The workflow
- Scan the original print or negative at 600 ppi or more, as a 16-bit TIFF or PNG, rather than photographing it.
- Keep the untouched scan as your archive master and work on a copy.
- Remove dust, scratches and tears first, with dust-and-scratch or inpainting tools.
- Correct the colour cast and fading.
- Upscale faithfully to the size you need, usually 1.5–2× from a good scan.
- Use face restoration sparingly, blending it in at partial opacity.
- Colourise last, if at all, and label colourised copies as such.
Start with the best possible scan
A flatbed scanner at 600 ppi records detail no upscaler can invent. A 6 × 4 inch print at 600 ppi is already 3600 × 2400 pixels, enough for a sharp 12 × 8 inch print at 300 ppi without any AI. Phone “scanner” apps are convenient, but they add perspective distortion, glare and their own sharpening. If you use one, shoot in even daylight with no flash.
| Original | Scan at | Resulting pixels |
|---|---|---|
| 6 × 4 in print | 600 ppi | 3600 × 2400 |
| 3.5 × 2.5 in wallet print | 1200 ppi | 4200 × 3000 |
| 35 mm negative (36 × 24 mm) | 3200 ppi | ≈ 4535 × 3024 |
Pixels = inches × ppi. A 35 mm frame is about 1.42 × 0.94 inches.
Repair before you upscale
Upscalers treat dust specks and scratches as detail and enlarge them faithfully, or worse, turn them into texture. Clean them up at the scan’s native resolution with a dust-and-scratch filter, a healing brush or an AI object remover, then fix the colour. Faded prints usually drift towards yellow or magenta; an auto-levels pass per channel gets you most of the way.
Upscale faithfully
Use a faithful model: Real-ESRGAN-family models, Topaz’s standard or recovery modes, or similar. Creative upscalers will happily add pores, eyelashes and fabric that weren’t there. On a family photo that’s rewriting history. For prints, 1.5–2× from a good scan is usually plenty. If the only copy is a small low-quality JPEG, upscale 2× and judge the result before going further.
Faces: restore gently
Small faces in group photos are where AI helps most, and where it misleads most. Face restoration models (GFPGAN, CodeFormer and their commercial cousins) rebuild eyes, teeth and skin from a handful of pixels. Blend the restored face at partial opacity over the faithful upscale, and check it against other photos of the person. More in the portrait guide.
Colourise last, and label it
Colourisation is interpretation, not restoration: the model guesses colours from context. Do it on the finished, repaired image, keep the black-and-white version as well, and name the colour copy clearly (“colourised 2026”) so future relatives know which is the original. If you want a quick one-off result, Upsampler has a free old-photo restoration tool that needs no account. For an archive, though, do the steps above yourself.
FAQ
What ppi should I scan old photos at?
For prints, 600 ppi is a good default and 1200 ppi is better for small wallet-size prints you want to enlarge. Film negatives and slides need far more: 2400–4000 ppi is common, because the frame is tiny. Scanning higher beats upscaling later, because the scanner records real detail.
Should I upscale or scan at a higher resolution?
Scan at a higher resolution whenever you still have the original. The upscaler can only guess; the scanner records what is actually there. Upscale when the scan, or a digital copy, is all you have.
Why do restored faces look like different people?
Face restoration models such as GFPGAN and CodeFormer generate a plausible face from very little information. At full strength they pull features towards an average face. Blend the restored face at 30–60% over the faithful upscale, and compare it with other photos of the same person.