Qwen Image 2.1 sizes (2752x1536, 2400x1792) vs Sume Qwen ratios

The Qwen-Image-2.1 model card gives three example sizes: 2048x2048 for 1:1, 2400x1792 for 4:3 and 2752x1536 for 16:9. On Sume you do not send those pixels to the Image API. You send an aspect_ratio from a list of 13 values for qwen/qwen-image, and Sume picks the size.

The sizes on the model card

The Hugging Face card (fetched 2026-10-07) says Qwen-Image-2.1 supports multiple aspect ratios up to 2048x2048, and gives 2400x1792 for 4:3 and 2752x1536 for 16:9 as examples. Those are the open-weights model sizes. 2752 divided by 1536 is 1.79 and 16:9 is 1.78, so the card sizes are close to the named ratios, not exact.

The 13 Sume values

The catalog lists these for qwen/qwen-image and qwen/qwen-image-max: 1:1, 16:9, 9:16, 4:3, 3:4, 4:5, 5:4, 3:2, 2:3, 21:9, 9:21, 1:2 and 2:1. Any value outside the list returns 400 unsupported_parameter, because Sume rejects a parameter value the model does not list rather than dropping it. The Image API docs describe that rule.

What to do with the difference

Treat the card sizes as a reference for open-weights runs, not as a promise about the hosted row. Sume does not say which Qwen release serves qwen/qwen-image, so measure the returned file size once per ratio and store it next to your job.

• Pick the Sume ratio first, then read the pixel size of the returned file.
• Use 4:5 or 9:16 directly; both are in the list, so no crop is needed.
• If you need an exact pixel size, resize in your own code after the call.