Paint over a watermark to erase it.

Most online erasers only handle one specific overlay — a single brand's logo, a fixed-position string, a known font. This manual watermark remover takes the opposite approach: you tell the tool exactly which pixels to erase by painting over them, and it reconstructs whatever was underneath. That means the tool is not tied to any particular AI generator, photo agency or social network — if you can see the watermark, you can paint it, and the tool can erase it.

We built it after fielding requests for watermarks our calibrated tools don't cover yet: Jimeng outputs, Midjourney sparkle marks, news-channel station IDs, freelance photographer signatures. Rather than ship a calibrated remover for every one, the brush-based approach gives the user direct control and works on every input.

After you drop an image, the tool decodes it on a canvas at up to 1024 px on the longest side (sufficient for most workflows and friendly to mobile memory). A second canvas sits on top of the image — that's where your brush strokes go. Each stroke writes a circular footprint into a binary mask buffer at the same resolution as the displayed image; a red overlay shows you what's marked.

When you click Erase, the manual watermark remover crops the painted region plus 32 px of context, hands the crop and the mask to the inpainting backend, and pastes the result back into the original image. On desktop, the backend is MI-GAN — an ICCV 2023 GAN-based inpainter — running locally via onnxruntime-web with WebGPU acceleration. On mobile, the backend is a neighborhood-average fill that needs no model and no GPU.

Use the manual watermark remover when the watermark you need to erase isn't one of the handful that have a dedicated calibrated tool, or when the watermark sits in an unusual position — top of the image, in the middle, repeated, partially overlapping a subject. Painting gives you exact control over what gets erased and what stays.

Conversely, if your image is from Google Gemini or Doubao, the brand-specific tools on this site will produce a sharper result with less effort — they know the precise pixel location and (in Gemini's case) the exact alpha map of the overlay. The brush is for everything else.

The tool never sees your image. Decoding happens in a Canvas element. Painting writes into an in-memory mask. Inpainting runs in your browser via onnxruntime-web (desktop) or pure JavaScript (mobile). Re-encoding happens via canvas.toBlob. None of these steps touches a network — open the browser's DevTools network panel and watch as you process an image, the only requests are static-asset fetches.

That privacy guarantee is the reason we limit ourselves to in-browser models like MI-GAN. Larger models would give marginally better quality at the cost of either uploading your image to a server or shipping a multi-hundred-megabyte download.

Paint a little wider than the watermark itself — that gives the inpainter more useful context, and faint watermark edges that you didn't quite cover would otherwise show up in the output. The 32 px crop padding helps but it isn't infinite.

Inpainting is reconstruction. On uniform backgrounds (sky, walls, gradients) the result is essentially indistinguishable from the original. On complex textures (text, faces, fine fabric) the inpainter will fabricate plausible patterns that may not exactly match. For those cases, paint a slightly smaller mask, run it, and only paint the residual if any remains.