There’s a particular kind of photo most families have somewhere: a creased portrait of a grandparent, the colors gone orange, a white crack running across a face. For years, fixing it meant a pricey restoration service or hours in Photoshop. Now you can upload it to a free AI tool and watch it come back to life in seconds — and people are, in droves. The before-and-afters of restored parents and grandparents are some of the most-shared, most-tearful posts on the internet right now.
But there’s a catch that almost none of the “best AI photo restorer!” pages will tell you, and it’s the difference between a keepsake and a quiet little lie. So before the step-by-step, the one thing you need to understand about what AI is actually doing to your photo.
The one rule: it’s restoration, not recreation
Here’s what’s really happening under the hood. AI doesn’t recover the detail that was lost — it can’t; that information is gone. Instead it reconstructs a plausible version, guessing the most likely pixels from everything it learned during training. For damage like scratches, dust, and fading, that guess is excellent and faithful. For faces, it’s a gamble. Researchers have actually proven the trade-off: the more flawless and realistic an AI-restored image looks, the more likely its fine details are fabricated. A NeurIPS 2024 paper put a number on it — pushing for a photo-perfect result mathematically forces the uncertainty in the output to at least double the information that was missing. Stunning results are statistically guaranteed to contain invented content.
You can see this go wrong in the most heartbreaking way. In one widely shared case this June, a woman restored a damaged photo of her late father — and the AI handed back “a completely new, stern-faced stranger.” As one observer put it: it didn’t restore her dad; it replaced him. The academic version of the same warning is gentler but identical: an AI-restored photo is “a plausible, aesthetically pleasing reimagining of what the original might have looked like — not a forensic recovery of what it did look like.”
None of this means don’t do it. It means: do it with your eyes open, trust your own memory over the AI, and never throw away the original. Now the fun part.
How to restore an old photo for free (step by step)
First, scan it well. Garbage in, garbage out — and with AI, a blurry scan invites the model to invent more. Photograph or scan the print in good, even light (no glare), and save it as a high-quality PNG or TIFF at 600 DPI or higher. A clean, sharp input gives the AI less room to make things up.
Then pick your tool. Both are free; they just have different strengths.
Option A: Google Gemini (free, fastest, the crowd favorite)
Gemini’s image model — nicknamed “Nano Banana” — is what most people now reach for, because it’s fast and free.
- Go to gemini.google.com and sign in with a Google account.
- Open the image tools (look for Tools → Create images) and upload your scanned photo.
- Use a restrained prompt — this exact phrasing is what the community has settled on:
Restore this old damaged photo without changing any details. Repair the
scratches, tears, and fading and correct the color, but keep every face and
feature exactly as it is. Output in high resolution.
- It generates in seconds. Download with the image’s download icon.
Option B: ChatGPT (free tier, a bit slower)
- Go to chatgpt.com or the app, click the attach/image icon, and upload your scan.
- Use a prompt that explicitly forbids identity changes:
Restore this photograph: repair scratches, folds, and discoloration, and
correct fading. Keep the people, their faces, the lighting, and the colors
realistic and unchanged. Do not add new elements or alter the identity of
anyone in the photo — only repair the damaged areas.
- Give it about a minute, then download the result.
The golden rule from experienced restorers: don’t over-restore. Ask for a light touch, then compare the result to the original side by side. If the AI “improved” a face into someone who looks subtly off, run it again with stronger “keep the face exactly as it is” wording — or accept a slightly less polished version that’s actually them.
Where AI restoration genuinely shines
For everything that isn’t a face or fine detail, AI is fantastic and faithful. Peer-reviewed surveys confirm it’s reliable for:
- Scratches, dust, tears, and creases — removed cleanly.
- Fading and yellowing — corrected to natural tones.
- Black-and-white colorization — great on clothing, backgrounds, and objects (skin and eye color are more of a guess).
- Denoising and mild blur — sharpened without much risk.
- Moderate upscaling — roughly 2× to 4× larger looks good.
Where it struggles, predictably: very low-resolution originals (a face only a few dozen pixels wide becomes an invented one), heavily damaged regions, unusual angles, and several kinds of damage stacked together.
What this means for you
- If you just want a nicer copy to share or print: go for it — this is exactly what AI restoration is best at. Use Gemini for speed, keep the prompt gentle, and you’ll get a lovely refreshed image for the family group chat.
- If it’s the one irreplaceable photo of someone you’ve lost: be extra careful, and protect the original above all. Restore the background and damage, scrutinize the face, and if the AI changes the expression or features even slightly, keep the real, imperfect version. A cracked photo that’s truly them beats a flawless one that isn’t.
- If you’re tackling a whole shoebox or archive: batch the easy wins (faded landscapes, group shots, undamaged scans) where AI is reliable, and flag the close-up portraits for a slower, careful pass. Always save restored versions as new files alongside the untouched scans — never overwrite the original.
What AI can’t do here
- It can’t recover lost detail. It reconstructs a likely version. The information that faded away is genuinely gone.
- It can’t guarantee the face is theirs. This is the whole warning. The better it looks, the more you should double-check it against your memory.
- It can’t fix a truly tiny or destroyed face. Too few real pixels means a fully invented one.
- It isn’t a historical record. A restored photo is a beautiful reimagining, not evidence of exactly how someone looked. Treat it as a keepsake, not a document.
The bottom line
Free AI from Google Gemini or ChatGPT can bring a faded, torn, century-old photo back to life in under a minute, and for damage repair and color it does a genuinely beautiful job. Just remember what it’s really doing: rebuilding a plausible version, not recovering the truth. Keep your prompts gentle, keep the face honest, keep the original safe — and you’ll get the magic without quietly trading away the very thing you were trying to save.
Want to get confident with AI images beyond restoration — editing, generating, and fixing photos the right way? Our AI Photo Editing and AI Image Generation courses walk you through it from scratch.
