Old Photo Restoration Negative Prompt: What AI Must Not Change

The Short Answer

A strong restoration prompt has two halves: allowed repairs and protected facts. Tell the AI to remove specific damage, then explicitly forbid changes to identity, age, expression, pose, clothing, objects, text, background, crop, and historical character. “Make this photo beautiful” is too open-ended for an irreplaceable family image.

No wording guarantees fidelity. OpenAI's official ChatGPT Images guide explains that uploaded images can be edited and that selection boundaries are not always precise. Treat every result as a candidate to inspect, not proof that lost details were recovered.

Copy This Constraint-First Prompt

Allowed repair: Restore this scanned photograph conservatively. Remove only visible dust, scratches, small stains, crease marks, and mild fading. Recover natural contrast and legibility while retaining the original photographic texture.

Do not change: Do not alter any person's facial geometry, identity, age, expression, gaze, hairstyle, body shape, hands, pose, clothing, jewelry, or skin texture. Do not add beauty retouching, modern makeup, new teeth, new eyes, or invented facial detail. Do not move, remove, or add people or objects. Preserve all text, borders, crop, camera angle, lighting direction, background, architecture, uniforms, and period-specific details. Do not colorize unless explicitly requested. If a detail is uncertain, leave it unresolved rather than inventing it.

Run one repair goal at a time when the tool allows it. Scratch removal, color correction, face reconstruction, colorization, and background repair introduce different risks; bundling them makes it harder to identify where drift began.

The Protected-Facts Checklist

• Identity: face shape, feature spacing, asymmetry, age, expression, gaze, hairline, and distinctive marks.
• People: number of subjects, position, pose, hands, clothing, jewelry, uniforms, and relationships in the frame.
• Scene: objects, signs, handwriting, buildings, vehicles, landscape, lighting, camera angle, and crop.
• History: period materials, wear, grain, borders, studio marks, and details that help date or explain the image.
• Uncertainty: missing areas that the source cannot verify. Plausible detail is not the same as recovered detail.

This checklist extends the copy-paste examples in our seven AI prompts for old photo restoration. Use the original article for task-specific prompts and this guide when preservation is the priority.

Why “Keep the Face the Same” Is Not Enough

Identity is distributed across many small relationships: distance between the eyes, nose width, lip shape, jaw asymmetry, wrinkles, hairline, expression, head angle, and the interaction between light and age. A model can preserve the general category of a face while changing the person.

Broad beautification language encourages exactly that drift. Avoid “perfect,” “flawless,” “cinematic,” “young,” “studio quality,” and “make them attractive” in a preservation prompt. Ask for legibility and damage repair, not an upgraded person.

Use Local Edits When Possible

1. Save an untouched scan before editing.
2. Select one damaged region that does not include a face when possible.
3. Describe the physical defect and the surrounding texture it should match.
4. Inspect the changed region and its edges at high zoom.
5. Compare faces, hands, text, and background objects with the original even when they were outside the intended selection.
6. Export a new version; do not overwrite the source scan.

For severely blurred faces, read how to unblur old-photo faces without changing identity. A conservative answer may retain softness because the source does not contain enough evidence for a truthful sharp reconstruction.

How to Audit the Result

• Place the original and result side by side at the same crop and scale.
• Check the face as a whole before zooming into attractive details.
• Count fingers, jewelry, buttons, insignia, windows, signs, and people.
• Read every visible word and number; generative edits often corrupt text.
• Toggle between versions to catch movement in eyes, mouth, jaw, hands, and object edges.
• Ask a relative who knew the subject to review likeness, but keep the original as the evidence.

If the result looks impressive but fails one identity check, reject it or label it as a creative reconstruction. Visual polish is not a substitute for fidelity.

Preserve an Evidence Chain

Keep the physical photograph, the untouched scan, a conservative restored master, and any creative version as separate files. Record the date, tool, prompt, and major edits. The U.S. National Archives digitizing guidance recommends keeping originals after digitization because digital copies can also be lost.

When a family photo matters more than experimentation, use the least transformative process that solves the visible problem. Explore Restore Old Photos AI for a focused restoration workflow, and always judge the result against the source rather than against an imagined modern portrait.