Author-Kun's First Foray Into Training AI
tl;dr it did not go so well
Where should I begin? Back in December, I was messing around casually with AI art so that I could have some cute Altergeist waifu art in Master Duel. I've been prompting and practicing in Stable Diffusion for some time, AI art is really fun.
So yesterday, I finally got around to trying to create my own models. Suffice to say... things did not go so well.
How Do I Create Model?
What I really wanted to do was to create something which was custom fit to my needs; i.e a model that can spit out /Conspiracy/ Girls characters faster than my opening up COM3D2. The problem is: where do I start?
For starters, there's a huge selection of youtube tutorials, but they don't all start at the same point. Some of them start from the Stable Diffusion installation, whereas I was somewhere in the "I know what a model is, but I want to make my own" intermediate range now.
In the end, I settled on AItrepreneur's selection of AI tutorials, starting with a promising extension in Stable Diffusion called Dreambooth.
A Dreambooth Adventure (Not)
SPOILERS: Read the requirements, Author-kun!
I initially thought that my laptop could handle things well, since prompting is no big problem for me, but boy was I wrong. Somehow, my computer registers itself as a 16RAM laptop, but Dreambooth doesn't! So it is constantly starved for RAM and doesn't load at all.
This was after hours of going through the Dreambooth extension and rewinding the tutorial to fit the parameters correctly, only to be met with... this does not work on your machine. I feel kind of pickled by this, but it's alright. My laptop is an old steed, it's been with me so long that the batteries have to be removed for potential fire hazards now. I can acknowledge that it is a little outdated, though it runs Elden Ring and other things well enough.
Textual Inversions
Alright, since Dreambooth doesn't work, we can safely DELETE it from the extensions tab.
but I still want a custom model to play with. So how do we do that? To do that, I had to look to another kind of training method: textual inversions. It's kind of similar to Dreambooth in function that you call out a style/character with a keyword and the best part is that textual inversions only take up a small amount of hard drive space.
Alright, I'm down. Once again, I went to AItrepreneur's handy tutorials, and the results were... okay. More like unusable amounts of okay, but that's mostly because I am training it on the wrong model.
PROTIP: Use the Stable Diffusion 1.5 model
It also automatically shut off my computer's power after about 800 steps, so that's about the limit old Leonard can take.
The Last Scene
Yesterday's training ended up in mostly failures, I don't have much to show for it. Prompting still works, it's just highly unspecific. Perhaps I should work on being able to prompt better first?
Written 230123
