How to Restore Old Family Photos with AI — A Step-by-Step Guide

To restore an old family photo with AI in 2026: (1) scan or photograph the original at the highest resolution, (2) remove damage with Nano Banana Pro, (3) colorize if needed with NBP or a specialist, (4) enhance the face with Topaz or NBP, (5) upscale for printing with Crystal. The whole workflow takes about 10 minutes per photo and costs roughly $0.50–$2 in API spend.

This guide is written for the photo in the shoebox under the bed — the cracked wedding portrait of grandparents you barely remember, the faded color print from a 1970s birthday, the only surviving picture of an aunt who died young. These photos matter. They deserve a careful workflow, not a one-click filter. The five steps below are the same ones working restoration photographers use in 2026, ordered to do the least damage to the original and the most justice to the subject. The full process takes about ten minutes per photo once you have the tools open, and the API spend lands between fifty cents and two dollars for a complete restoration. A working photographer offering this as a service typically charges between twenty-five and one hundred dollars per restored photo — the math is generous on either side of the relationship.

Why do old photos fade and tear?

Before we restore anything, it helps to understand what we are restoring against. Old photos do not "just get old" — they are slowly destroyed by five specific things, and knowing which one happened to your photo tells you which step in the workflow matters most.

Silver halide oxidation is the chemistry behind black-and-white prints. The image is made of microscopic silver crystals in a gelatin layer, and over decades those crystals oxidize and shift toward a yellowish or brownish tone. This is the soft, warm fade you see on a 1940s portrait. It is also why most really old prints have that distinctive sepia look — much of that is not the photographer's choice, it is chemistry.

UV exposure from sunlight is the single biggest accelerator of fade, especially for color prints made between roughly 1960 and 1990. The cyan, magenta, and yellow dye layers fade at different rates — magenta is the most fragile, which is why old color photos go yellow-green as they lose magenta first. If your color print looks like an Instagram filter from 2011, that is decades of UV damage, not artistic intent.

Acidic paper is the slow killer. Photo prints made before the 1990s often used non-archival paper that contains residual processing chemicals and lignin from wood pulp. Those acids attack the image layer from inside the print itself. This is the cause of the strange brown spots and overall yellowing on the white borders of an old photo.

Humidity and temperature swings cause physical damage — the gelatin layer expands and contracts, and over decades it cracks. This is the source of those fine, branching surface cracks you see on prints stored in attics or basements. Humidity also feeds mold, which causes the dark spotting called "foxing."

Fingerprint oils, water stains, and physical tears are the visible damage everyone notices first. Skin oils contain acids that etch into the emulsion over time, which is why old prints often have ghostly fingerprint marks. Tears and creases happen when prints are moved, stored, or rescued from a flood.

The good news is that 2026 AI tools can address all five categories — but each requires a slightly different prompt and tool. The damage-removal step below covers them in order of difficulty.

Step 1: Scan or photograph the original at the highest resolution

The single most important rule of photo restoration has nothing to do with AI: never work on the original. You scan once, you save the scan as a master file, and from that point on every step of the workflow happens on a copy. The physical print goes back into an archival sleeve and stays out of the sun.

You have two scanning options in 2026, and both are good enough.

The flatbed scanner option is the gold standard. Use 600 DPI for prints up to 4x6 inches, and 400 DPI for larger prints — that gives you enough resolution to upscale convincingly later without bloating the file pointlessly. Scan as TIFF if your scanner supports it, otherwise full-quality JPEG. Turn off all "auto-enhancement" features in the scanner software — you want the rawest possible file with no algorithmic interpretation, because every cleanup is going to happen downstream where you have proper tools. The Epson V600 and V850 remain the working photographer's favorites in 2026, and a used V600 runs about $200.

The smartphone option is fine if you do not own a flatbed. Use a phone with at least a 12-megapixel main camera, lay the print on a flat, evenly lit surface (a window-lit desk is ideal — avoid direct sun, which creates glare on the print's surface), and use Photoshop Express scan mode or Apple Notes document scan to capture it. Both apps automatically correct perspective and crop the print from the background. The result is good enough for most restoration work, though not quite as crisp as a flatbed scan.

Once you have the scan, save two copies in two different places. One on your computer, one in cloud backup or on an external drive. The five minutes that takes you now is the five minutes a future version of yourself will be enormously grateful for if the original print is ever lost.

Step 2: Damage removal — fix tears, scratches, water damage

This is the step where the photo starts to look like itself again. The goal is faithful repair — same composition, same subjects, same scene, just without the cracks, stains, fingerprints, and tears.

Nano Banana Pro (Google's Gemini 3 Pro Image, released October 2025) is the best general-purpose damage-removal tool in 2026. Its restoration mode preserves composition and identity rigidly while cleaning surface damage. The trick is naming the damage explicitly in your prompt: "restore this old family photo, remove the diagonal tear across the upper right, remove the fingerprint smudges on the left edge, repair the missing corner with appropriate background detail, preserve the faces and clothing exactly, keep the original composition." Vague prompts ("make this photo nice") produce vague results. Specific prompts produce surgical repair.

For prints with widespread surface cracks (the fine branching network from humidity damage), Nano Banana Pro handles them in one pass. For prints with a few large tears, you can ask the model to repair the tears specifically. For prints that are missing entire regions — a corner torn off, a face partially erased by water damage — the model will reconstruct plausible content, and at this point you cross the line between faithful and creative restoration.

Magnific (now part of Freepik) is the tool for that creative reconstruction stage. It uses a "creativity slider" that lets the AI invent plausible new detail — skin pores, fabric texture, individual leaves on trees. For photos that are too damaged to faithfully recover, this is the only path to a complete image. The honest tradeoff: Magnific is opinionated. It will invent texture that was not in the original. Use it for backgrounds and clothing freely; be cautious about using it on faces, where invented detail can drift from the actual person.

If you have a flatbed scan, run it through Nano Banana Pro first. If the result still has missing regions you want filled, run a second pass through Magnific with a low creativity slider (around 3 out of 10) applied only to the damaged areas via masking. This two-stage approach gives you the faithfulness of Nano Banana Pro with the reconstruction power of Magnific only where it is needed.

Step 3: Colorization — should you colorize a black-and-white photo?

This is the most personal decision in the workflow, and there is no universally right answer.

The case for colorization: a colorized photo is more emotionally accessible, especially for younger family members who have only ever known the world in color. Seeing a great-grandparent in color, even imperfectly colored, can feel like meeting them. It is moving in a way that black-and-white sometimes is not.

The case against colorization: it is always a guess. AI colorization models do not know what color your grandmother's wedding dress actually was. They infer plausible color from training data on similar photos from similar eras. The result is usually beautiful and emotionally true even when it is not literally accurate. Some families prefer the dignity of the original black-and-white. The history is what the history is, and the print has earned its monochrome.

If you decide to colorize, Nano Banana Pro is the 2026 tool of choice. Its colorization mode preserves identity, composition, and lighting while inferring period-appropriate color. The prompt to use: "colorize this black-and-white family portrait from approximately [year], use historically plausible colors for clothing and skin tones, preserve the original composition and facial features exactly." Naming the decade helps the model anchor its color palette — the colors fashionable in 1925 were not the colors fashionable in 1975. If you know specific facts ("the dress was navy blue, her eyes were green"), include them in the prompt and the model will honor them.

Always save the colorized version alongside the black-and-white master, never instead of it. Label the file clearly: grandma_wedding_1948_colorized_2026.tif. Future family historians will thank you for distinguishing the inferred color from the original capture.

For a deeper dive into colorization specifically — model comparisons, prompt patterns, and historical accuracy — read our companion guide on how to colorize black-and-white photos with AI.

Step 4: Face enhancement — keep your grandparent looking like themselves

Face recovery is where most restoration jobs succeed or fail. Get this wrong and the restored photo is technically improved but feels like it shows a stranger. Get it right and the person looks like themselves, only sharper.

Two principles guide this step: preserve identity over invent detail, and pick the tool that matches the damage.

Topaz Photo AI is still the leader for damaged faces with significant missing detail. Its face-recovery model is purpose-built for old portraits — it sharpens what is there, recovers nearby detail, and is conservative about inventing new features. The huge advantage of Topaz in 2026 is that it runs locally on your GPU. Nothing leaves your computer. For a family photo of someone who has died, the privacy of keeping the file off the cloud is meaningful. The tradeoff is that Topaz costs about $200 as a one-time perpetual license, though it pays for itself quickly if you do more than a few restorations a year.

Nano Banana Pro is the better choice for lightly faded portraits where the face is mostly intact but soft. Its facial reconstruction holds identity rigidly — same face shape, same eye color, same expression — while sharpening detail and recovering local contrast. The prompt to use here: "enhance the face in this portrait, sharpen the facial features, recover skin texture, preserve the exact facial identity, do not change the expression, do not add or remove any features." That last clause matters. Without it, the model occasionally adds glasses, removes a mole, or "fixes" a slightly asymmetric smile that was part of the actual person.

Avoid Magnific and other high-creativity tools for the face. They invent new pores, new hair patterns, new micro-expressions. For backgrounds and clothing, that invention is harmless. For the face of someone you loved, it is a quiet form of erasure.

If you are unsure which face-recovery tool to choose, our guide to the best AI photo restoration tools in 2026 compares them side-by-side on the same damaged portraits.

Step 5: Upscale for printing — make it gallery-ready

The final step is about size. A 600 DPI scan of a 4x6 print is about 2,400 by 3,600 pixels — enough for an 8x12 print at 300 PPI, which is fine for an album page but small for a frame on the mantelpiece. To make the restored photo print at gallery size, you need to upscale it.

Crystal Upscaler (Replicate) is the workflow's final tool. It scales 2x to 10x, produces up to 200-megapixel output, and costs about $0.025 per upscale. For most family restoration jobs, 4x is the sweet spot — it takes a 600 DPI scan of a 4x6 print up to roughly 10,000 by 14,000 pixels, which is enough for a 24x36 inch gallery print at 300 PPI.

The order matters. Run the upscale last, after all other restoration is complete. Upscaling an already-restored photo gives the upscaler clean detail to work with. Upscaling a damaged photo first just makes the damage bigger.

For prints destined for a frame, save the upscaled file as TIFF (lossless) and order from a quality lab. Whitewall, Bay Photo, Mpix, and most local pro labs print at 300 PPI on archival paper. Order the size that matches the wall, not the size that matches the file — a restored grandparent portrait at 16x20 inches in a quality frame on a fine-art paper will outlast you on the wall.

Should you keep or replace the original?

Always keep the original. This is the only step in the workflow that is not optional.

AI restoration produces a beautiful new file. It does not produce a new heirloom. The cracked, faded, fingerprinted physical print on your desk has been in the world for sixty or eighty or a hundred years. It was held by people who are gone. It carries the actual fingerprints of an actual ancestor. No restored digital file replaces that. The role of restoration is to give the family a sharing-and-display version — a print for the mantel, a file for the photo book, a digital share for cousins who live across the country. The original goes back into archival storage. Both versions are valuable; they serve different purposes.

If the original is severely damaged and continuing to decay (mold, deep crumbling), the right move is to scan it as soon as possible, then consult a physical-print conservator about stabilization. Do not "fix" the physical print with tape, glue, or rubber cement. Anything you do to the original is usually irreversible and reduces its value.

How much does AI photo restoration cost per photo?

The honest 2026 math, per restored photo:

The pro workflow needs multiple models — Nano Banana Pro for damage and colorization, Topaz for face recovery, Crystal for the upscale. Rangy gives you Nano Banana Pro, Crystal, and Skin Enhancer through your own API keys (Replicate + Kie + Freepik). A complete family photo restoration runs about $0.50–$2 in API spend. For working photographers offering this as a service, that's a $25–$100 billable job — restoration is one of the highest-margin services a portrait photographer can add to their menu in 2026, and the API spend is well under one percent of the billable rate.

The verdict — the complete 10-minute workflow

For the broader version of this workflow — covering all types of old photos, not just family — read our companion piece on how to restore old photos with AI. For photographers building this into a paid service, our guide for wedding photographers covers pricing, client conversations, and the legal considerations of selling restoration work. And for the print-quality side of the workflow specifically, see upscaling AI images for print.

Frequently asked questions

Last updated May 2026. Pricing and rankings change frequently — verify on each provider's official page before committing.