Socialism

I would be interested to see more examples of what the prompt is vs. the image outcome, just to analyze it more.

Well, by experience I would say a lot of it comes down to prompting differently if you can't get the exact result you want. Or adding keywords and keywords to just slightly change the outcome. There's some tricks and keywords you pick up on to get a certain result. I'll try to find a video that showcases all of this because there is a lot that goes into it beyond the commercial LLMs, what they do is take your text prompt and reformat it for the image generator. But you also lose some control and that's how we end up with yellow-filter GPT images (though you can absolutely fix that with some additional prompting).

I would say though the biggest factor is the seed, which determines the original gaussian noise that gets generated. The checkpoint (the image model) then denoises that incrementally over X many steps (all of this is decided by the user). But most models have a sampler and scheduler that is clearly superior and you would not use any other once you find it. The seed however completely changes how the picture looks, because what the checkpoint does is hallucinate patterns in the noise. This post is a good example: https://old.reddit.com/r/StableDiffusion/comments/1p80j9x/the_perfect_combination_for_outstanding_images/. If you click through the gallery quickly, you'll see it immediately.

You can reprompt though, even with the same seed. You could say, instead of a closeup of a wolf (or whatever keyword they used to get that picture of the wolf), "taken from afar". Some models even start to understand "taken from 10 meters away", "macro photography", etc. It really depends on what it's trained on, but you have to think like a descriptor - you're not describing what you want the picture to look like, you're literally describing what's in it. "Person, happy, smiling, in the pouring rain" - you have to add that happy otherwise the model will just "best guess" the expression, or might give them a blank expression.

People in the stabdif community (that subreddit I linked) generally share their prompts, you can explore a little and see how they got the results they did. But it's very dependent on the model itself, and then you can also add LORAs, which introduce purposeful bias. Loras can do a whole bunch of stuff, for example I have one that can produce pixel art. The model generates the picture, and then the LORA intervenes on some level to modify the weights of the neural network and make the output look pixel art. You have loras for everything, and anyone can train them. This is one of the example outputs from the pixel art lora:

edit: I forgot to add, this makes the process very different from other forms of illustrative work. But this is true of painting vs digital painting vs logo creation too, or sculpting vs 3d modeling. Imo image prompting is more akin to a lottery, since it depends on the seed so much you generate a bunch of images (people even generate a whole grid of 9 or more pictures at once and then select the best one), then once you find something good enough you lock the seed in or use img2img, then reprompt over and over again. I'm sure that's even still just entry-level stuff and the 'pros' do a whole bunch of more technical stuff to find exactly what they want.