Overview
GPT Image 1 is OpenAI's first natively multimodal image model, built on GPT-4o and released in April 2025. Unlike diffusion-based image models, it generates images token-by-token — the same autoregressive approach used in GPT language models. This gives it stronger prompt adherence and more coherent compositions. It went viral shortly after launch for its ability to produce Ghibli-style illustrations.
Use cases
Quick creative exploration and moodboarding. Social media visuals and thumbnails. Product concept sketches. Image editing with natural language instructions — upload up to 5 reference images to guide style or composition.
Inputs
All parameters are passed in the input object of the run request.
| Parameter | Required | Description |
|---|
| prompt | Yes | Text description of the image to generate (1–4000 chars) |
| aspect_ratio | No | Output aspect ratio. Default 1:1. Options: 1:1, 3:2, 2:3 |
| image_urls | No | Up to 5 reference images (max 4 MB each) for image editing |
Prompt tips
Lean into stylistic directions
GPT Image 1 excels at recognizable art styles. Mentioning a specific aesthetic — "Studio Ghibli watercolor," "vintage travel poster," "flat vector illustration" — produces more distinctive results than generic descriptions.
Keep prompts focused
With only 3 aspect ratios and no resolution control, simpler compositions tend to work best. Describe the core subject and mood rather than cramming in layout details.
Limitations
- Only 3 aspect ratios available (1:1, 3:2, 2:3)
- No resolution control — output resolution is fixed by the model
- Complex or small text in images may need iteration for accuracy
- Prompt length capped at 4000 characters
- Maximum 5 reference images per request, 4 MB each
FAQ
How does GPT Image 1 differ from DALL-E?
GPT Image 1 uses an autoregressive architecture (token-by-token generation) rather than diffusion. This generally produces better prompt adherence and more coherent compositions.
What aspect ratios are available?
Three options: 1:1, 3:2, and 2:3. The default is 1:1.
Can I edit existing images?
Yes. Upload up to 5 reference images via image_urls to guide the generation with your own visuals.