Restore Old Photo cleans up a damaged print: scratches, dust spots, small tears, noise and fading. Google's Gemini image model gets the photo with one instruction — repair the damage, fix fading and contrast, and keep the people's identity and the composition unchanged.
Because the repair is generative, the model does not just patch pixels; it redraws the picture. That is why it can close a crease convincingly, and also why it can smooth away texture or subtly change a face, especially one that is small or badly damaged. The result comes back at about one megapixel and is not coloured in — for colour, run Colorize B&W afterwards.
How to use
Scan or photograph the print flat, in even light and without glare — the model cannot repair what a reflection hides.
Sign in, add the image and start the job (3 credits).
Compare the result with the original at full size, especially faces and hands.
Download it, and keep your original scan as well.
FAQ
Will my grandparents still look like themselves?
The model is told to keep identity unchanged and usually manages that on clear faces. On tiny or heavily damaged faces it has to invent more, and the likeness can drift. Always compare with the original before you print or share.
How badly damaged can the photo be?
Scratches, dust, spots, light creases and fading are the typical cases. Large missing areas — a torn-off corner, a face lost to water damage — get filled with something plausible, not with what was really there.
Does it also make the photo bigger or add colour?
No. The result is about one megapixel and stays black-and-white if the original is. Use Upscale & Enhance for size and Colorize B&W for colour — each is a separate run with its own credit cost.
Is the photo uploaded, and what does it cost?
Yes, it is a cloud tool: the image goes encrypted to Google's Gemini API on a paid account (never the free tier, where data may be used for training) or, if that is down, to a backup provider. 3 credits per photo, refunded automatically if the run fails; you need an Abzar account.