REMOVE OBJECT FROM PHOTO

Remove overlaid objects and keep the original detail

Badges, stamps and corner marks composited onto an export can be reversed rather than painted over. Everything runs locally in your browser.

Upload image Open the editor

Runs entirely in your browser · Sign in required · Nothing uploaded

'Remove object from photo' usually means one of two very different jobs, and it is worth separating them before you pick a tool. The first is removing something that was physically in the scene — a bin, a passer-by, a power line. The second is removing something that was added on top of the finished image — a badge, a stamp, a corner mark, an export overlay.

The first job requires invention. There is no record of what was behind the bin, so any tool that removes it has to synthesise plausible replacement content and hope it matches. The second job requires no invention at all, because the overlay was blended with partial transparency and the original content is still partly encoded in the result.

This tool does the second job, and it does it by arithmetic rather than guesswork. That means the recovered area contains the actual original detail — the same grain, the same edges, the same colour — instead of a synthesised approximation that looks right at thumbnail size and falls apart at full resolution.

Reversal instead of reconstruction

Real detail, not synthesis

Recovered pixels are computed from the composited values, so the result contains genuine original content rather than generated filler.

No cloning repetition

Nothing is copied from elsewhere in the frame, so patterns and textures never repeat visibly around the treated area.

Consistent noise

Sensor grain and compression noise survive the correction, so the region matches its surroundings statistically.

Works over busy areas

Overlays sitting on faces, foliage or architecture come away cleanly because the underlying values are recovered, not invented.

Deterministic output

The same input always produces the same result — there is no random seed and no model drift between runs.

Private by construction

Photos are processed in the browser tab and never transmitted, so sensitive material stays on your device.

How it works

1
Open the editor
Click Upload image to start after signing in; no installation is involved.
2
Add the photo
PNG, JPEG or WEBP up to 25 MB. Video is supported too, in MP4, WebM or MOV up to 100 MB.
3
Automatic detection
The engine finds the overlay, works out its scale and opacity profile, and validates the match before applying anything.
4
Verify with the slider
Drag through the before/after comparison to confirm the treated region reads as untouched.
5
Export
Download the lossless PNG result, or process a queue and take everything as a ZIP.

Overlay removal versus generative object removal

Generative object removal has improved enormously, but its fundamental constraint has not changed: it is producing content that never existed. When you erase a person from a beach photo, the model paints sand where the person was. It might be excellent sand. It is not the sand that was actually there, and there is no way for the model to know whether a footprint, a shadow or a shell belonged in that spot.

Overlay removal has access to information the generative approach does not. Where an overlay is 40% opaque, 60% of the original signal is still sitting in the file. The job is not to imagine the missing content but to amplify and correct what remains. That is why the output is deterministic and why it stands up to pixel-level inspection.

The trade-off is scope. This approach only works on things that were blended on top. If the object you want gone was in front of the camera, you need a generative tool — and you should expect invented detail as part of the deal.

How the correction is validated

Applying an inverse blend with the wrong parameters can look worse than doing nothing, so the pipeline treats correction as a hypothesis to be tested rather than a step to be executed. For each candidate overlay position, scale and variant, it applies a trial correction to a working copy and measures the outcome.

The scoring looks at whether the treated region's statistics now resemble its immediate surroundings: local contrast, noise level, gradient continuity across the former boundary. A correct correction makes the region indistinguishable from its neighbourhood; an incorrect one leaves measurable discontinuity.

Only the highest-scoring candidate is applied to the real output, and only if it clears a confidence threshold. When nothing clears it, the tool reports that no supported overlay was found. That refusal is deliberate — a photo returned unchanged is recoverable, a photo returned badly corrected is not.

Candidates are tested on a working copy before anything is committed.
Scoring compares local contrast, noise and gradient continuity with the surrounding area.
Only the best-scoring candidate above the confidence threshold is applied.
Low-confidence cases are reported rather than silently processed.

Video as well as stills

The same engine handles video, which is where reversal really pulls ahead of generative approaches. Frame-by-frame inpainting has no memory between frames, so the invented content shifts slightly each time, producing a shimmer in the treated region that is far more noticeable in motion than any single frame suggests.

Because reversal is deterministic and derives its output from the actual frame content, consecutive frames stay consistent. The corrected region moves with the footage as it should, with no flicker and no crawling texture. Audio and container settings pass through untouched, so a clip comes back with the same runtime, frame rate and sound it went in with.

Frequently asked questions

Can it remove a person or object from the scene?

No. It reverses overlays composited on top of a finished image. Objects that were physically in the scene need generative fill, which invents replacement content.

How is this different from content-aware fill?

Content-aware fill synthesises new pixels from surrounding context. This recovers the actual original pixels mathematically, so nothing is invented.

Does it leave repeated texture?

No, because nothing is cloned from elsewhere in the image. Each corrected pixel is computed from its own composited value.

Will results vary between runs?

No. The process is deterministic — identical input produces identical output every time.

Does it work on video?

Yes, on MP4, WebM and MOV up to 100 MB, with consistent frame-to-frame results and audio preserved.

Is anything sent to a server?

No. All decoding, detection, correction and encoding happen inside your browser.

Reverse the overlay, keep the photograph

Deterministic, private and free. Load a file and see how the treated area holds up at full zoom.

Upload image