gpt-image-prompts

GPT Image Prompt Generator

Generate prompts for OpenAI GPT Image models, with gpt-image-2 as the default target. Treat the model like a production designer who follows a clear creative brief: define the goal, visual deliverable, subject, composition, constraints, and exact text.

Primary reference: OpenAI's GPT Image Generation Models Prompting Guide: https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide

Output

Unless the user asks for another format, return:

1. Ready-to-use prompt in a fenced code block.
2. Recommended API settings: model, size, quality, background, output_format, and notes for n or edit inputs if relevant.
3. Optional refinement prompts: one or two short follow-up edits that improve a likely first pass.

If critical details are missing, ask only the minimum clarifying questions needed. Otherwise, make tasteful assumptions and state them briefly.

Model defaults and constraints

• Default model: gpt-image-2.
• Quality: use low for fast drafts, thumbnails, and broad exploration; use medium for most production prompts; use high for dense text, small labels, infographics, slide/chart outputs, close-up portraits, identity-sensitive edits, and final assets.
• Size: use a concrete size when the output format is known; otherwise use auto.
  • Common sizes: 1024x1024, 1536x1024, 1024x1536, 2048x2048, 2048x1152, 3840x2160, 2160x3840.
  • gpt-image-2 size constraints: both edges multiples of 16px; max edge ≤ 3840px; long:short ratio ≤ 3:1; total pixels from 655,360 to 8,294,400.
  • Outputs above 2560x1440 are more variable/experimental; prefer 2K or below for reliability unless the user needs 4K.
• Background: gpt-image-2 does not support transparent backgrounds. Use background: "opaque" for product extraction and plain-background assets; use downstream background removal if transparency is required.
• Input fidelity: omit input_fidelity for gpt-image-2; it processes inputs at high fidelity by default.
• Output format: default png; choose jpeg for lower latency/photo workflows; choose webp for compressed web assets.
• Variations: use n when the user wants multiple directions, logos, thumbnails, or ad concepts.

Golden rules

1. Write a creative brief, not tag soup. Natural language and labeled sections are better than keyword piles.
2. Lead with purpose. Include what the asset is for: product ad, editorial hero image, UI mockup, classroom handout, pitch-deck slide, marketplace listing, etc.
3. Use a stable prompt order. Background/scene → subject → details → composition → style/medium → constraints → exact text.
4. Name the visual medium explicitly. Use phrases like photorealistic, vector-like logo, flat scientific diagram, mobile app UI mockup, pitch-deck slide, or comic strip.
5. Specify composition. Framing, viewpoint, subject placement, negative space, camera angle, lighting, mood, and color palette all matter.
6. Preserve invariants in edits. Use “change only X; keep everything else the same,” then list the preserved identity/layout/geometry/typography/lighting elements.
7. Quote exact text. Put in-image copy in quotes and specify typography, placement, contrast, and “no extra text.” For hard spellings, spell words letter by letter.
8. Iterate surgically. For an 80%+ good result, write narrow follow-up prompts instead of re-rolling: “make the lighting warmer,” “remove the extra label,” “restore the original background.”

Prompt structure template

Use labeled sections for complex or production prompts:

Create [DELIVERABLE] for [PURPOSE/AUDIENCE].

Scene / background:
[Environment, era, mood, context, world details]

Subject:
[Main subject with concrete attributes, materials, identity cues, pose/action]

Composition:
[Framing, viewpoint, camera angle, placement, negative space, layout hierarchy]

Style / medium:
[Photorealistic / editorial / vector / 3D render / UI / diagram / comic, color palette, lighting]

Text, if any:
Render exactly: "..."
[Typography, size, placement, contrast]

Constraints:
[No extra text, no watermark, no unrelated logos, preserve X, avoid Y]

For simple requests, a polished paragraph is enough. For production work, labeled sections are easier to debug and reuse.

Capability patterns

For guidance specific to the deliverable — photorealism, people and pose, text and infographics, ads, logos, UI mockups, slides and charts, product shots, edits and localization, multi-image compositing, comics — load PATTERNS.md and apply the sections that match the request.