sixgunzloaded wrote:Oh wow, it really takes quite a dramatic leap at 75% and 99% doesn't it? The rolling dice analogy clears it up for me. I can see where one might get really distracted playing around with those two percentages. Just seeing what it comes up with.
Even just running at the same strength produces different results. There's a certain gambling satisfaction from rolling good outputs. Though, it's easy to forget that you as the creator can manipulate the odds in your favour with the painting tools and changing the parameters, such as putting more emphasis on one specific tag and decreasing the priority of another.
When you start off with a text-only piece, how many descriptors does it take and what sort do you use to generate an image like "base.png"? Did it come up with that on the first try?
That one was literally the first try, so that was winning the pot right there. Often I might roll dozens of generations to get the right output. However, with the level of experimenting I've done, I've got my formula down to have more predictable and workable results.
For the descriptors: let's say it's "a lot". With the prompts, you can use prose or tags. For example, let's make the following:
"Blonde jungle explorer with long hair, wearing a green shirt, is sinking in quicksand in the jungle"
jungle3.png
jungle2.png
jungle1.png
This is where many new creators are stuck: it doesn't give us what we 'want", just what the AI interprets. Also keep in mind that the app I use (NovelAI) is trained on anime models, so it more closely defaults to anime, whereas other AI apps will use realistic mashups.
This is where knowledge of which prompts are effective distinguishes new creators from the good ones. Firstly, I don't use prose - the AI doesn't treat prose any better than tags, and prose can be hard to understand for AI. What exactly is a "jungle explorer", or "quicksand", and so on? So I lay out the features of the piece: long straight blonde hair, detailed hair, dynamic hair, blue eyes, green shirt, large breasts, sinking in quicksand, deep mud, muddy, oozing mud, ripples, jungle background, etc.
jungle4.png
We already have a very appealing output (three rerolls). As I said, NovelAI is an anime-based model, so the style is quite distinct. However, the style I use is photorealistic, so I add that tag in, run the prompt again, and after a few re-rolls to get something close to what I want:
jungle5.png
We're getting much closer to my ideal QS illustration. Part of this is putting the right prompts - these give the AI direction on what it should include, and I can add weight to individual tags. For example, if I add "navel", the character probably won't be chest-deep since the instructions must include a navel, so you manipulate particular things to show or not show them. Same with the negative prompts - by including things like "legs" and "hips" in the negative prompts, the AI is less likely to include them, hence it "sinks" the character deeper in most outputs.
The full list out prompts is what AI creators tend to protect a bit more. When creators share work, there's always a demand to find out what the exact prompts and settings are. But like a secret recipe, creators often give vague descriptions but might not want to share what makes theirs special. The consistent quality of outputs is the result of many failed experiments.