Editing One Element with AI — Without Everything Else Changing
ImageFactory Engineering · Published 2026-06-12
You asked AI to "just change the button" and the model also changed the face — because generative editing redraws the whole canvas even for a small request. Two things fix it: state what must stay the same, not just what changes; and specify the target by location, appearance, and what the spot should look like afterwards. In our measurements, name-only instructions ("remove the badge") applied 0 of 12 times; fully specified instructions applied 9 of 9 (100%).
Why everything else changes
Generative editing doesn't manipulate objects on layers like Photoshop. The model paints a new complete image, and any region you didn't mention carries no obligation to stay the same. The vaguer the instruction, the more discretion the model takes — and discretion is what changes your untouched elements.
Instruction patterns that decide the apply rate (measured)
| Instruction | Result |
|---|---|
| "Remove the sale badge" (name only) | 0/12 applied — the model can't pin the pixels |
| "Remove the red circular badge in the top-left (with SALE 50% text), and fill that spot naturally with the surrounding background gradient. Keep every other element unchanged." | 9/9 applied |
As rules:
- Pin the target with position (think 3×3 grid: top-left, center, bottom-right…) and appearance (color, shape, contained text).
- Describe the after-state — not "remove it" but "remove it and fill with the background".
- Say "keep everything else unchanged" explicitly. Preservation is not the model's default.
- If a face or logo absolutely must not change, pin it the same way — by description, not name.
How ImageFactory solves this
ImageFactory's workspace editor applies these rules automatically. Your image is analyzed on upload — element positions and appearances mapped — so when you type a short "remove this badge", it's internally expanded into a fully specified instruction with location, description and after-state. The 0%→100% measurement above is the experiment that validated this conversion. You write one short sentence; everything else stays put.