Quality and limitations
Photos processing uses MODNet, a small portrait matting model that also worked well on the product photos in our initial evaluation. Lightweight processing uses U²-Netp, a smaller general subject model. White background mode uses local color math for logos and signatures without downloading a model. BiRefNet Lite was evaluated and excluded after using much more memory and performing worse on our initial photo set. Large images use a small mask, upsampled at export to preserve the original pixel dimensions.
Full-resolution export describes the file size, not the segmentation detail. Fine hair, fur, glass, signatures and small text may lose detail. Erase and Restore are available to correct the result. Photos without a clear foreground subject may need another tool or manual editing.
This utility is under active evaluation. We do not claim that it matches a professional matting service on every image. The open model weights and browser runtime are reusable offline when the browser cache remains available.