Ideogram 4.5 edits keep the source shape; GPT Image 2.5 needs auto

For an edit on ideogram/ideogram-v4.5 through Sume, leave aspect_ratio out: the docs say an edit without it keeps the source image's shape. For other edit and image-to-image calls the same page says to send aspect_ratio: "auto" to match the reference, and adds that omitting the field is not the same as auto. Mixing the two habits is how a portrait photo comes back as a square.

What the docs say

The Image API page states both rules. For Ideogram 4.5: without input_references it generates from text; with them it edits the first image and uses up to 4 more as references (5 total), and an edit without aspect_ratio keeps the source image's shape. In the general options section it says, for edit and image-to-image calls, to prefer aspect_ratio: "auto" and that omitting the field is not the same as auto.

Those are two documented behaviours for two cases, so test your own pair rather than assuming one rule fits both. We have not measured how each model treats an unusual source shape.

Why the shape matters

A product photo for a marketplace often has a required shape. If the edit changes it, you crop or pad afterwards, and the crop can cut the part you edited. Keeping the shape in the request is cheaper than fixing it in a second step.

Size also touches cost. Ideogram 4.5 lists a flat price per quality tier whatever the size, per Fal's page, while GPT Image 2.5 output grows with pixels. An edit that keeps a large source shape is therefore free of a size penalty on Ideogram and not on GPT.

Two requests

Ideogram 4.5 edit, shape from the source:

{
  "model": "ideogram/ideogram-v4.5",
  "prompt": "Change the wall colour to sage green. Keep the furniture and light.",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/room.jpg"}}
  ],
  "quality": "medium"
}

The GPT Image 2.5 version

For GPT Image 2.5, the same edit sets the ratio to auto. If you also use a mask, remember the file rules from OpenAI's image generation guide; the mask must match the image in size.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Change the wall colour to sage green. Keep the furniture and light.",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/room.jpg"}}
  ],
  "aspect_ratio": "auto",
  "quality": "medium"
}

Check the result

Compare the output's width and height with the source after the first call; that is the only proof for your image. If the legacy Image 1.0 URL is still in your code, note that it uses image_urls and its own aspect_ratio notes, and is retiring.