How to write photo editing prompts that preserve the subject

When an AI edit goes wrong, the eye usually blames the new background, the new lighting, or the new colour grade. But the more common failure is quieter: the model regenerated the subject itself. The face is a little younger, the logo is a little off, the chair has one too many legs. The effect you asked for arrived — attached to a subject that is no longer quite the one you uploaded.

A good editing prompt is really two prompts braided together: the change you want, and the things that must not change. Most people write the first half with care and skip the second entirely. The fix is to make the preserve clause a first-class part of every prompt, not an afterthought.

Why models drift from the original

Image models do not edit pixels the way a retoucher moves a slider. They re-generate the picture toward your description, using the original as a strong hint rather than a locked layer. The more your instruction implies a fresh scene, the more freedom the model takes — and freedom near a face, a hand, or a brand mark is exactly where identity leaks out.

That means the words you choose set how much licence the model feels it has. "Make this look cinematic" invites a rebuild. "Keep the subject and pose exactly as they are; only regrade the colour to a cool cinematic tone" fences the change in. Same goal, very different blast radius.

Name what must stay the same, specifically

Vague preservation ("keep it realistic") does almost nothing, because the model already thinks it is being realistic. Preservation works when it is concrete and lists the parts you care about by name:

• Identity — "keep the same face, age, skin texture, and expression; do not slim or smooth the face."
• Geometry — "keep the same pose, framing, and crop; do not zoom, rotate, or re-centre."
• Product truth — "keep the label text, logo, colour, and proportions exactly; do not redesign the packaging."
• Scene anchors — "keep the same room, furniture, and window positions; only change the wall colour."

The pattern is the same every time: say the edit, then say the explicit do-not list. When the model knows which regions are off-limits, it spends its freedom on the part you actually wanted changed.

Edit in small passes, not one big leap

A prompt that asks for five changes at once gives the model five excuses to redraw the subject. You will get more faithful results by stacking small, reversible passes: fix the background first, confirm the subject survived, then adjust the colour, then sharpen. Each pass is a checkpoint where you can catch drift before it compounds.

If a single pass does mangle the subject, do not argue with it in follow-up text — that tends to drift further. Go back to the original image and re-run the edit with a tighter preserve clause. Re-anchoring to the source beats negotiating with a bad generation.

Check the edit against the original, not your memory

The last step is the one people skip: put the edit beside the source at full size and compare the parts you told it to keep. Faces and hands first, then text and logos, then proportions. Memory is generous and will wave through a face that is subtly wrong. Side-by-side comparison is the only honest check.

Key takeaways

• Every edit prompt is two prompts: the change, and the explicit do-not-change list.
• Preservation only works when it names parts by name — face, pose, crop, label, room.
• Stack small reversible passes instead of one all-at-once instruction.
• When a subject is mangled, re-anchor to the original; do not negotiate in follow-ups.
• Always compare the result to the source at full size before you trust it.