Almost everyone who tries AI watermark removal hits the same surprise: you select the area and click generate, the watermark is gone, but where it used to be there is a smeared patch of color or a faint shadow. The tool is not broken. There is a gap between what you expect from “removing” a watermark and what actually happens: it does not restore anything, it repaints.
This guide first covers which images are fair game, then explains how ToolGarden repairs an area inside the browser, then walks through two numeric examples showing how selection size decides the outcome, and finally lists how to diagnose ghosts, soft patches and seams, plus the cases where you should not use this approach. The AI models run locally in your browser, and the image is not uploaded to a server.
1. Draw the line first: only edit images you have the right to modify
A watermark usually stands for a copyright or source claim. Removing it from someone else’s work to republish, sell or pass it off as your own can infringe copyright and breaks the terms of most platforms. The tool is meant for images you are entitled to process, for example:
- Your own photos with a date-time stamp burned into a corner by the camera.
- Assets you made and watermarked yourself, when only the watermarked copy survives.
- Your own promotional images that still carry an old logo after a rebrand.
- Screenshots where you want to remove an account name, a temporary label or personal information.
If the image comes from someone else, get permission first or ask the rights holder for the unwatermarked original. The tool’s own page carries the same reminder: only process images you own or are allowed to modify.
2. How it works: repair is not restoration
The pixels under a watermark have been overwritten, so the original information is not in the file, and no tool can restore it. All it can do is guess what should be there from the surrounding picture and paint that in. The technique is called inpainting. ToolGarden’s pipeline is:
- You drag a rectangle on the image; that rectangle is the region to repaint (the mask). The mask is currently a rectangle, not a brush stroke.
- The tool crops a patch of the image that includes context around the selection and passes it to the model.
- The model repaints only the pixels inside the mask; pixels outside the selection stay unchanged.
- The result gets one more round of edge softening (the “Edge blend” control) to reduce the seam.
| Mode | How it works | Best for | Cost |
|---|---|---|---|
| High quality AI (default) | The MI-GAN ONNX model, run locally in the browser | Photos, textured or gradient backgrounds | The first run downloads about 28 MB of model assets |
| Compatible AI | The LaMa ONNX model, run locally in the browser | Photos and richer textures, as a compatible fallback | Also needs a model download |
| Fast local | No AI model; fills by distance-weighted estimation from nearby pixels | Flat colors, gradients, screenshot backgrounds | Smeared look on complex texture |
There is also automatic degradation between the modes: if you pick High quality AI and the model fails to load or run, the tool tries LaMa, and if that fails too it drops to fast local repair, with a progress message saying it is switching to fast local repair. So after generating, glance at the Mode field in the result details to confirm which one produced the output.
3. The procedure
- Upload an image. JPG, PNG, WebP, GIF, BMP, SVG and AVIF are accepted, up to 50 MB and 40 million pixels.
- Drag the selection. Once an image loads, a default box is placed at the bottom right: 32% of the image width, 14% of the height, with 5% margins from the right and bottom edges. On a 1920×1080 image that is 614×151 px with its top-left corner at (1210, 875). Drag it, resize it with the eight handles, or type values into the X, Y, Width and Height fields.
- Pick a mode. High quality AI is usually the first choice; on a flat color or smooth gradient, Fast local is far quicker.
- Adjust Edge blend. The range is 0 to 36 px, default 12. It grows the mask a few pixels outward and softens the seam.
- Choose the output format. The default follows the source: JPG gives JPG, WebP gives WebP, anything else gives PNG. JPG and WebP also have a quality slider from 50% to 100%, default 92%.
- Click Generate repair, zoom in on the preview, and download when satisfied. Moving the box or changing parameters clears the previous result, so regenerate afterwards.
4. Worked examples: how selection size decides the outcome
Example 1: a 1920×1080 screenshot with a text watermark at the bottom right
Say the watermark text occupies 220×40 px with its top-left corner at (1650, 1000). Do not box it exactly; leave about 10 px on every side: x=1640, y=990, width 240, height 60. The reason is that watermark edges contain anti-aliased semi-transparent pixels, and sometimes a stroke or shadow, so a box that is too tight leaves a faint outline of the letters. If the background is a flat color or smooth gradient, Fast local finishes almost instantly; on a textured background use High quality AI.
Example 2: a 4000×3000 photo, default box versus a tight box
This is the key to understanding result quality. The AI model input is a fixed 512×512 pixels, so the tool crops a patch with context around the selection and shrinks it to 512 before handing it over. Reading the source, the context size is the largest of three numbers: 48 px, 8% of the shorter image side, and 80% of the selection’s longer side; the crop is a square whose side is the selection’s longer side plus twice the context size, capped at the shorter image side.
| Selection | Context | Square crop side | Scale to reach 512 |
|---|---|---|---|
| Default 1280×420 | max(48, 240, 1024) = 1024 | 3000 (capped by the short side) | Shrunk to about 17% of original |
| Tight 500×140 around the logo | max(48, 240, 400) = 400 | 1300 | Shrunk to about 39% of original |
In other words, on the same 4000×3000 photo the default large box has the model painting on a version compressed to roughly one sixth, which is then scaled back up, so the repaired area ends up visibly softer than the grain around it. With a tight 500×140 box over the actual logo the model sees a much sharper picture. The takeaway is plain: the closer the selection hugs the watermark, the better, and for a large watermark it is better to work in several small passes, reloading the previous PNG each time, than to box a huge region at once.
5. Choosing an approach per watermark type
| Watermark situation | Recommendation | Expected result |
|---|---|---|
| Semi-transparent corner logo on a textured photo | High quality AI; box hugging the logo with a 5 to 10 px margin | Usually natural |
| Text on a flat or gradient background | Fast local is enough | Fast and hard to spot |
| Camera date stamp (orange digits) | Small box, High quality AI | Good on simple backgrounds |
| Repeated watermark tiled across the whole image | Not suitable | The selection covers nearly everything, leaving the model nothing to reference |
| Watermark over faces, text or product detail | Not suitable | Hidden detail cannot be recovered; you get plausible invented content |
| Watermark over regular patterns such as brick, grids or tidy text | Be careful; try several small passes | Regular textures tend to misalign or repeat |
6. Common problems and how to diagnose them
| Symptom | Most likely cause | What to do |
|---|---|---|
| A faint outline remains | Box too tight; anti-aliased edges and shadow not covered | Grow the box a few pixels, or raise Edge blend to 18 to 24 |
| The repaired area is blurry or soft | Selection too large relative to the image; model input shrunk heavily | Shrink the selection; handle large watermarks in several passes |
| Color differs from the surroundings | Selection crosses a clear color boundary | Keep each selection inside one color region and do them separately |
| Odd repeated texture appears | Too little context to reference, or the box is too large | Shrink the box; try the other AI mode |
| Slight blocky noise in the result | Output as JPG at low quality, or compressed repeatedly | Output PNG for intermediate results and export JPG only at the end |
| The first run is very slow | About 28 MB of model data is downloading | Wait for it; later runs in the session do not repeat it |
| Mode in the result says Fast local | Model loading failed and the tool fell back | Check the network, refresh, or choose Compatible AI |
Metadata and quality
The output is a new file re-encoded through a Canvas, so the original EXIF data (capture time, device, location) is not kept. If you need to know what metadata the original carried, look at it in the EXIF viewer first. Also, outside the selection, pixels are unchanged when you export PNG; exporting JPG or WebP re-compresses the whole image, producing slight overall changes, so choose PNG when you need the highest fidelity.
Choosing the export format
| Format | When to use it | Watch out for |
|---|---|---|
| PNG | Further editing, intermediate results, transparent areas | Largest files; photos grow noticeably |
| JPG | Final use on the web, social platforms or print | Lossy; transparent areas are filled with white |
| WebP | Web delivery with smaller files | Some older software cannot open it |
7. When not to use this approach
- Images you have no right to modify: a watermark is a rights statement, and republishing after removing it may be infringement.
- Images that must be kept as evidence or records: the repaired region is model-generated content and can no longer count as the original record.
- Watermarks that cover most of the frame or run across the whole image: there is not enough surrounding information, and you should ask the rights holder for a clean version.
- Invisible digital watermarks: this tool handles visible watermarks only and does nothing about hidden marks embedded in image data.
- Needing to recover obscured text or faces exactly: AI generates plausible content; it cannot recover the real information.