Grok Imagine on Sume: 1080x1350 snaps to 2:3, 4:5 is a 400

Ask Sume for x-ai/grok-image with aspect_ratio: "4:5" and you get a 400 invalid_request, because Grok Imagine's catalog list has no 4:5. Send the exact size image_size: "1080x1350" instead and the request is accepted, but it is snapped to the nearest ratio the model has, which is 2:3. Neither path produces a 4:5 image on Grok. If you need Instagram portrait, pick a model whose list contains 4:5.

The Image API page says 4:5 is Instagram portrait (1080×1350), not 4:3. The snapping behavior below comes from running Sume's normalizer on 2026-10-03; per-model lists are published through GET /v1/images/models, as described on the Image API page.

Why is one 400 and the other accepted?

Models whose catalog is a fixed list of ratios treat the two inputs differently. An explicit aspect_ratio must be in the list, so 4:5 fails. An exact pixel size is treated as a hint: Sume reduces it to a ratio and moves it to the closest entry in the model's list, keeping the same family when one exists. For 1080×1350 the ratio is 4:5, Grok has no 4:5 entry, and the closest is 2:3.

Because the request is accepted, nothing in the status code warns you. Check the returned image's dimensions, or read the model's aspect_ratio descriptor before you ask for a size.

Which models keep 4:5?

On the same run, both Ideogram rows kept 4:5 for a 1080×1350 request. GPT Image models list 4:5 too: the catalog comment maps it to a legal custom box of 1024×1280, because 1080 is not a multiple of 16 and so is not a valid GPT custom edge. Nano Banana Pro and Nano Banana 2 also list 4:5; the Image API page notes that exact 1080×1350 on Nano Banana is a documented post-step through the job's target_pixels.

So the rule of thumb is to match the target ratio to the model, not the other way round. For a 4:5 feed asset, choose a model that lists 4:5 and resize or crop to the final pixels afterward.

What should I do in a pipeline?

Fetch the catalog once, read supported_parameters.aspect_ratio.values for the model you plan to use, and fail your own validation when the target ratio is missing. For models with a fixed list, treat image_size as a preference rather than a guarantee, and compare width / height of the output with what you asked for.

If the shape drifts on an edit, pass aspect_ratio: "auto" on models that list it, so the result follows the reference. Grok's list has no auto entry, so send a ratio from its list.

What does this mean for a 1080x1350 deliverable?

If the file must be exactly 1080 by 1350, the order of work is: pick a model whose list has 4:5, request that ratio, then resize or crop the output to the final pixels in your own pipeline. Grok can still be the right model for another job; it just cannot be the source for a 4:5 asset without a crop.

A crop from 2:3 to 4:5 removes about 17 percent of the height (a 2:3 frame is 1.5 times taller than wide, a 4:5 frame is 1.25 times). Put the subject and any text well inside the middle of the frame if you must go that way, and check the crop before you publish.