Find the foreground, refine the edges, and take the original detail with you.
U²-NetP · 320×320 inference · single-thread CPU/WASM · measuring assets.
Apache-2.0 model; MIT runtime. Model license. Desktop Chrome tested; low-memory/mobile and WebGPU are unvalidated.
Hair, whiskers, glass and low-contrast edges need review. Upscaling a small mask cannot restore missing detail.
PNG, JPEG, WebP · 30 MiB · 20 MP · 8,192 px/side · no animated images
Choose an image and explicitly load the model.
No image is sent to a processing service.
Output uses original oriented dimensions and image colors, with your mask for alpha. JPEG always flattens onto the replacement color.
The lightweight foreground model was reviewed on a NASA portrait, coffee cup and cat image. It can retain background near the portrait, miss parts of the spoon, and lose whiskers/fur detail. Use compare, erase and restore; no perfect-edge or transparent-object promise is made.
Image dimensions are checked before decode and again after browser orientation. One image at a time; no batch queue. Original colors are used at export, not an enlarged 320×320 inference image. Model loading is explicit; there is no cloud fallback, analytics or persistent image/mask history.
Clear session closes the image bitmap, terminates the worker, and drops masks/history/export URLs. It cannot erase OS memory or previously saved files. Ordinary application/model requests still reach the host.
Validated in desktop Chrome 155 on a Windows host with about 16 GB RAM, using one single-thread WASM inference worker. Low-memory/mobile, WebGPU and other browsers remain unvalidated. Input is limited to 30 MiB, 20 MP and 8,192 px per side. Mask history is capped at eight snapshots or 32 MiB total. Large images retain fewer undo steps.