Photo restoration fails in a way that looks like success. The scratches disappear, the colours come back, and nobody can tell which of the faces the model invented. GPT Image 2.5 can repair a scan, so the useful skill is deciding what it may touch.
What GPT Image 2.5 changed
Nothing on the launch page describes a restoration feature, and the word does not appear there. What the page does describe is the behaviour restoration depends on: GPT Image 2.5 is “better at editing only what you’ve asked for, while keeping the rest of the details the same—even with more complex subjects and backgrounds.”
The second useful behaviour is continuity: over a longer conversation the model “follows specific editing instructions more reliably across multiple edits”, and “each new edit builds on the work you’ve already done”. That is the argument for repairing one scan across several passes in a single thread instead of one enormous prompt.
Two models are named. GPT-Image-2.5 Flare is “the default choice for most applications”; GPT-Image-2.5 Sunburst is “built for premium visual workflows that benefit from tighter control across edits”, which suits a multi-pass repair. No official page publishes a price per image or a usage quota.
Diagnose the damage first
Write down what is wrong before you write anything the model will read. Damage falls into four groups, and each wants a different order of work.
| Damage | What it is | Repair it in which pass |
|---|---|---|
| Physical marks | Scratches, dust, creases, torn corners | First, alone |
| Tonal decay | Fading, a colour cast, flat contrast | Second, after the marks |
| Missing colour | A black-and-white or sepia original | Last, once the tones are right |
| Destroyed information | A face, a sign or an edge the scan never had | Never — leave it out |
The last row is the whole discipline. A scratch is recoverable because the information around it survives; a face that was never in the scan is not, and any model asked to fill it will produce a plausible stranger. Decide which row each defect belongs to before you start.
Repair in passes, not in one prompt
One prompt that asks for everything at once gives the model permission to rebuild the picture. Split the work instead, and let each pass finish before the next one starts.
PASS 1 · clean: remove the scratches, the dust specks and the creases only. Change nothing else.
PASS 2 · tone: lift the contrast, neutralise the colour cast, keep the original grain and the soft focus.
PASS 3 · colour: add colour only where the original already shows a value or a texture to follow.Two rules keep the passes honest. Every pass repeats the same guard sentence: do not add anything that is not already visible in the scan. And check the boundary after a local repair, because the help page warns that “edits may extend beyond the area you selected” — a scratch removal that reaches into the cheek has quietly repainted skin.
Two examples: faded print and scan
Both prompts below are complete and copy-pastable, and both are written as single passes on purpose.
One: the faded print
A square scan of an old family photograph, faded and washed out.
Repair pass only: remove the dust specks and the two creases across the lower left, and keep every other mark.
Do not add anything that is not already visible in the scan: no new faces, no new objects, no invented background detail.
Keep the original grain and the soft focus of the print.
No colour changes in this pass, no text, no watermark.Read the result against the original at the same size. Fading is uneven, so the corner the model brightened most is usually where it started inventing. If the guard sentence was ignored, edit that region back rather than re-running the pass.
Two: the scratched scan
A vertical scan of a portrait print with a long scratch across the upper half.
Remove the scratch and the small tears along the right edge, and rebuild only the surface texture they interrupted.
Match the surrounding tone exactly; do not redraw the face, the hair or the clothing.
Leave the print border, the fold line and the age spots that belong to the original print.
No text, no watermark, no added props.Tears along an edge are the hardest case, because the model has to continue a texture rather than copy one. Keep the fold line and the age spots: they are evidence that the object is real, and removing them turns a restored photograph into a reconstruction.
Colourise without inventing
Treat colour as a decision made from evidence in the picture, not from what the era probably looked like.
Colourise last, after the tones are right, so you are judging colour instead of compensating for a cast. Then constrain it: name the colours you can defend from the scan and refuse the rest. A uniform whose shade survives as a mid-tone can take a colour; a coat recorded only as flat black cannot. Asking for “realistic period colours” invites a guess, and the guesses are what a viewer later mistakes for evidence.
Keep the grain as well. A pass that smooths everything leaves a picture that is clean and no longer photographic. The editing guide covers region-by-region work, and the quality levels guide explains why a higher setting is no substitute for a smaller correction.
What restoration cannot fix
| You want | Why it cannot be done safely | What to do instead |
|---|---|---|
| A face the scan never captured | The model can invent only a plausible stranger | Leave the area unresolved |
| Text that is illegible | Reconstructed glyphs become confident claims | Reproduce what is legible, no more |
| Original colour of a garment | The scan carries values, not hue | State the uncertainty by the image |
| A frame cropped away | Everything outside the scan is fiction | Reframe as a crop, or show the border |
Two things you cannot see travel with the file as well. The system card describes “a continued commitment to C2PA metadata” plus an invisible SynthID watermark, so a restored image still announces that it was processed. The same card states that ChatGPT Images 2.5 allows “heightened realism”, and that imagery of real people in sensitive or compromising contexts violates the usage policy — restoring a family photograph is not the same as rebuilding an official document.
Frequently asked questions
Is photo restoration free?
No official page publishes a usage quota or a per-image price, so treat both as unknown and budget your own review time.
One photo or several?
One photograph per thread, repaired over several passes. A reference image keeps the model anchored; the combine images guide covers what changes when you add a second.
How do I stop the face changing?
Ask for repair only, repeat the guard sentence in every pass, and inspect the boundary of each edit. If a pass has redrawn features, undo it and re-run on a smaller region rather than accepting the version that looks better.
Does colourising count?
It is a separate claim. Repair returns the scan toward the object it recorded; colour adds information the scan does not contain. Label the coloured version as a colourisation, and keep the repaired black-and-white file as the record.
One photograph, one checklist
Classify the damage, repair the physical marks alone, fix the tones, then colourise last — all in one thread on one scan. Compare the result with the original at the same size, check that nothing outside the damaged area moved, and record which parts are reconstruction rather than recovery. The transparent background guide finishes the job when the print is cut out for reuse, and the prompt examples show the guard sentence phrased elsewhere.