Restore or colorize an old photo: which first?

You found a box of family prints. Some are faded, some are blurred, most are black and white. You want to fix the faces and add colour. The short answer on order: restore first, colorize second, enlarge last. The longer answer is below, because the two jobs are done by different models that guess in different ways.

Two different jobs

People say "restore" to mean anything that makes an old photo look better. In practice, two separate models do two separate things.

Face restoration rebuilds blurry or damaged faces. The free old photo restorer uses GFPGAN v1.4 from TencentARC. A small detector (YuNet) finds up to 20 faces and five points on each: the eyes, nose tip and mouth corners. Each face is lined up to a standard 512×512 layout, rebuilt by GFPGAN, and pasted back through a soft oval edge so no seam shows. It works on faces only. Scratches, dust, grain and blur in the rest of the photo stay as they were.

Colorization adds colour to a photo that has none. The free photo colorizer uses DDColor, from a 2022 paper on colorization with two decoders. In the tool, DDColor looks at a 512×512 grey copy of your photo and predicts only colour. That colour is scaled up and laid over the photo's own lightness at full resolution. It does not sharpen anything or fix anything. Every scratch and every blur stays.

So neither tool does the other's job. A restored face on a black-and-white photo stays black and white. A colorized photo with a smeared face is a smeared face in colour.

Why restore comes first

There are three reasons, and the tools' own code points to all three.

1. The colorizer reads shapes. DDColor guesses colour from what it can recognise: a face, a collar, a sky, a lawn. It sees the photo at 512×512. A face that is a soft blur at that size gives it less to go on. A rebuilt face with clear eyes, lips and skin gives it more.

2. Restoring a colour photo can clash. When the restorer sees a black-and-white or sepia photo, it keeps the photo's own tones on the rebuilt face. When it sees a colour photo, including one you have just colorized, it keeps the skin colour GFPGAN draws. GFPGAN draws that skin from what it learned about faces, not from your colorizer's result. The face inside the oval can come out a different tone from the neck and hands around it. Restore the grey photo, and the colour pass paints face and body in one go.

3. The restorer already works this way. Its optional "Also colourize" switch runs DDColor on the finished, restored photo, never before it. If you want both jobs in one place, that switch is the restore-then-colorize order built in. It downloads about 224 MB more the first time you turn it on.

Use the separate colorizer when there are no faces to fix, such as a house or a street scene, or when you want its colour strength slider.

Why enlarge comes last

It is tempting to upscale a small scan first so the restorer "has more to work with". That buys nothing. The restorer rescales every face to its own 512×512 layout anyway, then rebuilds it. Upscaling first only adds pixels that were invented by a different model.

Enlarge at the end, and only if you need a bigger file for a print or a screen. Be aware of what the free image upscaler is. It runs Anime4K on WebGPU, which is built for graphics and anime. On a restored portrait it can make faces look over-sharpened. Check at 100 percent before you print. If you need a photo upscaler rather than a graphics one, Topaz Upscale Image is a paid row in Versely's catalog at 7 credits per image.

The better fix for a small image is not an upscaler. It is a better scan.

Scan the print properly

Most bad restorations start with a bad copy. The Library of Congress scanning guide for personal collections is a good baseline:

Clean the scanner glass. Dust on the glass becomes specks on every scan.
Remove clips, staples and tape from the print, as far as you can without damaging it.
Resolution. It suggests 300 dpi for 4×6 and 5×7 photos in general, and 400 or 600 dpi if you want to enlarge them to 8×10 or bigger. Slides need much more, around 1900 dpi.
Grayscale is its suggested setting for black-and-white photographs.

For restoration, lean towards the higher figure. A face that is 1 cm wide on a 4×6 print is about 120 pixels at 300 dpi and about 240 at 600. The restorer skips faces smaller than about 24 pixels, and gets better material from bigger ones. It scales photos above 4,096 pixels on the long edge down first, so there is no point going far beyond that for a restoration copy.

No scanner? A phone photo can work if it is sharp. Lay the print flat, shoot straight down so it is not skewed, use even daylight, and tilt it slightly if you see glare on a glossy surface. Crop the border off before you start.

Keep the untouched scan as your master. Do every step on a copy.

Using the restorer well

The restorer has two controls that matter more than anything else.

Strength. Restoration strength runs from 0 to 100 percent and blends the rebuilt face with the original. It starts at 70. GFPGAN does not uncover hidden detail. It draws a face that fits the blurry one. On a very small or very soft face it can change the eyes, teeth or jaw, or make someone look younger and smoother. The tool's own guidance is that 50 to 80 percent keeps more of the real face. Moving the slider does not run the model again, so try a few settings.

Per-face on and off. Each face found can be switched off. Turn off any face that no longer looks like the person. It is better to leave a grandparent slightly soft than to swap them for a plausible stranger.

If it says no faces were found, the face may be too small, in profile, in deep shadow or badly damaged. Crop tighter around the people and try again.

Colour is a guess, and you should say so

A black-and-white photo stores no colour. No tool can get it back. DDColor predicts likely colours from millions of colour photos it learned from. Skin, sky, grass and trees usually look right. Dresses, cars, painted walls and signs are guesses. A blue dress can come out red.

The free colorizer has a strength slider from 0 to 150 percent. It does not have a way to recolour one object. Lower strength gives a softer, hand-tinted look, which often reads as more honest on a very old print. If a relative remembers the real colour of a car or a coat, fix it afterwards in a photo editor, or leave it muted.

Two more cautions:

Do not run the colorizer on a photo that already has colour. It keeps only the lightness and paints its own guess on top, so the original colours are replaced, not restored. The page warns you when it sees colour. Faded colour prints are the grey area. Try a low strength first.
Label colorized images when you share them. "Colorized" in the caption stops a cousin treating your guess as the record of what Grandma wore.

For heritage and history work, where accuracy is the brief, colorizing archive photos for brand heritage goes deeper on research and disclosure.

The practical order

Scan the print at 600 dpi, or take a sharp, straight phone photo. Keep the master.
Run the old photo restorer. Start at 70 percent strength and compare. Switch off any face that changed too much.
Turn on "Also colourize" in the restorer, or take the restored PNG to the photo colorizer for the strength slider.
Only if you need a bigger file, enlarge last, and check faces at full size.
Save each stage as its own file. Share with "restored" or "colorized" in the caption.

For grain in the background, which the restorer leaves alone, the image denoise and deblur tool can help. It is tuned for graphics, so check photos closely.

Practical limits

All three free tools run in your browser. Your photos are not uploaded. The models download once and are cached: about 340 MB for the face model, 0.2 MB for the face finder, and about 224 MB for the colorizer. They use WebGPU when your browser has it and a slower fallback when it does not. A laptop or desktop is more reliable than a phone for this. The face model needs a lot of memory, and older phones can reload the tab.

Versely's main product is AI video and image generation, on the web and in the mobile app. Some families turn a restored portrait into a short animated clip for an anniversary. That is a paid generation and a separate decision; the model catalog lists the image-to-video options. If you want a paid colorizer instead of the free one, Topaz Image Colorization is 7 credits per image.

For most family prints, the free order above is enough: restore, colorize, enlarge, and keep the original.