GPT Image 2
OpenAI's flagship image model — prompt-faithful generation and reference-based edits, up to 4K with three quality tiers.
Capabilities
| Feature | Support |
|---|---|
| Text-to-Image | Yes |
| Image-to-Image (editing) | Yes |
| Max Resolution | 4K (3840 x 3840) |
| Reference Images | Yes |
| Aspect Ratios | 1:1, 1:3, 2:3, 3:1, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9 |
| Quality Tiers | low, medium, high |
| Size Tiers | 1K, 2K, 4K |
| Negative Prompt | No |
| Inpainting / Mask | Yes |
Prompt-faithfulness
GPT Image 2's strength is following the prompt literally — typography, layout, and explicit visual elements come through more reliably than with diffusion-based image models. Best for cases where the prompt names the elements directly ("a poster with the title 'Summer Festival' centered, in art-deco style on a teal background").
Quality vs. cost
Each size (1K / 2K / 4K) can be rendered at one of three quality tiers (low / medium / high). The credit cost scales with both axes. For iteration, 1K + low is the cheapest combination; switch to higher tiers once you've locked the composition.
Prompting Tips
- Spell out text exactly as you want it rendered. GPT Image 2 reads quoted strings literally.
- Describe layout, not just subject. "A 3-column infographic with section headers" lands more reliably than "an infographic".
- Use reference images for style transfer. Pass the look you want as a reference rather than describing it in prose.
- Use @-mentions when multiple references compete. Type
@to insert a reference inline so the model knows which image plays which role. See Referencing Images in Prompts.
Limitations
- No negative prompt support
- Rate-limited to 3 requests per minute on the upstream API