GPT Image 2 Complete Guide (III): Advanced prompting

Beyond structure: poses, typography in quotes, multi-image choreography, surgical constraints, and a four-step iterate loop—all for GPT Image 2.

Series：(I) Model selection & parameters · (II) Prompt basics · (IV) Generation use cases · (V) Editing use cases · (VI) Character consistency

1. People, poses, and actions

If people appear, specify:

• Figure scale / how much body is visible
• Gaze direction
• Interaction with props

Micro-examples

• “Full body; feet visible.”
• “Kid reads smaller versus the table.”
• “Eyes on the opened book—not the lens.”
• “Hands naturally on the grips.”

This cuts wrong proportions, puppet poses, or wandering eyes.

2. Constraints: change vs preserve

• Exclude: watermark, rogue lettering, trademarks
• Editing refrain: “Only change X; everything else unchanged.”
• Each pass repeats what stays: identity, geometry, grading, logos
• Surgical tweak: forbid touching saturation/contrast/arrows/labels/camera/context props

3. On-image typography

• Wrap required wording in quotes
• Describe hierarchy, palette, placement
• Spell tricky brands letter-by-letter
• Dense type / multi-font layouts → bump to medium/high

4. Multi-image inputs

Use strict ordering plus spoken roles:

• Image 1: product plate
• Image 2: style reference
• Image 3: environment

Composition sentences

• “Apply Image 2 style to Image 1.”
• “Seat Image 1’s bird atop Image 3’s elephant.”

5. Iterate in layers

Suggested cadence

1. Ship a minimalist baseline
2. Diagnose the miss
3. Adjust one lever (warmth, delete a prop, reopen negative space)
4. Re-affirm constraints before looping

Long prompts hide bugs—restart from clean scaffolds when drift snowballs.

Control is rarely “more words”; it is executable instructions.