The short version: Gemini can repair tears, fill in missing bits, and add colour to old family photos if you give it precise, scoped prompts rather than a vague "fix this photo" instruction. The trick is working in small steps, telling it exactly what to touch and what to leave alone, and accepting that colourisation is always a guess dressed up as a result.
Why Gemini, and what it's doing
Google's Gemini app lets you upload an image and talk to it like an editor sitting next to you. Behind the scenes it's using the Gemini 2.5 Flash Image model, the one people nicknamed Nano Banana because of how it handles image generation and editing inside a chat. You don't need a separate app or a subscription to try it; the free version of Gemini will take a photo upload and let you ask for changes in plain English.
What it's not doing is "restoring" in the archival sense. A conservator restoring a photograph is working to recover what was physically there. Gemini is predicting what probably was there, based on patterns from millions of other photographs. For a crease across a plain wall, that prediction is reliable. For half a face that's been torn away, it's inventing a face that fits the statistical pattern of faces, not necessarily your great-uncle's actual nose. Nobody selling you a tutorial on this wants to say that part out loud, but it matters once you're printing the result and handing it to your mum.
A worked example: the 1958 doorstep photo
Say you've found a black-and-white photo of your grandmother standing outside a terraced house in 1958, creased hard down the middle, with a patch of damage near her left shoulder and general fading across the whole print. Here's how I'd break that into prompts rather than firing one big request at Gemini and hoping.
- Step 1, assess before you touch anything. Upload the scan and ask: "Describe the damage you can see in this photo, including creases, tears, fading, and any areas where detail is missing." This gets Gemini to tell you what it thinks it's looking at before it starts guessing fixes, and it's a useful check, if it misreads a shadow as a tear, you'll catch it now.
- Step 2, fix structure first. Prompt: "Repair the vertical crease running down the centre of this photo and the damaged patch near the woman's left shoulder. Do not alter her face, clothing colour, or the background details. Keep the repair realistic for a 1950s photograph."
- Step 3, address fading separately. Once the structural damage is sorted, run a second prompt: "Improve the contrast and reduce the overall fading in this photo without changing the composition or adding any new objects."
- Step 4, colourise last. Only once you're happy with the repaired black-and-white version: "Add realistic, natural colour to this 1950s photograph. Use period-appropriate tones, skin tones, and clothing colours typical for the era. Keep the result looking like a colour photograph from the time, not a modern filter."
- Step 5, sense-check with family. Before anyone treats the colourised version as the "real" photo, show it to whoever remembers the original scene. Ask specifically about the clothing colour and the house. This is the step almost everyone skips, and it's the one that stops you printing a fiction and calling it history.
Five prompts, not one. That's the actual number that matters here: trying to cram repair and colourisation and sharpening into a single instruction gives Gemini too many jobs at once, and it tends to overcorrect, smoothing out texture that should have stayed or adding colour casts that look more like Instagram filters than 1958.
Prompts for specific damage types
Different damage needs different wording. Vague prompts like "restore this" get you vague, over-smoothed results. Specific prompts get you specific fixes.
- Torn or missing corners: "Fill in the missing corner of this photograph using the surrounding pattern and lighting as a guide. Keep the rest of the image unchanged."
- Water damage or staining: "Remove the brown staining and discolouration across the top third of this photo while preserving the texture and grain of the original print."
- Faded faces: "Sharpen and clarify the facial features in this photo without changing the person's expression, age, or facial structure."
- Double exposure or scratches: "Remove the thin white scratches running across the image and the faint double-exposure ghosting in the upper left, keeping everything else exactly as it is."
- Low-resolution scans: "Increase the clarity and detail of this image without adding artificial sharpening artefacts or changing any facial features."
Notice the pattern: every prompt names the damage, states what to leave alone, and avoids open-ended instructions like "make it look better." Gemini, like every generative model, fills gaps with its best guess when you leave the door open. Close the door as much as you can.
Colourising: what you're really choosing
Colourisation is the part people enjoy most and trust least, often without realising it. When you ask Gemini to colour a black-and-white photo, it has no record of what colour your grandad's shirt was. It's choosing the most statistically plausible colour based on the era, the lighting, and typical patterns in photos from that decade. A navy shirt and a dark green shirt look nearly identical in black and white, and the model will pick one. It might pick wrong.
That's not a reason to avoid colourising old photos. It's a reason to treat the result as an interpretation, not a document. If a colour matters to the family story, like a specific dress or a car, say so in the prompt: "Colourise this photo. The car in the background was red, a Ford from around 1955." Gemini will use that detail rather than guessing. For everything else, it's filling in plausible colour, and that's fine as long as everyone looking at the final print knows that's what happened.
If you want a side-by-side sense of how different free tools handle this exact trade-off, there's a fuller comparison on where to colourise old black and white photos with AI free, which is worth reading before you commit a whole shoebox of prints to one approach.
Working the free tier without wasting your photos
Google doesn't publish a hard daily cap for image editing on the free Gemini tier, but in practice expect the quality or speed to drop off if you push through a long batch of edits in one sitting. Work in smaller rounds: pick five or six photos, run your repair and colourise prompts on those, export them, then come back later rather than trying to clear an entire family archive in one go. It also means you're checking results as you go rather than discovering twenty photos later that Gemini decided your great-grandfather's cottage had a different roof.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
A few practical habits that save you re-work:
- Scan at the highest resolution your scanner allows before uploading anything, since Gemini can only work with the detail that's in the file.
- Keep the original scan untouched and always work on a copy, so you can go back if a prompt overcorrects.
- Ask for one change at a time on anything that matters, a wedding photo, a photo of someone who's passed away, a photo that's going into a memorial slideshow. Save the one-prompt-does-everything approach for snaps that don't carry weight.
- Export at the resolution Gemini gives you, then check it against the original before printing. Upscaling a result that's already slightly soft will only make the softness more obvious.
When Gemini isn't the right tool
For heavily damaged photos, physically torn into pieces, mould damage, or prints where most of the surface detail is gone, Gemini's guesswork gets less reliable the less original information it has to work from. In those cases a specialist restoration service, or a human retoucher working frame by frame, will do a more faithful job than any prompt, however carefully worded. There's a rundown of options, free and otherwise, including where the AI tools hold their own, at where to restore damaged old photos with AI for free, which is worth a look if you're dealing with anything beyond creases and fading.
The uncomfortable bit nobody writing these tutorials wants to put in a heading: some photos should stay damaged. A tear across a photo of a grandparent who died young isn't always a flaw to erase, it's part of the object's history, and a perfectly smoothed AI version can feel less honest than the creased original, even if it looks sharper. Worth deciding that on purpose rather than by default.
I have written more around this on the site: Where to Restore Damaged Old Photos With AI for Free, Where to Colourise Old Black and White Photos With AI Free.
For the rest of the series, see the ChatGPT Prompts and Limits: 30 Guides to Fun Prompts, Photo Edits and Daily Caps. To write your own, try the free tool that writes the prompt for you.
Frequently asked questions
Can Gemini restore a photo for free?
Yes, the free version of the Gemini app lets you upload a photo and ask it to repair damage or add colour directly in the chat, no separate account or payment needed to try it.
Does Gemini know the real colours in an old black and white photo?
No, it estimates plausible colours based on patterns from similar photos and the era you specify, so treat any colourised result as an informed guess rather than a factual record.
Why does Gemini sometimes change a face when restoring a photo?
When part of a face is damaged or missing, Gemini fills the gap with a statistically likely face rather than the actual person's features, so always prompt it to leave facial structure unchanged and check the result against other photos of the same person.
What's the best way to prompt Gemini for old photo restoration?
Break the job into separate prompts, assess the damage first, fix structural issues like tears and creases next, handle fading separately, and colourise last, naming exactly what to change and what to leave untouched each time.
