Apr 24, 2026 · 6 min read
AI Photo Restoration in 2026: What's Possible Today
AI photo restoration is suddenly excellent at color and clarity. It's still mediocre at faces. Here's the working line.
Restoration used to mean Photoshop, a graphics tablet, and three hours per print. Today, GPT-image-2 takes a damaged scan and returns a usable result in 20 seconds. But not every scan, and not every kind of damage.
What works in 2026
- Color restoration on faded prints — extremely good.
- Scratch / stain / fold removal — near-perfect at low magnification.
- Sharpening soft images — yields print-usable results from web JPEGs.
- Conservative re-coloring of black-and-white photos — coffee-table-book quality.
What still struggles
- Damaged faces of identifiable people — 'face hallucination' is real and hard to control.
- Heavy water damage where 80%+ of the image is reconstructed — feels invented.
- Multi-generation copies — too little signal to recover.
A working pipeline
- Scan at 600+ DPI. Resolution beats every algorithm.
- Run a single restore pass at HD quality. The prompt matters less than the original quality.
- Inspect faces. If they look 'invented', restore again with a tighter mask around the rest.
- Upscale 2× as a final pass for print.