Nano Banana Pro 4:5 is not 1080x1350: crop and resize in Python

Nano Banana Pro will give you a 4:5 picture, but not a 1080x1350 one. The Sume docs say Banana Pro sends aspect_ratio: "4:5" with a native size of about 928x1152 at 1K, and that exact 1080x1350 is a post-step. Resize the result yourself; the script below does it with a crop of a few pixels.

The Sume Image API page states it plainly: 4:5 is Instagram portrait (1080x1350), "not 4:3," and on Nano Banana "WxH maps to the native aspect_ratio" so 1080x1350 becomes 4:5. The docs also name a job-level target_pixels post-step, but they do not show its request syntax, so this post does not guess it.

Why is 928x1152 not exactly 4:5?

Four by five means a ratio of 0.8. 928 divided by 1152 is about 0.8056, a little wider. 1080 divided by 1350 is exactly 0.8. So a straight resize would stretch the picture by under one percent, which is why a centre crop first is the safe route.

For a feed that mixes formats, the same pattern covers other targets: compute the crop from the target ratio, then scale. Pinterest's 2:3 and a 9:16 story follow the same two lines of arithmetic.

How do you request it?

Send the ratio, not the pixels. Pixel strings on size are not served; use aspect_ratio and, on models that accept it, image_size. Google's guide lists 4:5 among the supported ratios for the Gemini 3.1 Flash Lite image model, and says 1K, 2K and 4K tiers exist across Gemini 3 models (Lite is 1K only), so ask for 2K on a model that offers it if you want more pixels to downscale from.

{
  "model": "<banana pro id from GET /v1/images/models>",
  "prompt": "Portrait product photo, centred, soft window light",
  "aspect_ratio": "4:5",
  "resolution": "2K"
}

The resize step

Crop to exactly 4:5 first, then scale to 1080x1350. Install Pillow and point the script at the downloaded file.

from PIL import Image

def to_1080x1350(src, dst):
    im = Image.open(src).convert("RGB")
    w, h = im.size
    new_w = min(w, round(h * 4 / 5))
    new_h = round(new_w * 5 / 4)
    left, top = (w - new_w) // 2, (h - new_h) // 2
    im = im.crop((left, top, left + new_w, top + new_h))
    im.resize((1080, 1350), Image.LANCZOS).save(dst, quality=95)

to_1080x1350("banana.png", "feed-1080x1350.jpg")

Any catches?

If the generated file is smaller than 1080 wide, the resize upscales and softens. Generate at 2K, or upscale with the Image Upscale route first. Keep important content away from the outer edge, because the crop trims it.

If you need the pixels to be exact without any local step, a model that accepts custom pixels is the other route. The Image API docs list GPT, Seedream, FLUX, Qwen and Recraft rows as taking image_size as WIDTHxHEIGHT, with the GPT rules of multiples of 16 and a 3840 maximum edge, so a 1080x1350 request has to be checked against the row's own descriptor.