Unmark

A minimal, on-device watermark eraser. Drop an image, brush over the mark (or tap ✦ corner for the standard AI-watermark position), hit erase — the region is reconstructed by a generative inpainting model running entirely in your browser. Nothing is ever uploaded.

How it works

• Model — MI-GAN (migan_pipeline_v2.onnx, ~27 MB), a lightweight generative inpainting network. Self-hosted from public/models/ and cached by the browser after the first visit.
• Runtime — onnxruntime-web, preferring the WebGPU execution provider and falling back to multithreaded WASM (cross-origin isolation headers are set in next.config.ts to enable threads). Runtime binaries are vendored into public/ort/ by scripts/copy-ort.mjs on npm install.
• Inference — src/lib/inpaint.ts feeds the model raw uint8 tensors (image [1,3,H,W] RGB + mask [1,1,H,W], where mask 0 marks the hole) on a ~512 px crop centered on the mask, then composites only the masked pixels back into the original — full-resolution context, no global resampling.
• UI — Next.js App Router + Tailwind. One screen: canvas, brush, corner preset, undo, hold-to-compare, PNG export. Keyboard: [ ] brush size, ⌘Z undo, hold C to compare.

Develop

npm install   # also copies onnxruntime wasm files into public/ort/
npm run dev

npm run build && npm run start for a production build. Deploys to Vercel with zero configuration; there is no server-side compute — the entire pipeline is static assets plus client inference.

Scope

Intended for images you have the rights to — e.g. cleaning the visible sparkle from AI images you generated. It does not (and will not) touch invisible provenance watermarks such as SynthID. Don't point it at other people's copyrighted work.

Credits

• MI-GAN — Picsart AI Research (ICCV 2023); ONNX export by andraniksargsyan.
• Tensor I/O conventions referenced from inpaint-web (Apache-2.0).