DALL-E 3 Image doesn't fill entire frame

Sometimes when I prompt GPT to make an image, it doesnt fill the entire frame. For example, I sent the following prompt to the dall-e-3 API:

        generated_image = client.images.generate(
            model="dall-e-3",
            prompt="Make a youtube thumbnail about "+thumbnail_prompt+". Make it a painting, dont put words on it, and use strong colours.",
            n=1,
            size="1792x1024"
        )

and it returned the attached image. How could I edit my prompt to make it always fill the entire frame (and not have those grey bars on the side)? Does anyone have a good prompt fix for this?

NOTE: The attached image is 1920x1080 because I did that after generating the image (which came out as 1792x1024)

1 Like

DALL·E 3 was trained to generate 1024x1024, 1024x1792 or 1792x1024 images. To create images more quickly or with lower quality and cost, you can use the new quality parameter. You can still generate 512x512 or 256x256 sized images if you select the dall-e-2 model.

This is just the image size DALL-E 3 generates.

The issue is about the underutilized space within the image that is generated.

Here for example is a wide image request, where the yellow image itself is exactly 1000 pixels wide, of 1792.

I suspect this is somewhat related to undisclosed techniques for making a wide image, such as using fill completion techniques on an initial square generation. It has had some acknowledgement, as much as the tall image creation aspect ratio requests sometimes being (remaining?) rotated were acknowledged as an issue, but is a symptom that still persists in similar probability - randomly coming about from identical inputs.

Fair enough.

It wasn’t immediately clear to me what the issue was the OP was referring to, it initially read to me as though they wanted the model to generate a 1920x1080 image.

Create a skinny cat image