BROWSER-BASED IMAGE INPAINTING
For visible marks on AI-generated images you own or may edit, paint over the mark and use Apply to all when it appears at the same relative position in each image. EraseKit does not detect every mark or remove invisible SynthID provenance.
Use this editor only on images you own or are authorized to modify, including your own AI-generated images. This independent tool is not affiliated with or endorsed by Google, Gemini, or Nano Banana. It edits visible pixels only; it does not remove SynthID or other invisible provenance data.
A MANUAL VISIBLE-MARK WORKFLOW
EraseKit does not automatically identify Gemini or Nano Banana marks. You choose the visible region and review the generated fill.
Add up to 30 PNG, JPG/JPEG, or WebP files in one session. The model may need to download and initialize in your browser on the first edit.
Select one image and use the brush or rectangle controls to cover the whole mark. Include nearby edges only when they need reconstruction too.
Choose Apply to all only when the mark occupies the same relative position in the other images. Otherwise edit each image separately, then inspect and download the results.
Apply to all repeats your painted coordinates proportionally across the batch; it does not find a watermark in each image. Changes in crop, aspect ratio, composition, or placement can make that mask miss the mark or cover unrelated details. Separate edits are safer when layouts differ.
A visible overlay is part of the displayed pixels and can be selected for inpainting. SynthID and other provenance mechanisms are not visible overlays that this editor can erase. Editing visible pixels does not promise to remove or preserve provenance information or file metadata.
Image processing happens locally in your browser; the website and model resources still load over the network. EraseKit uses ONNX Runtime Web with WebGPU when available and a WebAssembly fallback. Model initialization, image size, device memory, and browser support affect startup and batch time. Large selected areas are processed at reduced resolution and scaled back, so detail can soften.
Inpainting estimates what could fill the selected area; it cannot recover hidden pixels. Complex textures, faces, text, fine lines, and marks crossing several edges can leave artifacts. Try a tighter selection where appropriate, process images individually, or use a conventional editor for precise reconstruction.
No. You manually select a visible mark; Apply to all repeats that selection at corresponding relative coordinates.
No. The tool edits selected visible pixels and does not remove invisible SynthID or guarantee changes to provenance metadata.
Yes, for up to 30 images in a session. It works best when the mark is at the same relative position in similarly framed images.
PNG, JPG/JPEG, and WebP. PNG downloads remain PNG; JPEG and WebP are re-encoded in their respective formats.
No. The model generates an estimate, and complex backgrounds or large selections can produce visible artifacts.
