We took one portrait, degraded it to 3,889 bytes, and handed it to an AI photo restoration model. It came back sharp — and it came back as a different person. The prompt said do not beautify, do not slim the face. The mole moved inward, the jaw narrowed to a point, and the uneven eyebrows evened out. That is the failure mode nobody screenshots, and it is the one that decides whether AI photo restoration gave you your grandmother back or a stranger.
Sharpness is the easy part. What decides whether you can use a photo restoration is how much of it the model had to invent — and an invented face is a different face.
Our controlled pair: same model, same prompt, input dropped from 2,115,195 bytes to 3,889 bytes. Identity survived the first, not the second.
Best-documented free tier: Let's Enhance — 10 credits on signup, no card, and it publishes a size ceiling (512 MP). Most of the others publish nothing.
Before you upload: find the original file. A screenshot of a photo, or a copy that went through a messaging app, is the version that gets rebuilt.
Before you pick the best free AI old photo restoration tool: what one degraded input does
Same photo restoration model. Same prompt, word for word, both times. Only the input changed.
| Feature | Original | From the original file | From 3,889 bytes |
|---|---|---|---|
| Mole below the right eye | Outer, under the eye | In place | Moved inward |
| Brow asymmetry (left higher) | Present | Kept | Gone — thinner, arched |
| Jawline | Square | Square | Slimmed to a point |
| Lips | Thin | Thin | Smaller, reshaped |
| Recognisably the same person | — | Yes | No |
Why nobody else runs this test
Every free AI photo restoration comparison shows you an after. Showing a before and after proves the tool sharpened something. It cannot tell you whether the face that came back is the face that went in, because the damaged photo is the only reference anyone kept.
We generated the original ourselves and kept it. That is the whole trick: the ground truth exists, so "did it change her face" becomes a question with an answer instead of an impression.
What it means for a real old photo
The prompt forbade beautifying and slimming. It happened anyway — not because the model ignored the instruction, but because at 240 × 240 those features are not in the file. The model cannot preserve what it cannot see, so it fills in. In this pair the variable that decided the outcome was the input, not the model or the prompt — both of those were held constant.
What the best free AI old photo restoration tools publish before you upload
Every figure read from the vendor's own site on September 7, 2026. Not stated means the vendor does not publish it where you would look before uploading a family photo.
| Photo restoration tool | Works without an account? | Free allowance | What it repairs | Size ceiling | Watermark on free output | Checked |
|---|---|---|---|---|---|---|
| Let's Enhance | Signup, no card | 10 credits on signup | Enhance, upscale | Up to 512 megapixels | Clean file on paid plans | Sep 7, 2026 |
| Kavel | Yes — 15 credits as a guest | 15 guest + daily check-ins | 5 models, switchable | 1K / 2K / 4K | Clean downloads on any pack or plan | Sep 7, 2026 |
| restore-old-photos.com | Has a free try | Not stated | Repair · scan cleanup · colorization | 1K and 2K only — no 4K | Not stated | Sep 7, 2026 |
| Adobe Firefly | Not stated | Free tier | Restoration | Downloads capped at 2000 × 2000 | Not stated | Sep 7, 2026 |
| Nero AI | Not stated | Pay-as-you-go credits | Upscale, enhance | Not stated | Not stated | Sep 7, 2026 |
| Picsart | Not stated | Free tier | Restore old photos | Not stated | Not stated | Sep 7, 2026 |
| ImageHub | Not stated | "Free plan", amount not stated | Editor, watermark removal | Not stated | Not stated | Sep 7, 2026 |
| Bylo AI | Not stated | "Free" | Multi-model editing | Not stated | Not stated | Sep 7, 2026 |
| Canva | Not stated | Free templates | Restore old photos | Not stated | Not stated | Sep 7, 2026 |
Across nine photo restoration tools, two publish a size ceiling and two publish a free-credit number. Six of nine say nothing about whether the free file comes out watermarked — which you discover after the restoration, not before.
That gap matters more in photo restoration than in other categories. A watermarked logo generator costs you an afternoon. A photo restoration you cannot use is a photo you had already decided was worth saving.
The best free AI old photo restoration you can run offline
Two of the sub-queries an answer engine runs on this topic ask about local, offline restoration. No vendor answers that, because offline does not bill.
| Project | Stars | Last commit | Note |
|---|---|---|---|
| GFPGAN (TencentARC) | 37,672 | 2024-07-26 | The one every listicle names — and it has not been touched in over a year |
| CodeFormer (sczhou) | 18,127 | 2025-11-18 | Half the stars, still maintained |
Both are offline photo restoration models you run on your own machine, so the photo never leaves it. Both list their licence as NOASSERTION on GitHub, which means you check the repository before any commercial use.
The star count is the trap in offline photo restoration. GFPGAN is the name that shows up in every roundup, and its last commit predates most of the tools in the table above. If you are choosing a local model today, popularity is pointing you at the stale one.
Three things to do before any AI photo restoration run
- Find the original file. Not a screenshot of it, not the copy someone sent through a messaging app. Those are already the compressed version, and compression is what makes the model invent.
- Decide what is actually damaged. For photo restoration, scratches, tears and fading are physical damage sitting on top of intact detail — that is recoverable. Lost resolution is not damage, it is absence, and absence gets rebuilt.
- Run it more than once, on more than one model. Faces drift differently on each photo restoration pass. A single-pipeline tool gives you one attempt at one interpretation.
That last point is where Kavel's photo restoration page differs from the single-purpose tools in the table: Nano Banana 2, Nano Banana Pro, GPT Image 2, Seedream 5.0 and Qwen Image 3 sit in one dropdown, so the same scan runs through several models and you pick the one that still looks like her. It starts at 15 credits as a guest — no account, no card — and the credit cost of each run shows before you commit. Pricing is here.
How we checked the best free AI old photo restoration tools
Vendor pages were fetched directly on September 7, 2026; GitHub figures come from the repository API the same day. The restoration pair was generated with gpt-image-2 image-to-image, one original and one hand-degraded copy, identical prompts.
Two limits, plainly. One controlled pair, one model — it demonstrates that this model alters identity on a degraded input, not that every tool does, and it is not a score for any vendor in the table. And not stated is a gap in disclosure, not a verdict: a tool may well output 4K without a watermark while saying none of it on the page you read first.
FAQ
Can ChatGPT restore an old photo?
It can make one sharper. In our pair it also changed the face: from a 3,889-byte input the model returned a clean portrait with the mole in a different place, a narrower jaw and the brow asymmetry removed — with a prompt that explicitly forbade beautifying. That is one run on one model, not a measurement of every photo restoration tool. Sharpness is not the same as fidelity.
Which free AI photo restoration tool works without signing up?
Kavel gives 15 credits to guests with no account and no card. Let's Enhance gives 10 credits after a signup that does not ask for a card. The other seven in the table do not state their account requirement on the page.
Why does the restored photo not look like my grandmother?
Because the detail that made her recognisable is not in the file any more. Photo restoration models reconstruct missing regions from what faces generally look like, so the smaller and softer your input, the more of the result is invention rather than recovery.
Can AI recover detail that is not in the file?
No. It can produce plausible detail, which looks like recovery and is not. This is the difference between repairing a scratch — where the information beneath is intact — and upscaling a 240-pixel face, where it is gone.
Should I upscale the photo first?
Upscaling before photo restoration adds a second round of invented detail for the restorer to treat as real. Feed the largest original you have and let one model do one pass.
Is there an offline AI photo restoration tool?
Yes — GFPGAN and CodeFormer both run locally, so nothing is uploaded. CodeFormer has half the stars and a far more recent commit (November 2025 versus July 2024). Both list NOASSERTION as their licence, so check the repository terms before commercial use.
Do these tools keep my uploaded photos?
None of the nine states its retention policy on the photo restoration page itself. For a family photo that is worth reading the privacy policy before uploading, or using one of the offline models above.
Vendor terms, prices and repository figures read on September 7, 2026. Vendors change pricing without notice — check the current page before paying.
