“Make Them Look NEW Again” Is The Wrong Goal For A Family Photo
“Your family’s oldest photos deserve a second life.” The card is a single slide: “ChatGPT can RESTORE your old photos. Faded. Blurry. Damaged. Use these 6 prompts to make them look NEW again 👉” The caption promises to restore damage, sharpen blurry faces and bring black-and-white memories to life.
A note on sourcing before anything else: the six prompts are not visible in the post. It is a single cover image with a pointing finger, no comments, and no further slides we could reach. So this is not a review of six prompts. It is about the one word in the headline that decides whether any of them is a good idea — and about the sentence in the smallest type on the card, which is the best thing on it.
The source post — “Your family’s oldest photos deserve a second life”, BANI Academy. A cover slide for six photo-restoration prompts, with this fine print: “Results vary by photo and tool. Keep a copy of your original, and only upload photos you have the right to use.”
Two operations wear the same word
“Restore” covers two completely different things, and the difference is not about how good the tool is. It is about whether the answer is still in the file.
Sometimes it is. A scratch across a plain wall can be repaired honestly, because the wall continues on both sides of the scratch and the software can read it off the neighbouring pixels. Fading is the same kind of problem: the relationships between tones survive even when the overall contrast has collapsed, so stretching them back is arithmetic, not invention. Dust, a crease through a background, a yellow cast — all recoverable, because the information was never destroyed, only buried.
Sometimes it is not. A face that occupies forty pixels does not contain a face. The nose is two pixels. The eyes are smudges. There is no amount of processing that recovers a jawline from information that was never captured, and a model asked to sharpen it does not recover anything — it generates a plausible face and places it where the smudge was.
That second operation has a name, and the name is not restoration. It is reconstruction, and it produces a photograph of somebody who looks like they could have been your grandmother.
Diagram: is the answer still in the file?
| The damage | Is the information still there? | What the tool does | How much to trust it |
|---|---|---|---|
| Dust, specks, small scratches | Yes — the surrounding pixels continue through | Reads the neighbours and fills the gap | High. This is repair in the ordinary sense |
| Fading, colour cast, low contrast | Yes — the tonal relationships survive | Stretches the range back out | High. Arithmetic on what is already recorded |
| A crease or tear across a background | Mostly — the pattern continues either side | Infers the continuation | Good, with a glance to check the pattern lines up |
| A tear across a face or hands | No — what was under the tear is gone | Invents what it thinks belongs there | Low. This is a drawing, placed over your photograph |
| A blurry or tiny face | No — the detail was never captured | Generates a sharp, plausible face | Lowest. Nothing was recovered |
| Black and white | No — colour was never in the negative | Guesses every colour in the frame | It is an illustration. Often beautiful. Never a record |
Why this matters more for your photos than for anyone else’s
If a model invents a face in a stock photograph, nobody is harmed and nobody can tell. Your family archive is the opposite case on both counts.
Nobody can check it, and you may be the last person who could. A restored face of a public figure would be caught within a day by people who know the face. A restored face of your great-grandfather is checked by nobody, forever, because the two or three people alive who remember his face are the people deciding whether the restoration looks right — and a generated face is built specifically to look right.
And the output is the version that survives. This is the quiet mechanism. The sharp, colourful, undamaged version is the one that gets shared in the family group, printed for the wall, and attached to the memorial slideshow. The scanned original stays in a folder nobody opens. Thirty years from now the generated version is the photograph, and the thing it replaced was the evidence.
None of that is an argument against doing it. It is an argument for doing it the way the card’s fine print already tells you to.
The best sentence on the card is in the smallest type
Under the logo, in grey, at a size most people will not read:
“Results vary by photo and tool. Keep a copy of your original, and only upload photos you have the right to use.”
Four words — keep a copy of your original — settle the entire argument above. If the scan survives, the generated version is a nice print. If it does not, the generated version is now the only record and nobody will ever know which parts were invented.
The second half matters too, and in a different direction: photographs of other people’s families are other people’s. Uploading a shared family photo means uploading pictures of relatives who did not choose that.
What comes back is marked, on purpose
One more thing worth knowing before you put a restored image into an archive, because it is both reassuring and easy to be surprised by.
OpenAI’s own content-provenance documentation describes two signals on its images: C2PA signed metadata attached to the file, and a SynthID watermark embedded in the pixels. It notes that editing, converting or sharing a file can strip the metadata, while the embedded watermark may survive some transformations. OpenAI has also said that when a user edits an image, the credential is built to show that it was edited and how.
Google describes the same arrangement for Gemini: media created or edited with its apps carries an invisible SynthID mark, designed to survive cropping, filters and lossy compression. In August 2026 Google made the visible sparkle watermark optional while keeping SynthID and C2PA in place.
So a restored photograph arrives carrying, invisibly, the fact that a machine altered it. For a family archive that is a feature, not a problem — it is the archive recording its own provenance, which is exactly what you want a hundred years out. The failure mode is only the one above: an unmarked, unlabelled generated image that replaced a scan nobody kept.
The three promises in the caption, sorted
| The promise | What it really is | Do it? |
|---|---|---|
| “Restore damaged photos” | Repair where the damage is in a background; invention where it crosses a face or hands | Yes — and look specifically at what was under the damage |
| “Sharpen blurry faces” | Generation. The detail was never captured, so a new face is drawn | Carefully. Lovely as a portrait, worthless as a likeness |
| “Bring black-and-white memories to life” | Guessing every colour in the frame, from clothes to eyes | Yes, and label it. It is an interpretation, like a colourised film |
| “Make old images clearer” | Mostly honest — contrast, noise, sharpening of detail that exists | Yes. The safest of the four |
| “Without losing the details that make them special” | The claim that cannot be checked by looking, since a convincing invention looks like a preserved detail | This is the one to verify yourself |
A workflow that keeps the evidence
None of this requires restraint or extra software. It requires filenames.
| Step | What to do | Why |
|---|---|---|
| 1. Scan first, at the highest setting you can | Save as surname-1952-wedding-SCAN.tif or the best quality your scanner offers |
The scan is the evidence. Everything after it is a version of the scan, not a replacement for it |
| 2. Never overwrite it | Put the scans in their own folder and treat it as read-only | The single failure this whole article is about is the original quietly ceasing to exist |
| 3. Name what you did | …-RESTORED.jpg for repair, …-COLOURISED.jpg for colour, …-REBUILT.jpg where a face was generated |
A filename is the cheapest provenance record there is, and it survives being emailed to relatives |
| 4. Compare at full size before you share | Open scan and output side by side at 100%, and look at faces, hands, text, jewellery and patterns | Those are where invention shows. A background repair almost never surprises you; a face often does |
| 5. Say so when you share it | One line in the family group: “restored with AI, original scan is in the folder” | The relative who notices something is off deserves to know why — and so does whoever inherits the folder |
Step 4 is the one worth dwelling on, because it is the same check that catches generated images everywhere. Text, hands, jewellery and repeating patterns are where a generative model gives itself away — an inscription on a trophy becomes letter-shaped marks, a wedding ring gains a facet, a lace pattern stops repeating properly. We made the same point about a card whose captions turned out to be generated: zoom in on the smallest detail, because that is where the texture stops pretending to be information.
And it is the same distinction we drew about video: the difference between adjusting footage you shot and generating footage nobody filmed. Photographs just make the stakes personal.
Four honest caveats
- We could not read the six prompts. The post is a single cover image with no comments and no further slides reachable without a login. This article is about the operation the card describes, not about the quality of its six prompts, which may be excellent.
- Dedicated restoration tools are not all the same as a chatbot. Some denoisers and scratch-removers do the honest, local kind of repair described above and nothing else. Some face-enhancers generate. The useful question is not which brand but which of the two operations is happening to a given part of the picture.
- Provenance details move. The C2PA and SynthID points come from OpenAI’s own developer documentation and Google’s published statements, read in October 2026. Both companies have changed what is marked and what is visible within the last year, so check the current pages before relying on a specific behaviour.
- None of this says do not do it. A faded photograph that nobody can make out is not being preserved by being left alone. Restoration is a good thing to do with these tools. The argument is only about which copy you keep.
The one-line version
Some damage can be repaired because the answer is still in the file; some cannot, and there the tool draws something plausible instead. Both come back looking equally finished, your family photographs are the one set of images nobody else can check, and the generated version is the one that gets printed and shared.
Which is why the smallest sentence on the card is the only one that has to be followed: keep a copy of your original. Do that and everything else is upside. Skip it and, in thirty years, the guess is the record.
Do this one thing first. Before you upload anything, make a folder called SCANS — DO NOT EDIT and put the originals in it. That folder is the only part of this process that cannot be regenerated later, and it takes thirty seconds.
One Comment