11 common sizes vs GPT Image 2.5 custom rules: 6 pass, 5 fail

Of eleven common screen and social sizes, six pass the GPT Image 2.5 custom-size rules on Sume: 1280x720, 1920x1088, 2560x1440, 3840x2160, 1024x1280 and 3840x1280. Five fail: 1920x1080 and 1080x1350 and 1200x630 (not multiples of 16), 4096x2160 (edge over 3840) and 512x512 (under the pixel floor).

The Sume Image API docs set four conditions on custom pixel sizes for this model: both edges are multiples of 16, the longest edge is at most 3840, the area is between 655,360 and 8,294,400 pixels, and the aspect ratio is at most 3:1. The table applies all four to each size.

Eleven sizes, four rules

A size is listed as passing only when it breaks none of the rules. When a size fails several, all the reasons are shown.

Reading the failures

1080 is not a multiple of 16, because 1080 divided by 16 is 67.5. That one fact fails 1920x1080 and 1080x1350, the two most requested sizes. The nearest legal height for 1920 wide is 1088, which is 8 pixels taller, and you can trim the 8 rows afterward. 1200x630 fails because 630 is not a multiple of 16 either; 1200x640 is legal.

3840x2160 passes with no margin: it has exactly 8,294,400 pixels, the maximum. The 4096x2160 cinema size fails because 4096 is over the 3840 edge limit and its area of 8,847,360 pixels is also 6.7 percent over the cap, so it breaks two rules at once.

The panoramic 3840x1280 passes because it is exactly 3:1 and holds 4,915,200 pixels. Anything wider, such as 3840x1088, fails the ratio rule even though both edges are legal. Changing 1920x1080 to 1920x1088 is the fix you will use most often.

Snap a size before you send it

This function checks the rules and, for a target width and height, returns the closest legal pair that rounds each edge down to a multiple of 16. It does not fix the pixel cap or the ratio, so check the result again.

def snap(width, height):
    return width // 16 * 16, height // 16 * 16

def legal(width, height):
    pixels = width * height
    return (
        width % 16 == 0 and height % 16 == 0
        and max(width, height) <= 3840
        and 655_360 <= pixels <= 8_294_400
        and max(width, height) <= 3 * min(width, height)
    )

for target in [(1920, 1080), (1080, 1350), (1200, 630)]:
    size = snap(*target)
    print(target, "->", size, "ok" if legal(*size) else "still illegal")

After the render

Rounding edges down changes the shape slightly, so crop or scale to the exact target in your own pipeline. The price of a custom size depends on quality and pixel count; GPT Image 2.5 lists $0.02475 at low quality and 1K up to $0.2225 at high and 4K on Sume, and usage.cost on the response gives the billed amount. The Image API docs have the rules in full.