Controlling DALL-E 3 Output with Seeds

Did you know that you can control DALL-E 3 output not only with a prompt but also with a random seed?

In diffusion models, a "seed" is a number used to generate noise. This noise is a starting point for image generation. Thanks to that when using an identical prompt, different seeds can produce varied results. However, if you maintain the same seed and prompt combination, the outcome remains consistent and reproducible.

And we can use it as an advantage. Retaining the same seed while making subtle alterations to the prompt results in minor changes to the output. This makes DALL-E 3 a powerful and controllable image generator.

How to get a seed number?

Simply ask ChatGPT about that and it will return it to you.

How to use seed?

Ask ChatGPT to generate you an image with the selected seed.

For example, to generate the images below I used modifications of the following prompt:

                Generate me image with seed 2323376857 and prompt """Illustration of a orange long haired feline lounging amidst the organized chaos of a writer's desk. Strewn about are ink bottles, handwritten notes, and a vintage typewriter. The scene is presented in a lofi aesthetic with muted tones."""
              

Providing slight alterations to the prompt while keeping the same seed allowed me to generate different versions of the same image.

It was mentioned today on Reddit: https://lnkd.in/dkYemda7

You can read more about random seeds in diffusion models here: https://lnkd.in/dK_fDg-p

Original post: LinkedIn