Restore Old Photos With AI: A Practical 2026 Guide (No Hype)

You have a box of your grandparents' photos. Yellowed edges, a crack running across a face, the dull black and white of a print developed fifty years ago. Five years ago this was the work of a retoucher doing it by hand and charging by the hour. Today an AI does it in thirty seconds. The question isn't whether you can —you can— it's how true what it gives back to you really is.

This guide is about that: how to restore old photos with AI so the result looks like your family and not like an invented version. Real tools, a step-by-step workflow, prompts that work, and the part almost nobody tells you: when the AI is repairing and when it's lying.

What AI Can Do With an Old Photo (And What It Can't)

Here's the distinction that decides everything else. In 2026 two schools of restoration coexist, and confusing them is the number one cause of disappointing results.

Faithful restoration. Tools like Topaz Photo AI or Adobe Enhance start from the idea that the detail was already there, just hidden under noise, blur or low resolution. Their models are trained on millions of image pairs (clean ↔ degraded) and try to recover what the camera actually captured. They invent little. The result is more restrained, but faithful.

Generative restoration. Remini, Nano Banana or Magnific work the other way: they hallucinate new detail that's plausible given what they see. They invent skin pores, hair strands, fabric texture that were never in the original. The result is spectacular —and dangerous—: it dazzles on screen but may not represent the real person.

What does work well with either school:

• Removing scratches, dust, stains and creases.
• Reducing the grain and noise of old prints.
• Fixing yellowing and recovering contrast.
• Increasing resolution (upscaling) to print large.
• Reconstructing small missing areas.
• Colorizing black and white.

What it can't do, no matter what the marketing says:

• Know what was in a destroyed area. If half a face is missing, it doesn't recover it: it imagines it.
• Remember a person it never saw. It doesn't know the real eye color in a B&W photo; it bets.
• Bring back detail of something never recorded. There was no information there; what it adds is invention.

The Best Tools for Restoring Old Photos

There's no single winner. Each one shines at a different part of the job. Here's what people actually use in 2026.

Nano Banana (Gemini) — the most versatile by instructions

Google's image model, built into Gemini, is the most convenient for restoring by talking to it. You upload the photo and ask in natural language for what you want: remove scratches, recover detail, colorize. Its Pro version (Nano Banana Pro, on Gemini 3) reconstructs with surprising coherence. Access through the Gemini app; the powerful features come with the Gemini Pro subscription (around $20/mo). It's generative: flashy, but watch the fidelity of faces.

Remini — fast and built for mobile

The most popular app for instant "before and after". It repairs faces in old photos in seconds from your phone. It's generative to the max: it produces sharp, dramatic results that can drift from the original. It has free daily credits (with a watermark on export) and a paid weekly subscription to remove limits. Perfect for a fast first pass; less reliable if you need accuracy.

Topaz Photo AI — the faithful option for printing

The standard for photographers. It excels at noise reduction and upscaling without inventing too much. It includes a dedicated face-recovery module. Since late 2025 it's subscription only (the perpetual license ended); the price points to serious use, not a single photo. If you want fidelity and you're going to print large, it's the safe bet.

Magnific — extreme creative upscaling

The most aggressive upscaler: it multiplies resolution and re-imagines the image with "creativity" sliders. Brutal for digital art; risky for family photos because it invents eagerly. It's premium, with no free plan. Use it only when historical realism doesn't matter.

GFPGAN and CodeFormer — faces for free (if you're a bit technical)

Two free, open-source models specialized in repairing faces. GFPGAN is blazing fast and gives natural faces; CodeFormer preserves identity better. They run locally or on services like Replicate. The best value that exists for the most delicate part —the face— if running things doesn't scare you.

Google Photos — the free option you already have

Since 2025, Photo Unblur and Magic Eraser (erase objects) are free for everyone, not just Pixel. Careful: they are not damage-restoration tools. They don't remove scratches, colorize or recover a degraded portrait. They're for sharpening a blurry photo and little else. Useful as a single step, not as a complete solution.

Step by Step: How to Restore an Old Photo

The order isn't optional. Doing it backwards drags defects into the later stages.

1. Scan it properly (this sets the ceiling)

The result will never be better than your scan. Use a scanner at 600 dpi minimum or, if you're using your phone, diffuse light without reflections, the photo flat and a straight frame. A mediocre digitization forces the AI to invent more, and that's where it slips up. Clean the dust off the glass first.

2. Upscale (raise the resolution)

More pixels first, then the rest. Run the image through Topaz (faithful) or Magnific (creative, with care). With more resolution, the following steps have material to work with. Going 2x or 4x is fine; multiplying by more usually starts inventing texture.

3. Clean up damage and noise

Scratches, stains, creases, grain. This is where Remini (fast) or Nano Banana with precise instructions come in. To erase an object or a large stain, Google Photos' Magic Eraser. Remove it before touching the face and before colorizing: any noise you leave here propagates.

4. Repair the face (with judgment)

The sensitive part. Run it through GFPGAN or CodeFormer, or Topaz's face module. Compare the result with any other photo you have of the same person. If the AI changed the apparent age, the shape of the nose or the gaze, lower the intensity or go back. Better a face that's a little less sharp than a face that isn't theirs.

5. Colorize (if it's black and white, and last)

Last step, always. Colorizing before cleaning tints the noise. With Nano Banana, ask for historically realistic tones and natural skin. Accept that the colors are an educated guess, not a fact: the dress might have been a different color. If you have references (a family memory, another photo), tell it so in the prompt.

Prompts That Work (For Instruction-Based Tools)

With Nano Banana and similar, the prompt rules. These are task-specific and ready to copy.

General restoration for light or medium damage:

Restore this old photo: remove scratches, dust and creases, reduce grain and noise, fix yellowing and recover contrast. Keep the original features and composition EXACTLY as they are. Do not add or remove elements. Realistic, not beautified.

Face repair while preserving identity:

Recover the face detail in this photo: define eyes, skin and hair with natural texture. Preserve the EXACT identity, age and expression of the person. Do not smooth or rejuvenate. If an area is too damaged to reconstruct faithfully, leave it without inventing.

Colorize black and white with believable tones:

Colorize this black and white photo with historically realistic, period-appropriate tones. Natural, warm skin tones, muted and desaturated colors like a real print, not saturated. Do not change any detail of the image, only add color.

Severe damage (use it knowing it will invent):

Reconstruct the torn areas and missing edges of this photo coherently with their surroundings. Keep every area that IS intact untouched. Tell me conceptually which parts you had to reconstruct from scratch.

The Limits and the Ethics (The Uncomfortable Part)

This is where an honest guide parts ways with a "look at the magic" tutorial. Restoring photos with AI has a fundamental problem: the model doesn't know who it's restoring.

When you repair your grandfather's face in a badly damaged photo, the AI doesn't recover his face —it doesn't know it— it draws a plausible face from the millions it has seen. On light damage it's barely noticeable. On heavy damage, what it gives back might be a distant relative of your grandfather, not your grandfather. Sharper, yes, but a different person.

This isn't theoretical. Generative tools tend to:

• Beautify: smoother skin, symmetry that wasn't there.
• Rejuvenate: they erase wrinkles and years without you asking.
• Homogenize: faces that resemble each other because they come from the same model.

The practical rule: if you can't verify a detail against another real photo of that person, don't accept it just because the AI made it look nice. For a family archive photo, fidelity is worth more than sharpness. For a creative project, do whatever you want —but knowing it's interpretation, not a document.

And a simple legal note: restoring your own family photos is no problem. Restoring and publishing other people's photos, or using reconstructed faces as if they were real, moves into image-rights and honesty territory. Common sense.

Who Is This For?

It's more than enough for you if you want to rescue the family album, print an old portrait large, colorize a photo of your grandparents as a gift or clean up a blurry image. With free tools (Google Photos, GFPGAN, Remini credits) you cover almost everything without spending.

Make the jump to paid if you restore many photos, need professional print fidelity (Topaz) or want fast batch repair and color (Nano Banana on Gemini Pro, paid Remini).

Think twice if the photo is a document someone will take as historical truth. There, generative AI is a dangerous tool: it makes the invented look real. Restore conservatively, note that it's been retouched, and always keep the original scan untouched.

The honest promise isn't "AI brings your photos back to life". It's: AI cleans the damage time did, and fills in what's missing with an educated guess. If you use it for the first, it's an extraordinary tool. If you mistake the second for a miracle, you end up with a precious memory of someone who never existed.