How to remove AI watermarks
Remove AI watermarks without uploading your files
Most online watermark removers work the same way: you upload your picture to a company's server, they process it, and you hope their retention policy says what you think it says. Many then charge a monthly subscription, or stamp their own watermark onto the free-tier result.
WatermarkOut takes a different approach. The AI model runs inside your browser tab using WebAssembly, so the picture you are editing never leaves your device. There is no upload step, no account, no queue, and no server that could keep a copy — because there is no server at all. The site is a folder of static files, and the source is public under the MIT licence if you want to check that claim rather than take it on trust.
Which watermarks it can remove
The detector looks for a badge rather than assuming where one is, so it is not limited to the list below — a stock-photo tag, a channel bug or a logo you added yourself all work the same way. What the list tells you is which generators actually stamp anything in the first place, because several widely-repeated claims are simply out of date.
| Generator | Visible badge? | Notes |
|---|---|---|
| Google Gemini / Imagen | Yes — bottom-right | A white sparkle blended into the image. Since August 2026 paid tiers can switch it off at source. Size varies in fixed steps rather than scaling smoothly. |
| Grok (xAI) | Yes — bottom-right | xAI's Acceptable Use Policy prohibits removing it. Please don't use this tool on Grok output. |
| Bing Image Creator / Designer | Yes — bottom-left | Last confirmed by Microsoft in 2024; we have not found a 2026 update either way. |
| Meta AI | Older images only | The "Imagined with AI" label appeared on earlier output. Newer Meta models label invisibly instead. |
| Midjourney | No | Midjourney has never added a visible watermark on any tier. If you see a mark on a Midjourney image, something else put it there. |
| OpenAI / DALL·E | No | DALL·E 2 and 3 were retired in May 2026. The old five-colour corner bar only appears on archived DALL·E 2 images. |
| Adobe Firefly | Unclear | Reports conflict and we could not confirm the current behaviour. Run detection and see. |
| Stock photo sites | Yes | Repeating diagonal tiles. Mask these by hand — they cover the frame rather than a corner. |
Anything else — a logo, a date stamp, a caption, an unwanted object — you can mask by hand with the rectangle or brush tool. The tool does not care what the thing is; it only sees the region you marked.
How to remove a watermark
- Drop your file in — or click Browse. Nothing is uploaded; the file is read straight into the page.
- Mark the watermark. Choose a preset from the dropdown, or drag a rectangle over it. Use the brush for awkward shapes and the eraser to trim the mask back.
- Pick a quality setting. Balanced suits most images. Best runs a second refinement pass and takes roughly twice as long.
- Click Remove watermark and wait. Typically a few seconds on a modern machine.
- Compare and download. Press Space to flip between the original and the result, then save it as PNG, JPG or WebP.
A tighter mask almost always beats a loose one. Cover the watermark and a few pixels of margin — not half the picture — and the fill will blend better.
Supported formats and limits
- Images: JPG, PNG, WebP and BMP, up to about 100 MB.
- Video: MP4, WebM and MOV. Best for short clips; audio is preserved.
- Browsers: Chrome and Edge 90+, Firefox 90+, Safari 16.4+. WebGPU is used automatically where available and is considerably faster.
- Storage: roughly 250 MB free space for the cached model and app.
Honest limitations
Inpainting is a reconstruction, not a recovery — the model invents plausible pixels, it does not uncover what was hidden underneath. Where it struggles:
| Limitation | Why | What helps |
|---|---|---|
| Busy backgrounds — faces, small text, brickwork | A known weakness of this class of model | Use a tighter mask, or a desktop tool like IOPaint for stubborn cases |
| Video uses one static patch | The AI runs once on a frame, then the clean patch is applied across the clip | Ideal for fixed corner watermarks; moving content behind the mark may show a static patch |
| Long videos are slow | Every frame is decoded and re-encoded in the browser | Trim the clip first |
| Very large images | A browser tab has a memory ceiling | Resize before loading |
Is it really private?
Your images and videos never leave the browser — there is no endpoint to send them to. Two things are fetched over the network, and it would be dishonest not to name them: the app's own files, and a one-time 29 MB AI model downloaded from Hugging Face on first visit. The model is stored on your device and verified against a known checksum each time, so it is never downloaded twice and never silently swapped.
After that first load, the app works with no network connection at all. There is no analytics, no error reporting, no cookies and no fonts or scripts pulled from third-party trackers. You can verify all of this the same way we do: open your browser's developer tools, watch the Network tab, and remove a watermark.
What this does not remove
This matters more than anything else on this page, so it gets its own section rather than a footnote.
A visible badge is not the only mark on an AI-generated image. Google embeds SynthID, and OpenAI's images now carry it too. Meta's newer models use something equivalent they call Content Seal. These are not stamps in a corner — the signal is spread across the entire picture, encoded into the pixels themselves.
Removing the visible badge removes the visible badge, and nothing else. Google designed SynthID to survive cropping, filters, colour changes and lossy compression, and its own settings page states that the visible-watermark toggle "doesn't affect SynthID watermarks or Content Credentials". Because the signal is spread across the whole picture, editing a corner leaves the overwhelming majority of it untouched. Google does say SynthID "isn't foolproof against extreme image manipulations" — but the manipulations in question are ones that wreck the image.
One thing that is lost, unavoidably: saving a file from this tool rebuilds it from raw pixels, which discards the original file's metadata — including C2PA Content Credentials, EXIF and colour profile. That is a side effect of how browsers export images, not a feature, and it does not touch the pixel-level watermark either.
So if your goal is to make an image undetectable as AI-generated, this tool will not do that and neither will any other visible-watermark remover. If your goal is a cleaner picture for your own use, it will.
Fair use
This tool is meant for images you made or otherwise have the right to edit — cleaning a preview badge off your own generated artwork, removing a logo from a photo you own, tidying up your own footage.
Please don't use it to strip watermarks from someone else's copyrighted work, to get around a stock library's licensing, or to pass AI-generated images off as photographs. Provenance badges exist so people can tell what they are looking at, and removing one to mislead somebody is a choice about honesty, not about software.
One specific worth knowing: xAI's Acceptable Use Policy expressly forbids removing the watermark from Grok images — that is a term you agreed to, not a grey area.
More generally, some generators now let you turn the visible badge off at source — Google added exactly that option in August 2026. Where you have that choice, taking it is cleaner than editing the image afterwards, and it leaves the file's other provenance signals intact rather than looking like an attempt to hide something.
Questions
Is it actually free, or is there a catch?
Free, with no paid tier to upgrade to. It costs nothing to run because there is no server — the site is hosted as static files and all the computing happens on your own machine. There is nothing to meter and no usage to bill.
Does it put its own watermark on the result?
No. You get the picture back with the masked area filled in and every other pixel untouched.
Do my files get uploaded?
No. Files are read into the page and processed there. The only network requests are the app's own files and the one-time model download described above.
Does it work offline?
Yes, after the first visit. The app and the model are cached on your device, so you can disconnect entirely and keep working. You can also install it as an app from your browser's address bar.
Will it reduce my image quality?
Only inside the mask. Everything outside the region you painted is copied through unchanged, pixel for pixel. Save as PNG to avoid adding any compression of your own.
Why is the first visit slow?
It downloads the 29 MB AI model once. After that it loads from your own device and starts almost immediately. You can drop a file and mark up the watermark while the model is still downloading.
Which AI model does it use?
MI-GAN, an inpainting model from Picsart Research (ICCV 2023) built for running on phones and in browsers. It runs through ONNX Runtime Web, using your GPU via WebGPU where the browser supports it.
Can I remove a watermark from a video?
Yes, for short clips. The AI cleans one frame and that patch is applied across the video, which works well for a fixed corner watermark. Audio is kept. A mark that moves around the frame is not handled.