works best — contrast in brightness or colour matters more than resolution.
• Even, diffuse light beats hard shadows. A shadow that touches the subject is often read as part of it.
• Very fine detail — loose hair, mesh, fur against a busy background — is where the refined tier earns its keep.
• If nothing separable is found, you get an explicit message rather than a blank image. That is deliberate: a silent empty result is worse than a refusal.

• Input: JPEG, PNG or WebP, up to 12 MB.
• Output: PNG with an alpha channel, up to 10 megapixels.
• The free preview is capped in resolution. Full-resolution export is a paid, server-side feature.

Any current version of Chrome, Edge, Firefox or Safari will work. WebAssembly with SIMD is the minimum; WebGPU unlocks the refined tier. The page must be served over HTTPS — browsers only expose graphics acceleration in a secure context.