Remove a watermark from an Instagram screenshot.
Watermarks on screenshots are annoying but usually flat, low-frequency, and confined to a corner — exactly the case where LaMa (a diffusion-adjacent inpainting model) excels. This guide covers what to expect, what won't work, and the fastest path.
Legal note
Removing a watermark from copyrighted content you don't own or license is a copyright violation in most jurisdictions. This guide is for content you have rights to modify — your own screenshots with your own overlaid captions, screenshots from creators who have granted permission, or your own material re-imported from a service that added a watermark. See our terms.
Step-by-step
- Open /watermark-remover/. Choose AI mode at the top — it uses LaMa (~200 MB, one-time download, cached forever).
- Drop the screenshot. Screenshots up to 4K work. Larger source files get downscaled to 2K for inference, then composited back — usually invisible in the output.
- Brush over the watermark. Cover the watermark with a comfortable margin — 10-15 pixels of extra area on each side. Too tight and you get edge halos; too loose and the model gets less context to reconstruct from.
- Click Erase. First run downloads the model (~200 MB, one-time). Progress bar shows the download. Subsequent runs are inference-only, ~5-15 seconds depending on CPU.
- Inspect. Zoom in on the erased region. If the background was busy (a face, text, a complex pattern), the fill may be visibly off — try painting a larger mask (more context helps) or switching to Fast mode for content-aware fill on simple backgrounds.
- Save. PNG preserves transparency; JPG is smaller and fine for social sharing.
What works, what doesn't
Works well
- Small watermarks on flat backgrounds (sky, wall, single-color regions).
- Screenshot watermarks in corners with regular photo content behind them.
- Text overlays where the underlying content is smooth or repetitive.
Struggles
- Watermarks over faces or eyes — LaMa can invent features that look off.
- Watermarks over readable text — model can't regenerate the exact words.
- Watermarks that take up more than ~15% of the image area — too much to reconstruct.
Privacy note
Both modes run entirely in your browser. The model file is fetched once from HuggingFace CDN; after that, all inference is local. Your screenshot never leaves the device. If you inspect DevTools → Network during erase, you'll see zero POST requests.