Old family photos are full of scratches, stains and faded colors. I wanted to see how far a single consumer GPU can go at fixing them, without any cloud service or subscription. Here is what worked, what failed, and what I would tell anyone trying the same thing.

The setup

  • An RTX 3060 Ti (8 GB), ComfyUI, and an image-editing model (Flux Kontext, fp8).
  • About 3 to 4 minutes per photo.
  • Everything runs locally, so the photos never leave the machine.

What did NOT work

  • Classic face-restoration models: they happily invent a new face. Great-looking, but it is not the person anymore.
  • Inpainting + scratch detectors: they missed hairline cracks and erased handwriting on the back of prints.
  • One prompt for every photo: a 19th-century tintype and a 1970s color slide need very different instructions.

What works better

  1. Measure the noise first. A clean print should barely be touched: if I pre-filter it, the wrinkles and skin texture disappear. A newspaper halftone, on the other hand, needs a strong clean-up before the model sees it.
  2. Pick the instruction by photo type. Black & white gets repair + colorization; a color print gets repair + white-balance correction (the red or yellow cast of 60s-90s prints).
  3. Colorize first, derive the restored B&W from it. The colorized pass keeps faces closer to the original than a direct B&W restoration.
  4. Add a light film grain at the end, so the result still looks like a photograph and not a plastic render.

Examples

Original, restored, colorized (public-domain source, 1936):

Before and after: 1936 portrait restored and colorized

A 19th-century group portrait, with a deliberately muted palette because a bright blue suit looked wrong for the period:

Before and after: 1892 group portrait restored and colorized with a muted palette

The honest limits

When the original is almost unreadable (this newspaper halftone is the extreme case), the model has to reconstruct a lot: the face you get is plausible, but partly invented. I tell clients that before they order, instead of pretending it is a faithful restoration.

Extreme case: a newspaper halftone reconstructed, restored and colorized

If you would rather not run this yourself

I do it as a service on Fiverr: scratches, stains and tears repaired, and colorization included for every black & white photo. One photo, or packs of 5 and 10 for a much lower price per photo.

https://www.fiverr.com/crabe666/generate-hundreds-of-ai-images-in-one-consistent-style-in-bulk

P.S. I also build web scrapers for product data, prices and reviews from online stores: https://www.fiverr.com/crabe666/scrape-product-data-prices-and-reviews-from-any-online-store