AI spot clean
Paint a mask over a distraction and let MI-GAN fill it from the surrounding texture, colour and structure.
AI retouch · InpaintingLocal-first
AI Retouch & Object Removal
Remove distractions and rebuild a scene with MI-GAN running locally through ONNX Runtime WebAssembly.
The Remove brush is not a clone stamp with extra steps. Paint over an object, release, and MI-GAN 512 reconstructs the area using the surrounding scene. The model is a 29.5 MB ONNX network that downloads once and then runs in the tab.
Small distractions — a sign, a cable, a stray guest — are the sweet spot. For larger removals, Scene Rebuild repaints a wider selection and blends the result through a feathered mask. If the neural model cannot load, OpenCV Telea inpainting is available as a deterministic fallback.
No image uploads WebAssembly engine Works in the tab
What you can do
AI retouch, in detail.
Remove objects and rebuild backgrounds in your browser with MI-GAN 512 and ONNX Runtime WebAssembly. Paint a mask, run local AI inpainting, and keep every pixel on your device.
Scene rebuild
Select a larger region and repaint it with a coherent background rather than stretching nearby pixels.
Feathered compositing
The generated patch is blended through the original mask with a soft edge, so the repair does not announce itself.
OpenCV fallback
If the ONNX session cannot start, Telea inpainting still removes small marks without a model download.
Undoable edits
Every repair is a normal pixel edit in the history stack. Undo restores the exact previous pixels.
Runs offline after caching
The model is cached by the browser. After the first download, spot cleaning works without a network connection.
How it works
From file to finish in three moves.
No install step and no import queue. Open the editor, work on the document locally, and export the file your next tool expects.
Try it now- 1
Paint the distraction
Choose the Remove brush, size it to cover the object with a little margin, and paint a mask.
- 2
Release to run
The editor crops a square context region, runs MI-GAN at 512 × 512 and composites the result back through the mask.
- 3
Check and refine
Zoom in, compare with undo, and repeat on any edge the model did not resolve cleanly.
Who it is for
Built for real photo work.
These are the workflows the tool is designed around — not a list of theoretical features.
- Removing exit signs, cables and light stands from wedding reception photographs.
- Cleaning small distractions from real-estate and product images.
- Repairing a background after removing a subject from a group photo.
- Quick event-delivery cleanup without sending images to a cloud AI service.
Questions
AI retouch FAQ
Straight answers about formats, privacy and limits.
Is the AI running in the cloud?
No. MI-GAN runs through ONNX Runtime WebAssembly inside the browser tab. The image and mask never leave the device.
How large an object can it remove?
MI-GAN works best on small and medium regions with usable surrounding context. Very large removals can produce repeated texture or invented structure; Scene Rebuild and manual retouching may be better for those.
Does it work on RAW files?
Yes. Decode the RAW first, then run the Remove brush on the document. The repair is applied to the editable layer and can be exported at full resolution.
What are the model licensing terms?
MI-GAN code is MIT-licensed, but the published model was trained on Places2, whose terms are research/education-oriented. Review THIRD_PARTY_NOTICES.md before commercial redistribution of the model.
Put ai retouch to work.
Open the editor and bring a real file. Your photo stays on your device.