Most weak image-edit prompts fail in the same way. They describe a mood and hope the model will protect the one thing you cannot replace. "Make this look professional, cinematic lighting, ultra detailed" does not say whose face must stay, which bottle is the product, or whether the room is allowed to change. You get a prettier picture of a slightly different person holding a slightly different bottle.
A better edit prompt reads like a constraint list. This is a small set of patterns I reuse, with examples and the failure modes I now check for. I have been running them in an AI photo editor that takes one or more stills and a text instruction, because the patterns only work if you can point at a reference.
The lock-then-change sentence
Before I generate, I write two clauses. If either is vague, I do not spend a credit yet.
- Lock. The nouns that must survive: a face, hair, pose, a label, a silhouette, a room.
- Change. One verb: swap the outfit, move the background, lift the exposure, shift the season, restyle the picture.
Here is the same request written two ways: first the adjective pile, then the pattern.
Make her look professional in a modern office, highly detailed, perfect lighting, 8k.
Put her in a navy blazer over a white shirt in a bright modern office lobby, keeping her face, hair and pose.
The second sentence is useful because you can audit it. If the face moved, the lock failed. If the blazer never appeared, the change failed. If both happened, you asked for two jobs and got a compromise.
I keep the change clause singular on purpose. "Change the outfit, the city, and the weather" is three renders hiding in one prompt. When they land in one frame, you cannot tell which instruction aged the face.
Pattern 1: Outfit, with the body as a tripod
Use this when the person has to stay the same person.
Keep her face, hair, skin tone, and pose. Replace only the clothing with a navy blazer over a white shirt and dark trousers. Keep the camera distance. Do not add jewelry or change her age.
If the original photo is a hoodie on a couch, and you also want a lobby, split it: first change the clothes with the body locked, then place that result in the lobby and lock face, hair, and pose again. A single jump from couch to lobby often moves the shoulders.
Pattern 2: Background or season, with the subject nailed down
Use this when the product or person is already right.
Keep the person, the pose, and the bicycle exactly. Change only the scene to a winter street with falling snow. Late afternoon, same camera angle. Do not change the bicycle color or the rider's jacket fit.
The season line is easy to over-ask. Snow that also rewrites a red jacket into a parka is a failed lock. If I have a separate photo of the street I want, I attach it and say "use the second image as the location only." Café, beach, and lobby swaps follow the same template: name the new place in concrete nouns, and restate what is not allowed to move.
Pattern 3: Product on a new set
Use this when a phone packshot has to look like it belongs on a page.
Use the first image as the product. Keep the silhouette, the label layout, and the cap color exactly. Place it on the marble surface from the second image, soft top light, slight three-quarter angle. No extra bottles, no invented typography.
If you only have one photo, you can still name a set in prose, but the label is more likely to drift. A second reference that is just the surface, with the role spelled out, is worth more than a paragraph about luxury.
Pattern 4: Style change that is not a new identity
Use this when you want a 3D cartoon or an illustration, and you still need the same subject.
Keep the person's face, hairstyle, and the red backpack. Translate the photo into a 3D cartoon still, simple studio background, same pose. Do not change her age or add characters.
Style is greedy. If you do not lock the backpack, it becomes a more cartoon backpack. If you do not lock the face, you get a genre face. I treat style as a costume for the existing subject, not as a new commission.
Pattern 5: Enhance, without redesigning
Use this when the photo is already the right photo.
Keep the scene, the people, and the framing. Brighten the exposure, sharpen facial detail, and restore natural color. Do not change clothes, furniture, or expressions. Do not crop.
Enhancement is a repair verb. As soon as you add "and make it look like a magazine," you have switched patterns without admitting it.
Pattern 6: Combine, with jobs assigned
When several stills must become one frame, the prompt has to hire each file.
Image 1 is the character's face and hair. Image 2 is the product, including the label. Image 3 is the kitchen. Put the person at the counter with the product in the foreground, mid shot, window light from the right. Do not swap the product shape or invent a second person.
This only works in an editor that can take multiple references in one task. Attach them, then write the hiring list. An unlabelled stack of files is not a combine; it is a guess.
Failure modes I now check for
Stacked style words, no nouns. If you can delete half the prompt and lose no lock and no change, rewrite it.
Two changes in one job. Outfit plus city plus style is how faces slide. Split the jobs.
Identity described, not attached. "A woman in her thirties with brown hair" is not a person you have photographed. Attach the still.
Enhancement mixed with invention. "Sharpen and also put them in Paris" will sharpen Paris.
A previous scene used as the only lock. A scene already contains a background. Use a clean identity reference, or the background will leak.
Revising by starting over. When a take is almost right, restate the lock and describe only the remaining change.
Write the lock first, then the change. If a stranger could mark up your original from the prompt alone, the model has a fair chance.
Top comments (0)