Fidelity & sizing guardrails
Default new production flows to gpt-image-2 when your stack exposes it—especially for typography-heavy comps, composites, sensitive likeness edits, and shots where rerun cost hurts more than incremental spend. Larger canvases trending past ~2560×1440 can oscillate stylistically—treat oversized outputs as exploratory until regressions stabilize.
Prompt scaffolding
- Establish setting / background plates before hero subjects—this anchors diffusion attention.
- Enumerate materials & optics (paper grain, satin weave, sapphire edge highlights) sparingly rather than spraying adjectives randomly.
- Declare deliverable modality (UI mock sheet, billboard, infographic) so polishing heuristics align.
- Freeze invariants explicitly when iterating edits: typography blocks, silhouette geometry, branded glyph locks.
- Quote literal in-image strings with contrast + placement cues; escalate quality settings when density rises.
Generation archetypes you will reuse
- Infographics & classroom diagrams — specify audience reading level, maximum label count, arrow grammar, and forbidden embellishments (mascots, comic halftone, etc.) when clarity matters.
- Photoreal lifestyle & product — say photorealistic outright; pair with believable imperfections (wear, dust, edge flecks) to avoid wax-skin failure modes.
- UI mockups — describe shipped affordances, grid rhythm, and device frame generically; ban “fantasy HUD” language if you need App Store plausibility.
- Localized image translation — instruct “translate only copy; preserve composition, weight, line breaks except unavoidable reflow.”
- Marketing comps & sequential art — treat prompts like creative briefs: audience, cultural tone, palette bias, required tagline string once.
Edit / composite hygiene
For image → image flows, pair every destructive instruction with a preservation manifest (“do not alter camera yaw, skin tone mapping, label kerning”). When compositing multiple references, index them: “Image 1 = environment plate; Image 2 = subject insert; lock lighting direction from Image 1.” Iterate with single-axis tweaks after a strong base pass.
Legal & brand safety
Ask for original marks, refuse counterfeit trademarks, and keep sensitive likeness work human-in-the-loop. Model knowledge can infer real historical contexts—verify facts before publishing documentary-styled frames.