Free watermark and object remover
Brush over the thing you want gone. Choose Fast for instant results, or AI for photorealistic inpainting powered by LaMa. Files never leave your device.
Both modes work the same way from your side: paint a red mask over the watermark, logo, tourist, power line, or whatever else you want removed. Fast fills the region with content-aware pixel sampling in milliseconds. AI runs a full neural inpainting model in your browser for photo-quality results — no server, no upload.
When to use which mode
- Fast mode — instant. Best for solid or smooth backgrounds (skies, walls, plain product shots, gradients). Struggles on textured backgrounds like grass, faces, or repeating patterns.
- AI mode (LaMa) — ~10 seconds per image after the first-time 200MB model download. Handles textures, natural scenes, faces, patterns. First run downloads and caches the model; subsequent runs are instant to start.
Rule of thumb: try Fast first (it costs nothing), switch to AI if the result looks patchy.
Tips for cleaner erasing
- Cover the whole watermark plus a small border — 3-6 pixels beyond the visible edge. Partial coverage leaves ghosting.
- Use larger brushes for large objects — fewer, cleaner strokes beat many small ones.
- Smaller masks give better results — inpainters guess based on surrounding pixels. Less mask = more context = better guess.
- Fast mode: paint tight, not sloppy — every masked pixel gets averaged from neighbors, so extra mask = extra blur.
- AI mode: chunky watermarks are fine — LaMa is trained to handle large occlusions.
Frequently asked
Is my image uploaded anywhere?
No. Everything runs in your browser. Fast mode uses canvas pixel math. AI mode uses onnxruntime-web to run LaMa locally with the WebAssembly backend. Zero POST requests.
How big is the AI model?
Roughly 200MB. Downloaded once, cached in your browser's Cache Storage. Subsequent visits reuse the cached copy — no re-download.
Why does the mask sometimes leave a ghost?
Extend the mask 3-6px beyond the visible watermark. Anti-aliased edges of the watermark still contain original color, and if any of those pixels stay unmasked they bleed back into the fill.
Can I use this to remove watermarks from copyrighted images?
Only remove watermarks you have the right to remove — your own, licensed content, or content in the public domain. Removing watermarks to bypass copyright is illegal in most jurisdictions.
Does the AI mode work offline?
Once the model is cached (after first successful load), yes. The tool page itself is precached by the service worker.
What image sizes are supported?
Fast mode: no practical limit (tested to 4K). AI mode: images are resized to 512×512 for inference then composited back, so very large images take longer but detail outside the mask is preserved from the original.