A box of family photos rarely ages evenly. Some prints hold up for eighty years, others fade to a flat brown wash after twenty, and the ones that were folded, taped, or stored in a damp basement come out cracked. AI restoration models are now good enough that most of that damage is recoverable in minutes, using the same generative stack covered across the AI photo enhancement tools we compared this year.

The catch is that restoring with AI is not one button. It is four or five separate operations that have to run in the right order: capture, damage repair, detail reconstruction, optional colorization, and upscaling. Run them out of order and the model invents detail on top of noise, which is how you end up with a relative who no longer looks like themselves.

What actually breaks in an old photo

Before you pick a tool, name the damage. Restoration models are specialized, and the one that removes scratches is usually not the one that recovers a blown-out face. The same logic applies across the wider editing stack in our rundown of AI image editors and photo tools, where each model does one job well.

Old prints typically show some combination of:

• Physical damage: tears, creases, missing corners, tape residue, punch holes
• Surface damage: scratches, dust, mold spotting, silvering on the emulsion
• Chemical fading: the magenta or sepia shift you get from dye layers decaying at different rates
• Detail loss: soft focus, low original resolution, or a contact print that never held much detail
• Capture damage: glare, skew, and moiré introduced when you photographed or scanned it

Sorting the photo into two or three of these buckets tells you which passes to run and, more usefully, which to skip. A sharp 1970s print with a color shift needs a color pass and nothing else, and diffusion models like the ones behind FLUX will happily change a face that did not need changing if you let them.

Step 1: digitize the original properly

Everything downstream is capped by this step. A flatbed scan at 600 DPI beats a phone photo in almost every case, because you get even lighting, no perspective skew, and enough pixels that the upscaler is refining real detail rather than guessing. Scan to TIFF or PNG where possible, since repeated JPEG saves bake compression artifacts into exactly the fine texture a restoration model needs to read.

If no scanner is available, a phone works. Lay the print flat, use indirect window light rather than the flash, shoot straight down, and fill the frame. Crop the paper border off before uploading, because it confuses models that expect the whole frame to be image content, a quirk we also hit when changing photo backgrounds with AI.

Step 2: run the damage repair pass first

The first model pass should only remove damage. Scratch, tear, and dust removal is an inpainting task: the model reads the pixels around a defect and reconstructs plausible content to fill it. That works well when the surrounding area is intact and badly when the damage covers something structural, like half a jaw.

For large missing regions, mask the area manually and give the model a short text hint about what belongs there. Generative fill in this mode behaves like ordinary text-to-image inpainting, so the habits in our FLUX prompt library transfer directly: be concrete about material and lighting, and describe what should be there rather than what should not.

Work at full resolution and repair in sections if the file is large. Most hosted restoration tools downscale the input to fit a fixed model resolution and then hand back an upscaled result, which quietly costs you real detail on the first pass.

Step 3: rebuild faces separately

Faces are the part everyone judges, and they get their own pass for a reason. Dedicated face-restoration models are trained almost entirely on portraits, so they reconstruct eyes, teeth, and hairlines far more convincingly than a general enhancer. They also fail in a specific way: pushed too hard, they replace a real face with a generic attractive one, the same drift covered in our guide to realistic AI faces.

Keep the strength low, run it twice at 40 percent rather than once at 90, and compare against the scan at 100 percent zoom. If the nose shape or the distance between the eyes moved, you have gone too far. Where a modern model family like FLUX 1.1 Pro helps is in holding structural fidelity while still cleaning up texture, which is the tradeoff restoration lives or dies on.

Step 4: colorize only when you can defend the colors

Colorization is the most impressive step and the least factual one. The model is guessing. It reads context, skin, foliage, sky, fabric, and applies statistically likely colors, which means a wartime uniform can come back in the wrong service branch and a wedding dress can come back cream when it was blue.

If the photo is headed for a family archive, colorize with a light touch and keep the monochrome original alongside it. Feed in reference information where you have it: most colorization tools accept hint prompts, and a five-word note like “green wool coat, red brick” beats the default guess every time. Muted period-accurate palettes read as more believable than saturated ones, a preference that also shows up in photo generators tuned for realism.

Step 5: upscale last

Upscaling belongs at the end, after damage is gone and detail is rebuilt. Do it first and you spend the rest of the process fixing enlarged scratches. A 2x upscale is usually enough for print, while 4x is for wall-size output and starts to introduce plastic texture on skin. The API-level mechanics are the same ones in our walkthrough of upscaling images through ClarityAI.

Judge the result at the size it will actually be viewed. A 4x upscale that looks synthetic at 100 percent zoom often looks perfect as a 6×8 print, and the reverse is true for anything shown on screen at full resolution.

Turning the steps into a repeatable workflow

Doing this once by hand is fine. Doing it for two hundred photos out of a shoebox is not, and that is where chaining the steps pays off. Each pass is a discrete model call with a defined input and output, so the whole sequence can run as a pipeline: the scan goes in, the repair, face, color, and upscale models run in order, and the finished file lands in one place. Wireflow AI handles that kind of multi-model chaining without code, which matters when the same five steps have to run identically across a whole archive.

The other advantage of a fixed pipeline is that you can change one stage without redoing the rest. Swap in a different face model, rerun the batch, compare the two. Building that kind of chain has become routine, and the general pattern is covered in our guide to no-code AI workflows.

Common mistakes

• Running everything in one pass. Combined restore-and-enhance buttons apply every model at full strength and give you no way to tell which one caused a problem.
• Restoring from a compressed copy. Start from the highest-quality capture you have, not a screenshot someone texted you.
• Over-sharpening. Old photos have grain, and grain is not noise. Removing it entirely makes the image look like a modern render of a period scene.
• Trusting colorization as history. Always keep the original file.
• Skipping the comparison step. Flip between before and after at full zoom on every face before calling it done.

FAQ

Can AI restore a photo that is torn in half? Yes, if both halves exist. Scan them separately, align them in an editor, then run inpainting across the seam. If a half is missing entirely the model will invent that content, and it will not be your relative. Reconstruction of that kind is generation rather than restoration, closer to what a FLUX image generator does from a prompt.

How much does AI photo restoration cost? Most consumer tools offer a few free restorations then charge per image or by monthly subscription, typically a few cents to a dollar per photo at API level. Running the models yourself through an API is cheaper at volume and more predictable in batch.

Will restoration work on a photo with no visible face detail? Partly. If the face is a blur of a dozen pixels, no model can recover information that was never captured, and the output is an invention. Restoration works best when detail exists but is obscured by damage or softness, a distinction that also applies when creating portrait-quality headshots.

Should I restore the print or rescan the negative? If you still have the negative, scan the negative. It holds far more detail than any print made from it, and a good negative scan often needs almost no restoration at all. The quality gap is the same one that separates model tiers in our image generator comparison.

What file format should I save the result in? Keep a lossless master in TIFF or PNG, then export JPEGs for sharing. Save the untouched scan permanently and separately, so a future model can be run against the original rather than against your edit.

Do these models work on documents and film stills too? They do, with the caveat that text-heavy documents need a model that respects character shapes, since general restoration models will happily hallucinate plausible but wrong letters. Faster iteration models such as FLUX Realtime are useful for testing which pass helps before committing to a full-resolution run.

Wrapping up

Good restoration is mostly discipline. Capture the best scan you can, repair damage before touching detail, keep face models on a short leash, treat color as interpretation rather than fact, and upscale at the very end. Follow that order and the results hold up under zoom, which is the only test that matters.

For a handful of photos, a single hosted tool is enough. For an archive, the win comes from making the sequence repeatable, and a visual AI workflow builder lets you wire the same five stages together once and run them across every file in the box.