Before anything else: confirm you are allowed to edit this photo. If you shot it, made it, or hold a licence that permits modification, carry on. If it belongs to somebody else, stop. See the Acceptable Use Policy.
What you need before you start
- The highest-resolution copy you have. More surrounding pixels give the reconstruction more context, and the repair gets less obvious. Never work from a screenshot of the photo if the original exists.
- The original file, not a re-save. Every JPG round-trip adds compression artefacts that blur the boundary between mark and photo, which makes detection worse.
- A backup. Removal is destructive by nature. Keep the marked original.
Step by step
- Load the photo. In this tool that means dropping it onto the page; the file stays on your device.
- Let detection run. The tool looks for regions that behave like an overlay: consistent transparency, synthetic edges, repeated placement.
- Check the detected region covers the whole mark, including its soft fringe and any drop shadow. A mark’s anti-aliased edge is what causes the faint ghost outline in a bad removal.
- Reconstruct. The masked pixels are inpainted from the surrounding texture and then colour-matched so grain and tone carry through.
- Inspect before you use it with the checklist below.
Which photos come out clean
The difficulty is set almost entirely by what sits underneath the mark.
| What’s under the mark | Result | Why |
|---|---|---|
| Clear sky, studio backdrop, flat wall | Usually invisible | The surrounding pixels almost fully determine what belongs there |
| Even texture such as grass, gravel, sand, fabric | Usually good | Texture can be synthesised convincingly; exact detail doesn’t matter |
| Blurred background (shallow depth of field) | Usually good | There is little detail to get wrong |
| Gradients, water, sky with soft cloud | Fair | Banding can appear if the gradient is smooth and the patch is large |
| Architecture, horizons, straight edges | Poor | Lines must line up exactly across the repair or the eye catches it |
| Hair, foliage, patterned fabric | Poor | High-frequency detail turns to mush |
| Faces, hands, text, logos | Very poor | The reconstruction invents plausible detail, and on a face plausible is still wrong |
How to inspect the result
Do all five, in this order. Most bad removals fail on the first two.
- Zoom to 100%. Fit-to-screen hides everything. Look at the repaired region at actual pixel size.
- Look for a rectangle. A visible seam or a slightly softer block means the mask edge wasn’t feathered enough.
- Check for a ghost. A faint outline in the shape of the original mark means the mask was too tight and caught none of the fringe.
- Check the texture continues. Grain, noise and pattern should run through the patch, not stop at its border.
- Check the colour under different light. Some patches only show their tint on a bright screen or against a white page.
If it fails any of these and the photo matters, retouch the region by hand instead. An automated tool is a time-saver, not a replacement for a retoucher on a hero image.
When to stop trying
Give up on the automated route when the mark covers more than roughly a third of the frame, when it sits across a face, or when it is tiled across the whole image. In those cases there is not enough uncontaminated context left, and every result will look repaired.
The realistic alternatives are: get the unmarked original from whoever holds it, buy the licence, or use a different image.
Common mistakes
- Working from a downscaled copy. You throw away the context the reconstruction needs, then wonder why it smears.
- Saving back to JPG at high compression. The whole frame gets a second lossy pass. Export to PNG or lossless WebP if you can afford the file size.
- Assuming a clean result means a clean licence. The mark is gone; the copyright is not.