What is the proper way to learn how to use the webui to it's fullest?

[wh33t]

I've tried posting questions on reddit, searching around on the web, watching youtube videos ...

There's lots of responses about what sliders to adjust, things to download, and parameters to change, but such little information about why we should adjust them.

Take for example hypernetworks, wth even is a hypernetwork in this context?

What is the loss number? Is there a number I should be targeting? Do I want more steps with a lower learning rate? Higher steps with a higher learning rate? What the heck even is a learning rate?

What about textual inversion? Number of vectors per token? Wth is a token? Do I want more vectors per token? Or less? Is it situational?

Should I just ask specific questions here? I feel like this is probably the best place to do so. In which case, if you wanna burn off 10 minutes of your life answering me I'd be super obliged.

[grexzen]

Read the wiki.

[wh33t]

Ok, I get it. This is crazy experimental software and not many people actually understand it.

[ghost]

I feel like hypernetworks are guides for model to specific style, at least that's how i use them. It's really important to have them trained on right tags that represent pictures of your choice, then it will guide the model to represent them close to refernece images i feel like. That's my experience at least.

Loss - dw about it. If it's NaN then hypernetwork/embedding is dead.
Obviously you need more steps with lower learning rate and less with higher, but high learning rate make cause loss to go wild and kill your embedding/hn.
Read wiki, most of relevant information is there, as Grexzen suggested.

More tokens - more information can be represented with embeddings i think, but would require bigger dataset to train on, and more steps, obviously.

[aoirusann]

Feeling the same way as you, I realize it's probably because I know too little about AI. Those parameters and methods are not created in this project.
In fact papers on textual inversion and hypernetwork can be easily found: https://arxiv.org/abs/1609.09106, https://textual-inversion.github.io/
However, of course, I don't understand :3

I'm learning about Machine Learning now, starting from cs229, hoping I can catch up with this trend one day.

[scorbon]

Simple stuff you can pick up by watching Youtube videos. I just watched one that explained how to use the scripts for X-Y and all that stuff. For the more complex AI, u-net, layers, latent diffusion, fast.ai has a couple useful courses to go through. First a more general then a more SD specific:
https://course.fast.ai/
https://www.fast.ai/posts/part2-2022.html