Use a repeatable prompt template
For Text to Image, fill in these slots in order:
- Subject — who or what is in frame
- Action or pose — what is happening
- Environment — where the scene takes place
- Lighting — time of day, key light direction, mood
- Style/medium — photo, illustration, 3D render, etc.
- Camera — lens feel, angle, depth of field
Prompting for Image to Image
Lead with preservation clauses, then changes: “Preserve face and pose. Change outfit to navy blazer. Studio gray backdrop.”
If the model drifts too far, narrow the change list to a single edit per generation.
Iterate without burning credits
Duplicate what worked in the last prompt and adjust one variable (lighting, color palette, or camera only).
Use Nano Banana 2 for exploration; switch to GPT Image 2 for the final hero asset.
What to avoid
These patterns often cause weak or inconsistent results:
- Vague superlatives only (“amazing,” “epic”) with no visual detail
- Conflicting styles in one prompt
- Requesting readable text or logos without specifying placement and font style
- Extremely long prompts that bury the main subject
Category: Image generation