GPT Image 2.5 infographic prompt: order, text, and quality setting

For a GPT Image 2.5 infographic, write the prompt in a fixed order of scene, subject, key details, then constraints, and set quality to high when the layout carries a lot of text. That is what OpenAI's cookbook recommends for dense layouts. Keep the labels short, because OpenAI's image guide says precise placement in structured layouts can still be hard.

The structure OpenAI recommends

The OpenAI Cookbook prompting guide advises writing prompts in a consistent order: background or scene, subject, key details, constraints. The image prompting page adds that complex requests are easier with labelled sections.

Set quality and size for text-heavy layouts

OpenAI says to set quality to high for dense layouts or heavy in-image text. Sume's docs say an omitted quality on GPT Image 2.5 defaults to high, and that auto reserves the max price, so a fixed high is the predictable choice for a layout job.

For the canvas, OpenAI recommends 1024x1536 for portrait, and allows custom sizes as WIDTHxHEIGHT where both edges are multiples of 16 and neither exceeds 3840. Sume's catalog states the same custom-pixel rules, with a total of 655,360 to 8,294,400 pixels and an aspect ratio of at most 3:1. 1024x1536 passes every rule: both edges are multiples of 16 and the area is 1,572,864 pixels.

Send it on Sume

Use image_size for pixel sizes; Sume's docs say not to put custom pixels on size. The request below is a complete call, and the prompt carries the structure above.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "quality": "high",
    "image_size": "1024x1536",
    "prompt": "Scene: flat white portrait poster, wide margins. Subject: a three-step water filter routine. Details: three numbered boxes top to bottom with the labels \"1 RINSE\", \"2 FILL\", \"3 WAIT\". Constraints: each label exactly once, no other text, no logos."
  }'

Where it will still fall short

Expect to check every label. OpenAI lists text placement and layout-sensitive composition as things the model can still get wrong, and says complex prompts may take up to 2 minutes, so a dense infographic can return as a Sume job rather than a direct image. If the response status is 202, read the result from the job as described in Jobs and results.

For charts with exact numbers, draw the chart in code and use the image model for the surrounding artwork.