Skip to main content

GPT Image

Provider · OpenAI

What OpenAI's GPT Image models (2.5 and 2) are good at, what changed in 2.5, and how to write prompts they follow.

What is GPT Image

GPT Image is OpenAI's image generation and editing family. The current release is GPT Image 2.5 (September 2026); GPT Image 2 is the release before it, and most prompts in this library were first written for it. Both read a prompt the way a language model does — long, structured instructions land better than keyword soup — and both are among the strongest widely available models for rendering readable text inside an image: posters, packaging, UI mockups, signage, diagrams with labels.

What changed in GPT Image 2.5

  • Two API models instead of one. GPT-Image-2.5 Flare is the everyday default — OpenAI quotes higher quality than GPT Image 2 at about half the latency. GPT-Image-2.5 Sunburst is slower and aimed at work where editing precision matters most: campaign creative, polished product shots.
  • Better reference fidelity. Subjects from a reference photo survive edits more reliably.
  • Steadier multi-turn edits. Follow-up instructions ("now make the jacket red") are followed more reliably across several turns.
  • Lighting and texture. More natural light and richer surfaces out of the box.

Prompts written for GPT Image 2 carry over unchanged. Start with the same prompt; reach for Sunburst only when an edit series keeps drifting.

What it is good at

  • Text in images. Headlines, labels, price tags, book covers, UI copy. Quote the exact string you want.
  • Instruction following. Multi-clause prompts with explicit constraints ("three objects, centered, hard shadow, seamless background") come back close to spec.
  • Product and poster work. Clean studio lighting, layout-aware compositions, brand-style graphics.
  • Editing with references. Feed an image plus an instruction to restyle, extend, or swap elements while keeping the layout.

Why developers reach for it

The prompt is the interface. There are no seeds to hunt, no LoRAs to fetch, and no sampler settings to tune — a well-written paragraph plus an aspect ratio is the whole request. That makes it easy to template: the parts of the prompt worth exposing as controls (subject, palette, camera angle, medium) are exactly the parts you would have written by hand.

How to use the templates below

Each card is a prompt that produced the image shown. Open one to read the full prompt, copy it, and swap the parts you care about. The prompts work on both GPT Image 2 and 2.5, and usually transfer to other models with minor edits — the text-rendering instructions are where the difference shows most.

Supported settings

Capabilities
Text to imageImage editingText in imagesQuality tiersUp to 4K
Resolutions
1k2k4k
Aspect ratios
1:12:33:23:44:39:1616:9

Pitfalls to avoid

  • Long paragraphs dilute instructions. Lead with the subject and the shot, then add style, then add constraints — the model weights the opening of the prompt most heavily.
  • Text inside images works best when you quote it exactly ("a sign reading `OPEN 24H`") and say where it goes. Describing text loosely gets you plausible-looking gibberish.
  • Negative phrasing ("no people") is unreliable. Describe the scene you do want ("an empty street at dawn") instead.
  • Quality tiers change price, not composition. If a render is wrong, fix the prompt before you raise the tier.
  • Reference images steer style and layout, but conflicting instructions in the text usually win. Keep the two aligned.

GPT Image prompt templates

Featured GPT Image prompts

Editable templates are on the way. Until then, these are the cases worth starting from.

Latest GPT Image cases

Browse all GPT Image prompts