Use it on files you own
Use this tool only on images you own or have permission to edit. Removing a credit or a mark from someone else's work can break copyright or licence terms, and that responsibility is yours.
Free · Runs in your browser
This free image watermark remover lets you paint over a logo, caption or stamp on a photo and fill the area in with the MI-GAN inpainting model. The model runs inside your browser, so the picture itself never leaves your device.
A small inpainting tool with the controls that matter: paint, repair, compare and download.
There is no fee, no credit balance and no mark added to your result. Paint, repair and download as many images as your device can handle.
The repair model is a 28 MB ONNX file executed by WebAssembly in a background worker. Nothing about your image is sent to a server, and the page stays responsive while the model works.
Instead of guessing where the mark is, you paint it. A brush from 4 to 100 px, an eraser, undo and a 100 to 300 percent zoom let you cover a thin line or a large logo precisely.
The repaired colour is copied back only under your brush strokes. Every other pixel stays exactly as in the original, and the result is saved as a PNG without another round of JPG compression.
Four steps from upload to download.
Click the button or drop a JPG, PNG or WebP file onto the tool. Files up to 20 MB, 16 megapixels and 4096 px on the longest side are accepted.
Select the brush and cover the whole watermark, including its soft edges. Use the eraser for stray strokes, undo to step back and the zoom slider for small details.
Press the repair button. The first run downloads the model, then the area is filled in. A timer and a cancel button stay on screen while it works.
Switch between original and repaired, then download the PNG. If a trace remains, choose keep editing and paint over it again on top of the result.
Clean a date stamp or channel logo from a frame you captured from your own footage before you use it as a cover image.
Remove a supplier code or a stray label from a photo you are allowed to edit, before you list the item in your shop.
Tidy a scanned print, a draft stamp or a signature from your own artwork so the gallery shows the clean version.
Take a page number, sticker or leftover caption out of a screenshot or diagram before you paste it into a presentation.
Wipe a username overlay or a timestamp from a screenshot you took, right on your phone and without installing an app.
Try inpainting on a sample photo to see how a model fills in texture, and how mask size and surrounding context change the outcome.
Use this tool only on images you own or have permission to edit. Removing a credit or a mark from someone else's work can break copyright or licence terms, and that responsibility is yours.
Plain sky, walls and soft gradients repair well. Faces, text, fine patterns and busy scenes can come out blurred or invented. Check the result at 100 percent before you use it.
Input is JPG, PNG or WebP up to 20 MB, 16 megapixels and 4096 px per side; output is PNG. The browser needs WebAssembly, Web Workers and a secure HTTPS page, and a large image needs enough free memory.
Yes. You do not pay anything and you do not need credits. The processing runs on your own device, so it costs us no server time per image either.
No. The image is decoded, painted on and repaired in your browser. The only thing the page downloads is the repair model, a static file served from this site. We never receive your picture or your brush strokes.
JPG, PNG and WebP files up to 20 MB, with at most 16 megapixels and no more than 4096 px on the longest side. The result is always a PNG.
The first time, the page downloads the model of about 28 MB and starts it. Your browser can keep it in its cache, so later repairs on this device usually skip the download. The repair itself then depends on your processor and the size of the marked area.
You paint a mask. The page cuts out a square region around it, at least 512 px wide with extra context around your strokes, scales it down to at most 1024 px and passes it with the mask to MI-GAN, an inpainting network trained on the Places2 dataset. The generated pixels are placed back under your strokes only.
No. It fills the painted area with plausible texture and does not recover what was under the mark. Small marks on smooth backgrounds work best, while large, semi-transparent or repeating marks across the whole picture usually leave visible traces.
Choose keep editing. The repaired image becomes the new starting point, so you can paint a slightly larger area, or only the leftover spot, and repair again. Each round works on the previous result.
Pixels outside your brush strokes are copied unchanged. The file is saved as a PNG, so no new JPG compression is added, but a PNG is often larger than the JPG you started from, and metadata such as EXIF is not kept.
The model needs WebAssembly, Web Workers and a secure HTTPS page. Update to a current Chrome, Edge, Firefox or Safari, turn off extensions that block workers and try again. If memory runs out, close other tabs or use a smaller image.
Yes. Paint with your finger, zoom with the slider and download the PNG. Phones have less memory, so very large photos may fail; resize them first if that happens.
Describe a scene or start from an image, and BestAIVideo's AI video generator turns it into a short clip.