I Restored My Grandfather's 1927 Portrait with AI and My Family Saw His Face Clearly for the First Time in 40 Years

My grandfather was born in 1923 in a small village in Hunan. There is exactly one photograph of him as a child. It is a black and white print, maybe two by three inches, taken around 1927 when he was four years old. For as long as I can remember, that photo has been taped inside a yellowed album in my grandmother's apartment, slowly fading under decades of fluorescent light.

By the time I got my hands on it last spring, you could barely make out his face. His eyes were two dark shadows. His mouth was a faint line. The left third of the photo was covered by a brown water stain from a leak that happened sometime in the 1980s. My grandmother told me she had not been able to see his face clearly in that photo for over forty years.

I spent three weeks learning how to restore it. I made plenty of mistakes along the way. I over-sharpened him into a wax figure on my first attempt. On my second attempt, I let the AI fill in too much missing detail and he came out looking like a completely different child. But by the fourth attempt, I had a workflow that actually worked. When I showed the final version to my grandmother, she touched the screen with her fingertips and did not say anything for a long time.

Start with a good scan, not a phone photo

My first mistake was trying to restore a photo I had snapped with my phone camera. I held my iPhone over the album, tried to keep it steady, and took a picture of the picture. The AI restoration tools I tried could not do much with it because phone photos of photos introduce reflections, uneven lighting, subtle angles, and their own compression artifacts on top of the original damage.

What actually worked was using a flatbed scanner. I borrowed a Canon CanoScan from my local library. It cost me nothing. I scanned the original print at 600 DPI, saved it as a TIFF, and that became my starting point. The difference was immediate. The scan captured detail that my phone camera had completely missed. Tiny texture in the paper. Subtle variations in the fading that the AI could later use to reconstruct what was underneath.

If you cannot access a scanner, a phone photo can still work as a last resort. But you need to put the photo flat on a table, use natural daylight from a window, hold the phone perfectly parallel to the photo, and use an app that lets you save as TIFF or at least high quality PNG. Do not use the default camera app JPEG. The compression will cost you detail you cannot get back.

Fix the big damage before you feed anything to AI

AI restoration models are designed to enhance what is there. They are not designed to reconstruct information that has been completely destroyed. If a photo has a tear across it, a water stain covering a third of the image, or heavy creases, the AI will try to restore those damaged areas the same way it restores everything else. It will sharpen the edges of the tear. It will enhance the texture of the water stain. It will make the damage look sharper and more permanent.

I learned to do basic cleanup first in a regular photo editor. For my grandfather's photo, I used the clone stamp tool to fill in the worst of the water stain with texture from the undamaged parts of the background. I did not try to make it perfect. The goal was just to remove the harsh edges of the damage so the AI would not treat it as important image detail. Ten minutes of clumsy clone stamping saved me hours of fighting the AI later.

For tears and creases, the same principle applies. Smooth out the physical damage as much as you can with basic editing tools. Give the AI a clean but blurry starting point rather than a damaged but detailed one. The AI handles blur much better than it handles damage.

Two passes, not one

The biggest breakthrough in my workflow came when I stopped trying to do everything in one step. Early on, I would run a photo through a general restoration tool and hope it fixed everything at once. It never did. The background would look great but the face would be waxy. Or the face would look sharp but the clothing textures would be obliterated into smooth gradients.

What actually worked was splitting the job into two separate passes. First pass, ClarifyPix old photo restoration. This handles the overall cleanup. It removes dust, scratches, fading, and general degradation across the entire image. It brings back the overall contrast and sharpness that the photo had when it was new. This pass costs 10 credits and takes about ten seconds for a standard scanned image.

Second pass, face restoration. Once the overall photo looks decent, I run a dedicated face restoration model on just the faces. This model was trained specifically on millions of human faces, so it understands facial anatomy. It knows where eyes sit relative to each other. It restores pupils, eyelashes, lip lines, and skin texture without touching the background or clothing. This pass costs 4 credits per image and takes about five seconds.

The two passes together do what one pass never could. The general restoration fixes the medium the photo is printed on. The face restoration fixes the people in the photo. They are fundamentally different problems and need fundamentally different tools.

Colorization is a separate decision

After I had restored the black and white image to something that looked like a well-preserved 1927 photograph, I had to decide whether to colorize it. This is trickier than it sounds. Black and white photos have a certain gravity to them. They feel historical. They feel real in a way that colorized versions sometimes do not.

I tried colorizing my grandfather's portrait. The AI did a technically good job. The skin tones looked natural. The background foliage came out in appropriate greens. His clothing was a plausible shade of dark blue. But something was off. The colorized version looked like a still from a period drama, not a real photograph. It had lost the texture of being old. My grandmother preferred the black and white version.

For my own family archive, I keep both. The restored black and white version is the one I printed and framed. The colorized version lives in a digital folder labeled "experiments." Some photos take colorization beautifully. Especially outdoor scenes and group photos where the color information adds useful context about clothing, setting, and season. But portraits, I have found, often work better in their original black and white. Your mileage will vary by photo.

What to do with the original scan

I made a mistake I want to warn you about. After my fourth attempt produced a version I was happy with, I almost deleted the earlier failed attempts and the original scan. My thinking was that I had the good version now, so why keep the bad ones. That thinking is wrong.

AI restoration is interpretation. Every restored photo is the model's best guess at what the original looked like. It is a very educated guess, trained on millions of real photos. But it is still a guess. For genealogy purposes, the original scan is the historical record. The restored version is a visual aid. Future AI models that do not exist yet might do a better job with the same original scan. If you delete the original, you lose that possibility permanently.

I now keep three versions of every photo I restore. The original high-resolution scan as a TIFF, stored in a folder called "originals." The restored version as a PNG in "restored." And a small JPEG copy in "shared" that I send to family members. The original scan is backed up to two different cloud services. My grandmother's ability to see her husband's childhood face again started with a scan from a public library scanner. That TIFF file is now the most valuable digital file I own.

Printing restored photos for family

Once I had a restored version I was happy with, I wanted to print it so my grandmother could have a physical copy. This introduced a whole new set of problems that I had not considered.

The original photo was about two by three inches. I wanted to print it at five by seven so she could see the details without a magnifying glass. That meant I needed to upscale the restored image to 2100x1500 pixels at 300 DPI. The restoration had improved the quality significantly but at the original size, the pixel count was still only about 1200x800 from the 600 DPI scan. I needed to upscale it further to get a clean 5x7 print.

A second round of AI upscaling on the already-restored image worked surprisingly well. The 4x upscale took the restored version to 4800x3200 pixels, which is more than enough for 5x7 at 300 DPI. I had it printed at a local photo lab on matte paper. The matte finish was important because glossy paper tends to highlight any remaining imperfections, while matte paper is more forgiving and looks more period-appropriate for a vintage photo. Total cost including the print was about eighteen dollars and the look on my grandmother's face was worth significantly more than that.

How to handle a whole album

After the success with my grandfather's portrait, I got ambitious. My grandmother has four albums full of old family photos, probably two hundred images in total, spanning from the 1920s through the 1970s. Restoring them all one at a time with the full two-pass workflow would take me months.

I triaged the collection. Photos with faces got the full two-pass treatment. General restoration plus face restoration. Cost was about 14 credits each. Photos without faces, like landscapes and buildings, got a single pass of general restoration at 10 credits. Photos that were already in decent shape I left alone entirely. Not every photo needs AI restoration. If the original scan looks good at normal viewing size, restoration is a luxury, not a necessity.

I also learned to batch scan efficiently. Most flatbed scanners can fit three to four small photos at once. I scan them all as one large TIFF, then split them into individual files later. This cut my scanning time from several hours to about forty minutes for a hundred photos. The scanning is the tedious part. The AI restoration is the fun part. Do not let the scanning bottleneck keep you from getting to the fun part.

If you have old family photos sitting in albums that you have not looked at in years, pull them out and scan the ones that matter most. Start with the photos where faces are hardest to see. Those are the ones where AI restoration makes the biggest difference. You might be surprised by what is still hiding in those faded prints, waiting for the right tool to bring it back.