AI Image Editing: The CHANGE / PRESERVE / MATCH Framework
How to edit AI images without drift, identity loss, or the 'pasted-in' look.
By Depikt Team
Why editing is harder than generating
Generating an image from a prompt is a single-shot operation. The model has total freedom — it just needs to satisfy your constraints.
Editing is fundamentally different. You're asking the model to change one thing while keeping everything else identical. That's a much harder problem, because "everything else" is a massive set of visual properties: lighting angle, color temperature, skin tones, fabric texture, shadow direction, grain structure, focal plane, and thousands of other details you don't consciously notice until they change.
Without structure, edits drift. Skin tones shift. Backgrounds bleed into subjects. Hair changes texture. Shadows point the wrong way. The result looks "pasted in" — which is worse than just generating from scratch.
The framework: CHANGE / PRESERVE / MATCH
Every edit prompt should have exactly three blocks:
CHANGE: The one thing (or small set of things) you want different. Be specific about what replaces what.
PRESERVE: An explicit, exhaustive list of everything that must stay exactly the same. If you don't name it, the model may change it.
MATCH: The lighting, color, and grain logic that bridges the old and new elements. This is what prevents the "pasted in" look.
Why three blocks, not two
Most people write edit prompts with just CHANGE and PRESERVE. They skip MATCH — and that's where the uncanny results come from.
When you swap a studio background for a rainy street, the subject's lighting shouldn't stay studio-flat. The new environment should cast subtle color spill onto the subject — cool blue from neon, warm orange from streetlights. Without the MATCH block, the model either keeps the original lighting (subject looks pasted) or re-lights everything (subject's face changes).
MATCH tells the model: keep the original lighting direction and intensity on the subject, but let the new environment influence color temperature and reflections naturally.
Example 1: Background swap
Rough idea: "Put this portrait in a rainy Tokyo street at night"
Example 2: Object removal
Rough idea: "Remove the trash can and power lines"
Example 3: Style transfer
Rough idea: "Make this photo look like a watercolor"
Multi-turn editing: the drift problem
Single edits are manageable. The real challenge is multi-turn editing — making 3, 4, 5 changes to the same image across multiple generations. Each turn introduces small drift, and drift compounds.
By turn 3 or 4, the subject's face has subtly changed, the lighting is inconsistent, and the image has lost coherence.
The fix: restate PRESERVE every turn. Don't rely on the model remembering what to keep from previous turns. Copy-paste your full PRESERVE block into every edit prompt, updated to include anything new you added in previous turns.
The second fix: work from the best previous output, not the original. Each edit should reference the most recent good result, not the original image. This keeps the edit chain grounded.
When to edit vs. when to regenerate
Edit when:
- You like 80%+ of the image and want to fix one element
- The composition, pose, and lighting are right but a detail is wrong
- You need to swap a background or remove an object
- You're doing style transfer on a specific composition
Regenerate when:
- The composition is fundamentally wrong
- Multiple major elements need changing
- The pose or framing doesn't work
- You're on edit turn 4+ and drift has accumulated
Regenerating with a refined prompt is often faster than fixing a broken edit chain.
Common editing failures and fixes
"Pasted in" look: You're missing the MATCH block. Add color spill, reflection, and grain matching from the new environment.
Face drift after 2-3 edits: Your PRESERVE block isn't specific enough. Add: "facial bone structure, eye color, skin texture, exact hairline."
Shadow inconsistency: Name the light direction in MATCH. "Shadows cast to lower-left, consistent with key light from upper-right."
Background bleeding into subject edges: Add to PRESERVE: "exact silhouette edges, no feathering or blending at subject boundary."
Grain mismatch: Add to MATCH: "Match ISO grain structure and intensity of the source image."
The shortcut
Writing CHANGE / PRESERVE / MATCH blocks manually is tedious, especially the exhaustive PRESERVE list. Depikt's Critique mode can analyze your edit prompts and flag missing PRESERVE elements before you waste a generation on drift.
Build a prompt that follows this guide.
Paste a rough idea. Get back a structured prompt.
More in Guide
Depikt Can Now Generate Images
Prompt Build and Critique made it easier to write a good instruction. The step that was still missing was making the image. Generate closes that gap: create, edit, and regenerate images from a prompt and optional references, without leaving Depikt.
ProductYour AI Assistant Can Now Use Depikt
Depikt is now available inside AI assistants. Connect it through MCP and let ChatGPT, Claude, and other compatible assistants search the prompt library, retrieve complete prompts, browse templates, and use Depikt’s guides while helping you create.