Repair dust and selected areas of a film scan locally
Negative Converter includes a standalone AI repair brush and AI-assisted dust removal. Paint an area in Retouch and release the pointer to reconstruct it with MI-GAN. Photo processing runs on your device.
Brush repair or automatic dust removal?
Use the AI repair brush for a specific mark or damaged area you can identify yourself. Automatic dust removal first detects candidate spots and then repairs its selected mask. AI is the default repair option when dust removal is enabled; importing a photo does not automatically apply dust removal.
The shared MI-GAN model begins loading in the background after you open a photo. Wait for the model-ready status before painting. The initial model download requires a connection in the web app; inference then uses local WebGPU or WebAssembly according to device support.
How to use the repair brush
- Import and convert a scan, then open Retouch.
- Enable the AI repair brush and wait for the model-ready status.
- Set a brush size slightly wider than the defect. Zoom in and paint the area with a mouse or touch.
- Release the pointer and inspect the repaired texture. Undo an unsuitable result and try a smaller selection.
- Export the photo after processing completes. Saved repair selections follow the photo’s crop, rotation and mirror settings.
What does AI reconstruction change?
MI-GAN predicts plausible replacement texture from the surrounding image. It does not recover a hidden historical original and may invent detail. Be especially careful with faces, lettering, fine branches, wires, stars and documentary evidence. Preserve the source scan and compare the result at full resolution.
Automated spot detection can mistake real subject detail for damage. Inspect the mask and use manual selection when a detector cannot separate a defect from the scene.
Precision and export
Pixels outside the selected repair region retain their existing 16-bit values in the AI repair stage. The model operates on 8-bit input tiles, so reconstructed pixels are not newly recovered 16-bit sensor data. PNG and TIFF can carry the final 16-bit image; JPEG is an 8-bit format.
For groups of scans, batch conversion keeps each photo’s own repair selections. Copying a color treatment does not stamp one photo’s defect locations onto another.
Model and source
The shipped model is the MI-GAN pipeline from Picsart AI Research. Model provenance, license and checksum are documented in the project’s model directory. The application is published by NeoAnalogLab and its source is available for inspection.
Questions and answers
Does AI repair upload my film scan?
No. The model runs on the device. The web app downloads model and application assets, but does not send selected photo pixels to an inference server.
Can I use AI repair without automatic dust detection?
Yes. The standalone AI repair brush lets you paint the area you want to repair without running automatic detection.
Can I undo an AI repair?
Yes. Use Undo to restore the previous editing state. Keep the original scan when you need an unmodified archival copy.
Is the reconstructed detail guaranteed to be accurate?
No. Inpainting generates plausible content. Inspect the result and avoid treating generated texture as recovered evidence.