How to restore an old photo
Drop the scan or a phone photo of the print on the page. Check the outline and press Next, check the marked cracks and press Repair, press Enhance to restore faces and enlarge, colourise a black-and-white photo if you like, then save. Every step starts automatically and can be corrected by hand.
Frame
The print's edge is found inside a white mount, a scanner lid or a table, and cropped. A print shot at an angle, or one that bows or curls, has four corner handles and four edge handles: drag them onto the print's edges and it comes out flat and square.
Repair
Cracks, scratches, dust and spots are marked in colour. The Find slider changes the marks at once. Paint roughly along missed damage with Mark (it snaps to what stands out), click a wrong mark with Unmark to take it off, undo when needed, then press Repair. The marked areas are filled by an AI model from what is around them.
Enhance
Faces are rebuilt sharp with GFPGAN, and the photo enlarged two or four times with Real-ESRGAN, which also takes out grain and blur. Keep grain for a lighter touch.
Colour
Shown when the photo is black and white or sepia. DDColor guesses likely colours: Best colours (231 MB) holds up on grainy and busy photos, Fast (56 MB) is a quick download. Each pixel keeps the photo's own lightness, so no detail is lost.
Finish
Turn on Auto fix to stretch a faded print back to full contrast and take out the yellow or sepia cast. Face restoration starts at half, which keeps more of the person; set how much to keep, adjust the light and colour, and save JPG, PNG or WebP.
The photo check
As soon as a photo opens, it is checked, and the result says how damaged it looks and what will help:
| Result | What it means |
|---|---|
| Looks clean | Little or no damage found. Repair can be skipped |
| Light damage | Dust, spots and short scratches. Repair takes care of them |
| Damaged | Cracks or long scratches. Check the marks before repairing |
| Heavily damaged | Cracks across the photo, often across a face. The repair rebuilds what is missing from what is around it, so large areas and faces may not come back exactly as they were |
It also says when a print is faded or yellowed (Finish's auto fix), black and white (the Colour step appears) or small (worth enlarging). Busy scenes, such as gravel, leaves or branches, can look damaged to the check: the marks in Repair show what it found.
Why this order
Cropping first means the slow AI steps only work on the photo. Repairing before enhancing matters most: an enlarger sharpens whatever it sees, cracks included, and a face model shown a cracked face rebuilds the cracks as features. Filled first, the cracks are gone before the face is rebuilt. Colour and tone come last, so they are set on the finished picture.
What it cannot do
- A torn-off corner or a face mostly lost is made up, not brought back
- Glare from a phone photo of a glossy print cannot be removed: photograph it again in soft light from the side, or scan it
- Fine dust on a busy scene and faint cracks across a light face may not be found: paint them with Mark
- A print that curls in more than one direction straightens only along its edges
Where does the photo go?
It stays on your device. The AI models download from this site once and run right in your browser, on your graphics card if it can, otherwise on your processor.
- Downloaded once, then kept in your browser for next time
- The same model files as the image upscaler, remove object and colouriser: one download serves all of them
- Saved without metadata like location or camera
Specs and browser support
| Input | JPG, PNG, WebP, GIF (first frame), BMP, AVIF, up to 24 megapixels |
|---|---|
| Output | JPG (92 % quality), PNG or WebP, up to four times the size of the photo and 64 megapixels |
| Frame | Border found on a 640-pixel copy; perspective with a bend for each edge (a Coons patch); no download |
| Repair | Cracks and scratches found by the scratch detector of Microsoft’s Bringing Old Photos Back to Life (MIT), 38 MB, dust by grey-scale top-hat filters, eyes kept safe with YuNet; filled with MI-GAN (MIT), 28 MB, in 512-pixel tiles |
| Enhance | Real-ESRGAN general x4v3 (BSD-3-Clause), 5 MB; GFPGAN 1.4 (Apache-2.0) and YuNet, 86 MB |
| Colour | DDColor (Apache-2.0): large, 231 MB, or tiny, 56 MB; only for black-and-white photos |
| Runtime | onnxruntime-web (MIT), 14 or 17 MB, once |
| Browsers | WebGPU in Chrome and Edge. Other browsers use the processor. A computer with 8 GB of memory is best for 4× results |
| Price | Free, no watermark, no account |
Questions
It fills cracks, scratches, spots and small missing areas from what is around them, and it rebuilds faces. A torn-off corner or a face mostly lost cannot come back as it was: the AI makes up something likely, not the real thing.
No. The AI models are downloaded to your browser and run there. Your photo stays on your device.
When you open a photo it is checked for cracks and scratches, fading, a colour cast and its size. It says how damaged the photo looks, from clean to heavily damaged, and which steps will help most.
Face restoration draws a sharp face from the blurry or damaged one. The more damaged the face, the more it has to guess. Lower Face restoration in the last step to keep more of the original face.