Using masked inpainting for authorized image cleanup

The watermark remover lets you paint a mask over an unwanted image region and asks an inpainting model to synthesize replacement pixels. Use it only on images you own or have permission to alter. Watermarks frequently communicate ownership or licensing; erasing one does not remove copyright, attribution requirements, provenance, or restrictions attached to the underlying work.

How it works

The browser sends the original JPEG, PNG, or WebP together with a mask derived from the area you mark. On the server, the mask is resized if necessary and thresholded so white pixels identify the replacement region. A SimpleLama inpainting model fills that region using surrounding visual context and returns a PNG. If no usable mask is present, the script saves an unchanged PNG copy. The model invents texture; it does not recover pixels that existed beneath opaque lettering or logos.

Steps

  1. Confirm the right to alter the image. Locate the original license or owner approval. When possible, obtain an unwatermarked source from the creator instead of reconstructing hidden content.
  2. Upload the best available original. Choose a supported image with enough surrounding texture for context. Avoid first shrinking or repeatedly recompressing the image, because artifacts make reconstruction harder.
  3. Mask narrowly around the obstruction. Paint over all watermark pixels but keep the marked area close to the unwanted element. Excessively broad masks force the model to invent more scene content.
  4. Review the synthesized patch. Download the PNG and inspect repeated patterns, warped edges, faces, text, and object boundaries. Undo through the original source and revise the mask if the replacement is misleading or visibly inconsistent.

Practical use cases

A photographer can clean a proof overlay from their own image after retrieving the licensed master, using the master to verify that the inpainted preview is not mistaken for original detail.

A designer can remove an obsolete date stamp or temporary internal label from company-owned artwork when the editable source is unavailable and the communications owner approves the change.

Limitations

Inpainting produces plausible pixels, not authentic recovery. It can smear text, duplicate textures, bend architecture, alter faces, or create misleading details, especially across large masks and structured patterns. The current workflow is not appropriate for evidence, journalism, restoration records, identity documents, or any use where invented content must not be presented as fact. Removing third-party attribution can violate law or contract. Results are flattened PNG files and may not preserve metadata, animation, layers, or original color characteristics.

Privacy and file retention

Both the image and painted mask are uploaded to the server for model processing. The resulting PNG and temporary inputs are scheduled for deletion after they are older than one hour, while the download link expires after one hour. A mask can reveal exactly what someone intended to conceal or alter. Avoid personal, confidential, evidentiary, or licensed third-party material unless processing is authorized.

Frequently asked questions

Does the model reveal what was under an opaque watermark?
No. The covered pixels are unavailable; the model generates a visual guess from neighboring content. That guess should never be described as recovered evidence.
Why should the mask be narrow?
A focused mask limits how much content must be invented and gives the model more untouched context. It should still cover every unwanted watermark pixel.
May I remove a watermark from an image found online?
Only if the rights and license permit it. Public visibility is not permission to remove attribution or reuse the work; request a licensed original from the owner.