AI Photo Restoration

AI photo restoration repairs the damage time does to a photograph: fading, blur, grain, cracks, tears, stains and the soft focus of an old camera. Instead of retouching every scratch by hand, you let models rebuild the missing detail and the lost tone, and you stay in charge of how far they go.

On Mage, restoration is a short workflow rather than one button. Image Enhancer upscales a scan to 2K, 4K or 6K and refines its detail. Refine fixes faces and hands with its Fix Eyes, Fix Hands and Add Detail presets, or with an instruction you write. Inpaint repairs a tear or a stain you paint over. And an edit model such as Nano Banana 2.1 or Mango 3 can repair or colorize a whole photo from a plain-language instruction.

Be clear about one thing: AI restoration reconstructs detail it cannot see. A face that was a blur of a few dozen pixels comes back sharp because the model invents a plausible face, and plausible is not always the person. Compare every result with the original, keep the passes that stay true to it, and treat the restored photo as an interpretation, not a record.

Order matters. Start with Image Enhancer at a moderate size to clean up grain and softness and give the later tools more to work with. Repair physical damage next, while the photo is still close to its original state. Fix faces after that, colorize if you want to, and save the largest upscale for the very end, so you enlarge the finished image rather than every intermediate flaw.

Change one thing at a time. A single pass that sharpens, repairs, colorizes and modernizes a photo all at once gives the model the most room to drift. Separate passes let you see exactly which step changed a face, and go back one step when it does.

• Scan at high resolution, or photograph the print flat with no glare.
• Enhance first at 2K or 4K; save 6K for the final image.
• Repair tears and stains before touching faces.
• Keep the original open and compare after every pass.
• If a face stops looking like the person, step back and try a lighter pass.

For a single crack across the sky or a stain on a sleeve, Inpaint is the precise tool. Paint a mask over the damage only, describe what should be there ('clear sky', 'the same grey wool sleeve'), and the rest of the photo is left exactly as it was. Small masks on plain areas restore almost perfectly.

For damage spread across the whole photo, a prompt edit is faster. Attach the scan to an edit model and write the job plainly: 'Restore this photo: remove the scratches, dust and creases, and keep every face, the clothing and the composition exactly as they are.' Naming what must stay matters as much as naming what to fix. Missing corners and torn-off pieces can be rebuilt the same way, but anything the model fills in is a guess, so keep those areas simple.

Mage has no colorize button. You colorize by asking an edit model, and that turns out to be an advantage, because you can tell it what you know. 'Colorize this 1950s photo. Her dress was navy blue, the car is pale green, natural skin tones, soft daylight' gets you closer to the truth than an automatic guess.

Color is the step most likely to look modern and overdone. Ask for muted, period-appropriate tones and natural skin, and compare a colorized version with the original black and white before you print it. For portraits, a light pass of Skin Enhancer afterwards brings back texture, and the AI image upscaler guide covers the final enlargement.