Screenshot Redactor
Hide private details with solid black boxes. Review and download a flattened PNG.
Paint over a small distraction and let MI-GAN estimate the image underneath.
Choose an image, then paint over the object, including its edges. Use Undo or Clear selection to correct the mask. Keyboard users can add a rectangle with percentage coordinates. Process the selection, compare the result, then download PNG.
Drop files here or choose below
Self-made sample files. Processing still runs locally.
PNG / JPG / WebP · One file · 10 MiB per file · 2048 px per side · 4,000,000 pixels · PNG output. First-use model download, plus engine: 26.78 MiB.
A filled area is an estimate, not recovered original content. Large objects, faces, text and structured backgrounds can look wrong. Transparency stays unchanged, so an erased item over transparent pixels is not made opaque. Browser decoding can normalize hidden RGB values in fully transparent pixels before processing. Pixels outside the selection are preserved from this decoded image.
PNG / JPG / WebP · One file · 10 MiB per file · 2048 px per side · 4,000,000 pixels · PNG output. First-use model download, plus engine: 26.78 MiB.
A filled area is an estimate, not recovered original content. Large objects, faces, text and structured backgrounds can look wrong. Transparency stays unchanged, so an erased item over transparent pixels is not made opaque. Browser decoding can normalize hidden RGB values in fully transparent pixels before processing. Pixels outside the selection are preserved from this decoded image.
Paint over a small distraction and let MI-GAN estimate the image underneath. The selected area receives generated RGB pixels; every unselected pixel and the original alpha channel are preserved.
Use small strokes with surrounding context. Select no more than half the image. The rectangle’s left/top coordinates refer to the physical image, also in RTL interfaces. Undo keeps the last 10 mask edits.
A blank or overly large mask cannot run. Clear it and make a smaller selection. If model loading or inference fails, retry after checking the connection or choose a smaller image.
A filled area is an estimate, not recovered original content. Large objects, faces, text and structured backgrounds can look wrong. Transparency stays unchanged, so an erased item over transparent pixels is not made opaque. Browser decoding can normalize hidden RGB values in fully transparent pixels before processing. Pixels outside the selection are preserved from this decoded image.
No. Your files and text are processed locally in this browser, without uploading your input. Engines and models download when needed. Clearing or leaving the page discards the current session; download or copy any results you want to keep.