AI infographic generator: draw the layout, add the numbers

An AI image generator can draw an infographic's look, a tall layout with icons, illustrations, and color blocks, but the numbers, labels, and chart proportions it draws aren't guaranteed to be right. Generate the visual frame with empty panels, then place the text and figures yourself in a design tool. On Sume, ChatGPT Image 2.5 takes tall custom sizes up to 1:3, exactly 1280 × 3840 pixels, and today returns up to 4 layouts per call.

Sume facts come from the Image API docs and the per-model lists that GET /v1/images/models serves, read on 2026-09-28.

Why can't an image model make the whole infographic?

Because it draws; it doesn't calculate. POST /v1/images takes a text prompt and optional reference images, with no field for a table of data, so a bar labeled 42% is only as tall as it happens to look, and small labels can come out misspelled. Use the model for the illustrations, the icons, the colors, and the rhythm of the page, and keep every number and word under your own control.

How do I generate an infographic layout?

Describe the frame, not the facts. Send the prompt to POST /v1/images with openai/gpt-image-2.5 and a tall image_size:

• The format: "a tall vertical infographic with a header band and five stacked sections".
• The empty parts: "a blank white panel beside each icon for text, and an empty circle in the third section for a chart".
• The illustrations: one flat icon per section, such as a seedling, a watering can, a sun, a basket, and a jar.
• The style and colors: "flat illustration, rounded shapes, green and cream".
• "No text, no numbers, no letters", so every panel stays empty.
• n for up to 4 layouts per call. Each completed image is billed.

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",
    "prompt": "A tall vertical infographic layout about growing tomatoes: a header band, then five stacked sections, each with one flat icon on the left and a blank white panel on the right. An empty circle in the third section. Flat illustration, rounded shapes, green and cream. No text, no numbers, no letters.",
    "image_size": { "width": 1280, "height": 3840 },
    "n": 4
  }'

Which tall sizes can I generate?

It depends on the model. A model only accepts the ratios its catalog lists, and today a ratio it doesn't list is refused with 400 invalid_request, naming the ratios it takes. These three go taller than 9:16, and all three list 9:16 as well; AI image generation API aspect ratios has every model's full list. The docs don't give pixel sizes for Ideogram V3 or Nano Banana 2 at these ratios, so check the size of what comes back.

Can I turn an article into an infographic?

Yes, but the summary is your job, not the image model's:

• Pull out five or six short points and the numbers behind them. A chat assistant can draft them; check each one against the article.
• Ask the image model for exactly that many sections, in the order of your points, with one matching icon each.
• Keep the points out of the prompt. The prompt only needs the topic, the layout, and the icons; the words go in afterward.

How do I add the numbers and text?

In a design or slide tool, on top of the image:

• Place the image as the background, then type the headings, figures, and sources into the empty panels.
• Build each chart from the real data in a chart tool and drop it into its panel, so every bar and slice matches its value.
• Check every figure against your source before you publish.
• If you still want a drawn headline, AI image with text covers the prompts and the proofreading.

What are the limits?

• No data in and no chart out: any chart in the image is a picture of a chart.
• The file is a flat PNG, JPEG, or WebP. No model lists svg today, so there is no editable vector version.
• Each completed image is billed, and a failed one is not. usage.cost in the response is the USD amount billed.
• High quality (ChatGPT Image 2.5's default) and a large n are two of the slow settings most likely to return 202 with a job to poll instead of the images.
• Results are Sume-hosted, signed URLs, so download the layouts you keep.