Demand Gen image assets in every ratio from one reference

Run one loop over four sizes and pass the same product photo as an input_references entry each time. Google's Demand Gen spec, read 2026-09-30, lists 1:1, 1.91:1, 4:5 and 9:16 images, and its September 24 2026 update adds one-click image ads on 9:16 full-bleed in Shorts.

Which size do I request for each ratio?

Two of Google's four recommended sizes cannot be sent as written, because custom pixels must be multiples of 16. The table shows the nearest sendable neighbor.

What does the loop look like?

Each request gets its own Idempotency-Key. Responses are jobs; the default mode is async, so store the returned job id and poll it.

const sizes = ["1200x1200", "960x1200", "1200x640", "1088x1920"];
const ref = "https://cdn.example.com/product.png";

for (const image_size of sizes) {
  const res = await fetch("https://api.sume.com/v1/images", {
    method: "POST",
    headers: {
      Authorization: "Bearer " + process.env.SUME_API_KEY,
      "Content-Type": "application/json",
      "Idempotency-Key": "dg-set-" + image_size + "-v1",
    },
    body: JSON.stringify({
      model: "openai/gpt-image-2.5",
      prompt: "Product on a clean background, same product as the reference",
      input_references: [{ type: "image_url", image_url: { url: ref } }],
      image_size,
      output_format: "png",
    }),
  });
  console.log(image_size, res.status, await res.json());
}

Why not keep aspect_ratio auto on edits?

The docs advise aspect_ratio: "auto" for edits so the output matches the reference, and say omitting the field is not the same as auto. Here you want fixed shapes, so you set image_size instead, which wins over aspect_ratio. The cost is that a very different reference shape may need a prompt that says how to recompose.

What should I check before uploading?

Open each result and check composition at every ratio; a centered product in 1:1 can sit too high in 9:16. Check Google's own file rules from the snapshot (.jpg or .png, 5 MB for images). For larger runs, see bulk image generation.