Instead of shining our flashlight on a few images at a time, we want to illuminate an arbitrarily-sized area of this latent space of possibility at any given resolution.

We can treat the inputs to our generative image model as levers for directing image generation, each with their own range of possible values. Existing tools let you change one or two of these levers, occasionally allowing for a small number of additional prompts or seeds, but we want to enter our prompt and generate an image for every point within our area of interest. To do this, we will conduct a grid search by providing a list of values for each parameter, and then generating an image for every combination of those values.

When it comes to output image variation, some model parameters act like coarse-grain levers that dramatically change the result, while other parameters act like fine-grain levers for subtle variation. For our purposes, we will use two coarse-grain levers (Prompt and Seed), two fine-grain levers (ETA and Scale), and we will add an additional lever next to the prompt as a list of modifiers that will be appended to the end of the prompt. Now we have 5 parameters: Prompt, Modifier, Seed, ETA, and Scale. Let's start with a sample job composed of a single prompt, a single modifier, 9 seeds, 9 ETAs, and 9 scales:

Here is the set of all possible combinations of these parameter values:

Now let's generate the results and check them out: