hu-qi/clearmatte-studio

ClearMatte Studio is a small web app that removes the background from an image. The whole thing runs inside the browser, so the picture never leaves your computer. It is built as a static site using a JavaScript library called Transformers.js and an open model named birefnet-lite-512, which has been converted to the ONNX format so it can run on the web. The feature list is short. You can upload a PNG, JPEG, or WebP image. The app then runs alpha matting, which is the technique that figures out which pixels belong to the subject and which belong to the background. You can preview the cut-out result against a checkerboard grid, a white background, a dark background, or a custom color, and there are sliders for cutoff, softness, and cleanup before you export. The final image can be saved as a transparent PNG or flattened onto a color background. Running it locally is as simple as starting a Python HTTP server in the project folder and opening localhost:5173. The README notes that WebGPU needs a secure context, which localhost counts as, and that browsers without WebGPU support fall back to WASM, the WebAssembly runtime. The model itself is not bundled. The app first looks for it in a local directory under models/studioludens/birefnet-lite-512, which you populate by downloading the model from Hugging Face. Serving the files from the same site avoids cross-origin issues and means visitors do not need to reach Hugging Face from the browser. If the local copy is missing, the app falls back to fetching from the Hugging Face Hub. Visitors from mainland China are routed to the local copy first. URL parameters such as model=local, model=remote, model=hub, model=mirror, and mirror=0 or mirror=1 let you override the loading behavior for testing. Deployment is simple: upload the repository as static files behind HTTPS, with no build step or server runtime needed. The repository has 1 star at the time of writing.