Seedream 5 Lite 4:5 portrait: render 1920x2400, resize to 1080x1350

The short answer

Yes, Seedream 5 Lite can make a 4:5 portrait on Sume. Ask for aspect_ratio: "4:5" (or a 1920 x 2400 image_size) and resize the result to 1080 x 1350. The scale is exact: 1080 / 1920 = 0.5625 and 1350 / 2400 = 0.5625, so nothing is cropped and the shape is the same. The Sume price is $0.04375 per image, so 100 portraits cost $4.375, or $4.38 rounded.

Why 1920x2400

The Image API docs say 4:5 is Instagram portrait at 1080 x 1350, not 4:3. They also say that exact 1080 x 1350 can come from a documented post-step through job target_pixels, and that Image 1.0 documents custom pixels for Seedream among other models. Sume's contract source keeps a Seedream custom-pixel box for 4:5 at 1920 x 2400 (and 2400 x 1920 for 5:4), with each edge in the 1920 to 4096 range.

A render at that size is larger than the Instagram target, so you resize down, which keeps detail. That is better than asking for 1080 x 1350 directly and getting something near it.

The numbers

The table gives the render, the resize, and the cost for the Seedream rows as read on 2026-10-09. The shared price list shows seedream-5-lite at $0.04375 per image.

The resize step in code

In Pillow the resize is one call, img.resize((1080, 1350), Image.LANCZOS), after you download the result from the Sume-hosted URL. Because 1920 x 2400 and 1080 x 1350 share the 4:5 shape, there is no crop and no letterbox. Save as JPEG at a quality near 90 for a feed upload, or as PNG if you need lossless files.

Keep the original 1920 x 2400 render. It is useful for a print crop, a story-sized crop, or a second pass later, and it costs nothing extra once you have paid for the image.

Checking the result

After the download, open the file and confirm the size with img.size. It should read (1920, 2400) before the resize and (1080, 1350) after. A mismatch means the model returned a different ratio, and a crop would then be needed instead of a plain resize.

Watch the file weight too. A 1080 x 1350 JPEG at quality 90 is usually well under a megabyte for a simple scene, but busy textures grow quickly. If an upload target has a size cap, lower the quality in steps of 5 until the file fits.

Check the model's own supported_parameters with GET /v1/images/models/{model_id}/endpoints before you rely on a size, because the docs say a parameter a model does not list returns 400 unsupported_parameter. Do the resize with a high-quality filter (Lanczos in Pillow), and export as JPEG or PNG for the feed.

If you would rather not resize at all, GPT Image 2.5 accepts a legal 1024 x 1280 4:5 box, but it bills on its own quality tiers and not at a flat $0.04375.