Sources and review notes
Last reviewed August 22, 2026
Reviewed against: Nostalgia's current restoration-choice and safety contract and a labeled August 2026 Museum pass whose rejected and unchanged results remain public.
Scope: No restoration model can recover evidence that is absent from the source. Colour, missing areas, softened faces, and unreadable text require explicit human judgment and may need to be left unresolved.
Try a restore free, no sign-up, and see the limits on your own photo.
Restore one photo freeNo sign-up. One photo, on the web.
What Works and How Well
Not all restoration tasks are equal. Here is how AI performs across the most common types of work.
| Area | Quality | Notes |
|---|---|---|
| Scratch & dust removal | Often useful | Surface marks may be reduced, but inspect faces, text, borders, and repeated patterns for invented or erased detail. |
| Fading & contrast recovery | Often useful | Tone and contrast can improve visibly. Compare the full frame so clothing, skin tone, highlights, and documentary marks are not silently changed. |
| Upscaling (2×) | Presentation aid | Makes a file larger and may infer plausible detail; it cannot recover information the source never captured. Keep the source file as the record. |
| Face enhancement (mild blur) | High-review | May make a face clearer or may alter identity. Compare every face against the Source and reject a version that changes the person. |
| Colorization | Creative interpretation | Produces plausible—not historically recovered—color. Keep it separate from a faithful Chosen restoration and label it as colorized. |
| Tear & crease repair | High-review | Fills gaps with plausible content. Tears through faces, text, hands, or identifying details may require manual work or withholding the result. |
| Severe damage (water, mold, missing areas) | Often unsafe | AI can only invent plausible content where evidence is missing. Keep the Source, expect a leave-as-is or withheld decision, and consult a conservator for valuable or fragile material. |
| Deblurring phone captures | Recapture first | A sharper capture is more trustworthy than generated sharpness. Use repair only when the original print cannot be recaptured or rescanned. |
| Glare reduction | Recapture first | Change the light or use multi-shot capture before asking AI to fill washed-out areas. Fully obscured information cannot be recovered faithfully. |
| Heavily blurred faces | Use with caution | May generate a plausible but incorrect face. Always compare with the original. |
Surface marks may be reduced, but inspect faces, text, borders, and repeated patterns for invented or erased detail.
Tone and contrast can improve visibly. Compare the full frame so clothing, skin tone, highlights, and documentary marks are not silently changed.
Makes a file larger and may infer plausible detail; it cannot recover information the source never captured. Keep the source file as the record.
May make a face clearer or may alter identity. Compare every face against the Source and reject a version that changes the person.
Produces plausible—not historically recovered—color. Keep it separate from a faithful Chosen restoration and label it as colorized.
Fills gaps with plausible content. Tears through faces, text, hands, or identifying details may require manual work or withholding the result.
AI can only invent plausible content where evidence is missing. Keep the Source, expect a leave-as-is or withheld decision, and consult a conservator for valuable or fragile material.
A sharper capture is more trustworthy than generated sharpness. Use repair only when the original print cannot be recaptured or rescanned.
Change the light or use multi-shot capture before asking AI to fill washed-out areas. Fully obscured information cannot be recovered faithfully.
May generate a plausible but incorrect face. Always compare with the original.
The Face Hallucination Problem
This is the most important limitation to understand. When a face in a photo is severely blurred, damaged, or very small, AI fills in plausible features based on patterns it has learned, not based on the actual person.
The result can look like a convincing photograph of someone who resembles the original subject, but is not them. Restored communities describe this as getting back “someone who might be the cousin of the person in the photo.”
Common Misconceptions
AI can restore every photo without review
AI is a best-effort approximation. Severely damaged or very low-quality photos may produce results with artifacts or invented detail. Quality of the input scan is the single biggest factor.
Colorization is historically accurate
AI infers colors from patterns. It does not know what color your grandmother's dress actually was. Treat colorization as a plausible suggestion, not a historical fact. Occasionally the AI produces muted or unnatural tones; if the result looks off, try running colorization again for a different interpretation.
One click fixes everything
The best results come from using the right tools in the right order, starting with a good scan. The photo check before restoration and review after restoration produce more trustworthy results than a single blind pass.
AI restoration replaces professional conservators
For archival-grade preservation of extremely valuable or fragile originals, a professional conservator is still appropriate. AI restoration is excellent for family photo collections: the 99% of photos that need care but not museum-grade treatment.
Enhancement vs. Restoration vs. Generation
These terms are often used interchangeably, but they mean different things:
Most AI tools, including Nostalgia, do all three to varying degrees. Nostalgia's photo check tells you what kind of work each tool will do on your specific photo, so you know what to expect before you start.
How to Get the Most Trustworthy Results
- 1Start with the clearest scan you can get
AI cannot recover detail that was never captured. Use a clean, full-frame Source and rescan glare, blur, or perspective errors whenever the print is available.
- 2Read the photo check report
It tells you the photo's condition, what damage was detected, which tools are safe, optional, or blocked, and whether conservation guidance applies. This saves unnecessary steps and sets expectations.
- 3Follow the recommended tool order
Restore first, then optional extras. Each step builds on the previous one for better results.
- 4Review faces carefully
Zoom in and compare before/after. Faces are where AI is most impressive and most fallible.
- 5Keep your Source scan
Your accepted Source is the reference record. Nostalgia preserves it as a separate version, but keep local backups too.
A worked example, including what went wrong
We built a public exhibit from fifteen Farm Security Administration photographs and ran a documented restoration pass in August 2026, then judged the results by eye against conditions written down before the run. That historical pass is not the current Automatic Restore route. Seven Hungry Children records the outcome plainly: two pairs earned a before/after slider, three were rejected because the model invented detail on a face — including freckles on Allie Mae Burroughs that do not exist in the negative — and one came back byte-for-byte identical to the original.
That is one bounded example of AI restoration on real archival material, and it is why the wall shows the photograph as it survives, with the restoration as something you pull across it rather than the thing you are handed. For family prints rather than archival negatives, the gallery shows restored examples you can compare at a glance.
Try it yourself
Upload one old photo free, no sign-up. Compare the result with the Source and keep only a version you trust.