How to Write AI Image Editing Prompts That Preserve the Rest of the Photo
The phrase ‘change only this’ is rarely enough. A dependable editing prompt names the intended change, the protected details, and the visual quality that makes the new content belong.
The change–preserve–quality structure
Write the prompt in three parts. Keep each part concrete and observable.
- Change: name the object, region, or visual property and its desired result.
- Preserve: list the identity, layout, text, product, pose, or lighting details that matter.
- Quality: explain how the new content should integrate—natural edges, matching perspective, consistent light, or the same grain.
“Change the jacket to a tailored navy blazer over a plain white shirt. Preserve the person’s face, body shape, pose, hands, watch, background, and camera angle. Keep fabric folds realistic and match the existing light direction and image grain.”
Name what the model might otherwise reinterpret
Models often treat an editing request as permission to improve the entire image. The protected list narrows that permission. For portraits, name identity, expression, hairstyle, skin tone, body position, hands, clothing that should remain, and the camera angle. For products, name silhouette, proportions, labels, logo placement, materials, reflections, and shadows.
Background replacement
“Replace only the background with a bright modern office, softly out of focus. Keep the subject, hair edges, face, clothing, pose, foreground, and camera framing unchanged. Match the office light to the subject.”
Object removal
“Remove the cable running across the lower floor and continue the wooden floorboards with matching direction, texture, perspective, and shadow. Keep every person and piece of furniture unchanged.”
Hairstyle change
“Change the hairstyle to a short textured crop with a natural hairline. Preserve the face, identity, expression, facial hair, ears, head angle, clothing, background, and lighting.”
Avoid conflicting instructions
“Make it dramatic but keep the lighting unchanged” forces the model to choose between two goals. Decide whether the drama should come from color, weather, camera perspective, or light. Likewise, “keep everything identical but turn it into a cartoon” is contradictory because a cartoon transformation must reinterpret texture and edges.
Build the preservation list from the image type
“Keep everything else unchanged” is easy for a person to understand and difficult for a generative model to execute. The model needs observable anchors. Choose the details that carry identity, commercial meaning, or composition, then name them in a compact list. The best list changes with the source.
For portraits and hairstyle previews
Protect identity, facial proportions, expression, skin tone, eye color, head angle, hands, body position, clothing, accessories, camera framing, background, and light when they matter. If the hair is the requested change, describe length, texture, fringe, part, volume, and color separately. A useful preservation line might say: “Keep the face, expression, skin tone, head angle, ears, clothing, background, and lighting unchanged.”
For products and packaging
Protect silhouette, proportions, viewpoint, scale, materials, hardware, seams, labels, logo placement, color, reflections, contact point, and shadow. Product prompts deserve stricter review because a visually small change can create a commercially inaccurate representation. If label text must be exact, plan to composite the original label or use a manual editor instead of relying on generated text.
For interiors and architecture
Protect walls, windows, doors, floor lines, ceiling lines, built-in elements, perspective, vanishing points, camera height, and permanent fixtures. Name furniture that must stay. When removing clutter, state the surfaces that should continue behind it. “Reconstruct the rug border, baseboard, wall, and floorboards with matching direction and perspective” is more useful than “fill naturally.”
For style transformations
Protect composition, subject placement, silhouette, horizon, camera angle, and recognizable landmarks. Give the new medium its own vocabulary: line weight, palette, texture, material, brush or ink behavior, realism, contrast, and light. A style prompt cannot preserve every original surface, so focus on the structure that makes the source recognizable.
Use nouns and relationships instead of praise words
Words such as beautiful, premium, professional, dramatic, clean, and high quality express preference but do not define the image. Replace them with visible relationships. “Professional product photo” can become “neutral seamless background, soft contact shadow, controlled highlight on the left edge, accurate material color, and no extra props.” “Dramatic city scene” can become “cobalt and amber storefront light, wet asphalt reflections, dark blue sky, and preserved building geometry.”
Quality terms are most useful at the end of a prompt when they explain integration: matching perspective, natural edges, consistent grain, realistic shadow, the same light direction, or believable depth of field. They should not carry the entire request.
A prompt can be short and still be complete
Long prompts are not automatically precise. A dependable instruction can be three sentences: one for change, one for preservation, and one for integration. Remove repeated adjectives and any instruction that cannot be checked in the resulting image. If two sentences conflict, split the task into versions.
“Replace the plain studio with a coral architectural set and an acid-yellow platform.”
Preserve“Keep the cobalt chair, silhouette, seams, legs, material, scale, and camera position unchanged.”
Integrate“Add hard editorial light and a contact shadow that follows the platform perspective.”
Every sentence now creates a review step. The set and platform can be judged, the product can be compared with the source, and the contact shadow can be checked against geometry. If the result fails, the next instruction can address the failed part without rewriting the entire brief.
Troubleshoot by the kind of drift
| Observed problem | Likely prompt gap | Next action |
|---|---|---|
| The requested object changed, but the face changed too | Identity was not protected clearly enough | Return to the parent; add face, identity, expression, and skin tone to the preservation line |
| The product looks attractive but has different hardware | The product was treated as a style reference | Name silhouette, hardware, seams, material, viewpoint, and scale; use a tighter editor task |
| A removed object leaves a blurry patch | The hidden surface was not described | Name the wall, floor, fabric, pattern, line, perspective, and shadow that should continue |
| A new background looks pasted behind the subject | Lighting and boundary integration are missing | Specify light direction, depth of field, color spill, contact shadow, and natural edge treatment |
| A style change loses the composition | The prompt describes the medium but no structural anchors | Protect camera angle, framing, subject placement, horizon, and major silhouettes |
Separate the first pass from the correction pass
The first prompt should solve the largest visual decision. Once the result is usable, switch to marked refinement for the remaining local problems. A marker supplies location, so the marker text can concentrate on the visible outcome. Put shared preservation rules into the whole-image direction.
“Continue the rug border through this area with the same weave, width, perspective, and shadow.”
Whole-image direction“Keep the cleaned room, chair, table, plants, window, floor, camera angle, and daylight unchanged.”
After generation, review the marker, its immediate context, and the complete image. If the model repeatedly changes a detail that must be exact, return to the parent version or use a traditional mask. More prompt length cannot turn a generative system into a pixel lock.
Keep a reusable prompt brief
For recurring product, portrait, or campaign work, save four lines outside the editor: the primary change, the protected details, the integration standard, and the review checklist. Update only the task-specific nouns. This keeps instructions consistent across versions without producing dozens of identical pages or treating one generic prompt as suitable for every image.
- Change: one observable transformation.
- Preserve: the identity, geometry, text, composition, and light that carry value.
- Integrate: edge, perspective, shadow, grain, color, and depth relationships.
- Review: the exact details that would make the result unusable if they drifted.
Keep the brief specific to one workflow. A portrait preservation list should not be copied into a product photo, and an architecture checklist should not be added to every style transformation. Reuse the structure, then rewrite the nouns and relationships from the current image. This is how a prompt remains consistent without becoming generic.
Use a second pass for precision
When the broad result is usable, stop expanding the original prompt. Open marked refinement and point to the exact unresolved areas. Give each marker one instruction and add a short whole-image preservation line. This reduces the number of decisions in a single request.
A strong editing prompt sets boundaries for creative freedom. It tells the model where invention is welcome and where continuity matters more.
Review the details that carry risk
Before downloading, zoom in on faces, hands, product labels, text, jewelry, transparent edges, architectural lines, and reflections. Generative editing can produce a convincing overall image while changing a small detail that matters commercially or personally.