# Old Photo Revive

> Free, on-device old photo restoration at https://old-photo-revive.skillsafe.ai/ . Open a scanned or phone-captured old photo and repair it in the browser: remove dust and specks, clean film grain and blur (optionally enlarge 2x or 4x), fix fading and colour casts, brighten under-exposed shots, add local contrast, and colourise black-and-white photos. Before/after slider, PNG/JPG download. Nothing is uploaded; no sign-up; no credits.

## What each step does
- Dust & specks: pixels far brighter or darker than the median of their 5x5 neighbourhood are refilled from clean neighbours (strength slider).
- Grain & blur clean-up: Real-ESRGAN general x4 v3 (BSD-3-Clause) runs tile by tile with onnxruntime-web (WebGPU, WebAssembly fallback); output is kept at the same size or enlarged 2x/4x; a strength slider blends with the original.
- Fading & colour cast: each RGB channel is stretched to its own 0.5-99.5 percentile range.
- Brighten: a luminance gamma chosen from the photo's mean brightness; already-bright photos are untouched.
- Local contrast: luminance plus a share of its difference from a wide box blur.
- Colourise: DeOldify "artistic" (MIT) predicts colour from a grey copy at 256/384/512 px; only the chroma is used, the photo's own full-resolution luminance is kept. Colour is a plausible guess, not a record.

## Limits
- Photos are worked on at up to 3000 px on the long side; enlarged output is capped at 4096 px.
- Models download once from models.skillsafe.ai (5 MB clean-up, 255 MB colour) and are cached; every file is SHA-256 checked.
- For blurry faces use https://face-restore.skillsafe.ai/ on the result.

## Credits
- Real-ESRGAN by Xintao Wang et al., BSD-3-Clause.
- DeOldify by Jason Antic, MIT; ONNX export by the facefusion project.
- onnxruntime-web by Microsoft, MIT.
- Example photo: State Library of Queensland, c. 1925, public domain.