Bring damaged photos back to full detail.
Studio-grade restoration, running locally.
Nothing to install and no account to create. The models run on your machine.
Six tools in one studio
Restoration, sharpening, black-and-white colorization, gap filling, ID photos, and subject-aware crop — the full repair workflow on a single page, with batch processing for the modes you run most.
100% private
Family photos are personal. Everything runs on your device — no uploads, no servers.
WebGPU acceleration
Inference uses your GPU when one is available and falls back to WASM everywhere else, automatically.
Real models, not filters
Swin2SR super-resolution, GFPGAN face restoration, LaMa inpainting, DDColor colorization, BiRefNet and MODNet matting — all local, MIT/Apache licensed.
From damaged photo to download.
- 01
Drop in a photo
Drag and drop, paste, or browse. JPG, PNG, or WebP — scans and phone snapshots alike.
- 02
Pick a mode
Restore faded photos, sharpen blur, colorize black-and-white shots, brush out objects, swap ID backgrounds, or crop to the subject.
- 03
Compare, then download
Slide between before and after, then export a full-quality PNG. Models stay cached for the next run.
Restore
Scratches, noise, and faded color repaired
Sharpen
Swin2SR 2×/4× super-resolution
Colorize
DDColor adds natural color to B/W
Fill
LaMa inpainting removes objects
ID Photo
Cutouts and passport sizes
Crop
Subject-aware framing
About the photo restorer
A free AI toolkit for restoring old and damaged photographs: scratch removal, blemish repair, black-and-white colorization, face enhancement, and upscaling in one place, running fully on your device. It is designed for digitized family archives, historical scans, and damaged prints where uploading sentimental images to a third party feels wrong — and usually costs a per-photo fee anyway.
Six modes cover the full workflow — Restore, Sharpen, Colorize, Fill, ID Photo, and Crop — and the three heavy-lifting ones (Restore, Sharpen, Colorize) also run as a batch queue, so a shoebox of scans can be processed unattended.
How it works
The toolkit chains several on-device models. Swin2SR super-resolution denoises and enlarges in overlapping 512 px tiles, which is how inputs up to 4096 px stay inside browser memory limits. GFPGAN restores portrait faces, with a 0–100 % strength slider that blends its output against the sharpened original; at 0 % the 340 MB model never even downloads. A LaMa-style inpainting network fills scratches, tears, and dust spots you brush over, and DDColor colorizes black-and-white photos by predicting their color channels while keeping the original luminance untouched. WebGPU is used whenever the browser supports it, with an automatic WebAssembly fallback; weights download once per model and are cached.
Colorizing black-and-white photos
Colorize runs DDColor-tiny, an open model from the DDColor project (Apache-2.0 licensed). The graph works at 512×512 in Lab color space: it predicts the two chroma channels (a/b), while the lightness channel comes straight from your photo. Predicted colors are resampled back to the original resolution, so output detail and sharpness are the scan’s own — the file you export has exactly the dimensions of the file you dropped in. Colorization is free at full resolution on every plan, with no watermark; Pro’s Restore advantages live in input size and batch size, not colorization.
Photos that already look color are detected up front on a small thumbnail, and the run button is disabled — colorizing a color image would only shift its colors. If a scan is merely faded rather than black-and-white, Restore mode’s auto-levels revive saturation without invoking the color model.
Brushing out scratches & damage
In Restore mode, switch on “Repair scratches & damage” and paint the defects with the brush; an eraser tool refines the mask and Clear starts over. On run, masked areas are filled by LaMa first, then the whole image passes through denoising/super-resolution, optional face restoration, and auto-levels. The mask only guides that first inpainting pass, so scratches disappear without the rest of the photo being repainted. Fill mode offers the same engine for removing whole objects — people, wires, watermarks — without the restoration chain afterward.
Batch processing
The Single/Batch switch queues files for Restore, Sharpen, or Colorize — the modes without per-image brushwork. Files are processed strictly one at a time, because parallel inference is the quickest way to exhaust browser memory, and every file uses a snapshot of the current settings: scale, face toggle, and strength. Free plans queue up to 4 images and Pro up to 16. Results download individually or as one ZIP, encoded PNG by default; single-image mode additionally offers JPEG/WebP with a quality slider. One corrupted file fails on its own and can be retried without restarting the queue. Fill, ID Photo, and Crop stay single-image because they depend on per-photo interaction.
Limits & requirements
Restore/Sharpen accept up to 2048 px on the long edge for free accounts and 4096 px on Pro; larger drops are pre-scaled with an on-screen notice. Colorize and the scratch-repair brush have no resolution gate and always run at the photo’s full resolution, free. Model weights are sizable — Swin2SR ~200 MB, GFPGAN ~340 MB, LaMa ~208 MB, DDColor ~129 MB — and each downloads only the first time its feature runs, then stays cached in the browser. Chrome or Edge 113+ use WebGPU; Safari and Firefox run WebAssembly, which is slower but produces the same output. There is no file-size cap and no watermark, and both free and Pro output is licensed for commercial use.
Privacy
Family photos can be deeply personal. Here they never leave the browser: no upload, no server-side processing, no retention policy to worry about.