Synthetic aged-photo restoration

Start with the right photo.

A photo with readable facial features and limited surface damage; this test uses an internally generated aged print.

What to expect

Restore this deliberately simulated damaged photographic print.

Keep an eye on

Synthetic restoration demonstration, not an authentic archival recovery. Missing details and colors are inferred; exact historical appearance cannot be recovered from damaged pixels.

Make it yours, step by step.

1. Adjust the recipe to describe your subject and preferred background.
2. Copy the complete recipe, then open ChatGPT.
3. Upload your photo in ChatGPT and paste the recipe. The short code alone is not an official command.
4. Inspect identity, text and proportions. Ask for a targeted correction if needed.

Choose the change deliberately.

Restoration can plausibly fill scratches and damaged regions, but it cannot recover information that is absent. Keep the original available so viewers can distinguish visible detail from reconstructed detail.

Review and refine the result

Inspect eyes, lettering and boundaries crossing damaged areas. If the result becomes a newly invented portrait, reduce reconstruction and retain more source texture. The current demonstration uses simulated damage.

A related direction, and why

/VINTAGE — A vintage treatment adds an aged appearance; restoration instead attempts to reduce visible damage.

These are editing suggestions, not additional test observations. Read the practical guide or use the before/after checklist.

Where this effect helps — and its limits.

Restore this deliberately simulated damaged photographic print. Synthetic restoration demonstration, not an authentic archival recovery. Missing details and colors are inferred; exact historical appearance cannot be recovered from damaged pixels.

Input advice and use cases are editorial guidance, not additional test results.

Notes from recorded tests.

These notes describe the named runs, which may use a different subject from the comparison above. They do not guarantee results on another input or a future model. See our testing method.