AI photo edit came back a different size: four causes on Sume

If an edit on Sume came back at a different size than your source, one of four things happened: you sent a shape, you sent a pixel string that was snapped to a ratio, the 1K edit size for that ratio is larger than text-to-image, or the model needs aspect_ratio: "auto" and you left it out. Find yours in the table, then fix the request instead of resizing the result.

The four causes

I checked the Ideogram rows by calling the Sume size resolver directly on 2026-10-05, with no paid call. The GPT row is from the Sume docs.

What stays the same

On Ideogram 4.5 an edit with no aspect_ratio and no image_size is sent as image_size: auto, which keeps the source geometry, and a lone resolution: 2K is ignored. That matches the Ideogram edit reference, which states that output resolution matches the input image exactly (read 2026-10-05). On GPT Image 2.5 and other edits, the Sume docs say to use aspect_ratio: "auto" to match the reference, because omitting it is not the same (Image API docs).

Find it in your logs

Log the request body next to the output size. If the body has aspect_ratio, image_size or size, the size change was asked for. If it has none and the size still changed, check the model: only some models keep the source shape on their own.

• Request has a shape field: expected, not a bug.
• Request is a pixel string: expected snap to the nearest ratio.
• Request has nothing and the model is GPT Image 2.5: add aspect_ratio: auto.
• Request has nothing and the size is still wrong: check the model's aspect_ratio values in GET /v1/images/models and compare with the table.

What not to do

Do not resize a wrong-shaped edit back to the source size. A stretch changes the product, and a crop can lose the part you edited. Fix the request, re-run, and assert the size in code before the file goes anywhere.

A habit that prevents it

Keep one function that builds the edit request, and make it refuse to add a shape field unless the caller sets a flag. That one rule removes the first two causes in the table. Then add the size check after the call so the other two cannot pass silently.

Record the model id with each result. Behaviour differs by model, and a pipeline that switches models for cost reasons can change the output size without any change to your own code.