Why does the model still re-style my image sometimes?

Image-to-image models have a creative bias. To force strictness, repeat the preservation clause, list edits as imperatives, and use the polish-neurips-diagram template style.

PromptsPolish

PolishImage → ImageLow creditEN

Strict Label Rename / Value Edit

Surgical text edits to an existing diagram while preserving every other element.

When to use this prompt

When you need to fix a typo, swap a value, or rename labels in an existing figure without redesigning anything.

The prompt

Apply only the following surgical edits to this image. Do NOT redesign, do NOT reinterpret, do NOT add new elements.

Edits:
1. Replace the text "creatinine clearance < 30" with "creatinine clearance > 30" in the same font, size and color as the surrounding labels.
2. Replace the caption text "Tabaco use" with "No tobacco use" in the same font and size.
3. Increase the overall font size of all labels by 10% for accessibility, while keeping all box positions and proportions unchanged.

Preserve everything else exactly:
- Box positions and proportions
- Arrow directions and connections
- Colors and styles
- Background

Repeat: do NOT change anything besides the listed edits. This is a strict micro-edit, not a creative regeneration.

Variations

Numeric value swap

Apply only this edit: change every occurrence of "< 30 mg/dL" to "≥ 30 mg/dL" in the table. Preserve fonts, colors, alignments, table structure exactly. Do NOT add anything else.

Tips

List every edit as a numbered, atomic instruction. Vague edits invite redesigns.
Repeat the preservation clause near the end of the prompt. Models otherwise drift toward creative changes.
For accessibility tweaks (font size), give an exact percentage rather than "make readable".