stable diffusion simplified for non technical people | How stable diffusion and training works

#stablediffusion #AI #Lora #koyass #sd #regularization #simpletutorial
#character #charactertraining #simple #simplified
Hello, This is a theoretical video about how stable diffusion works, for non technical people, in a simplified manner, it aims to give some simplified explanation of its concepts, how it was trained, its components such as VAE, U-Net, Clip, and differences between training methods and some notes about regularization.
I think it is useful to have some understanding of what happens in the background, so that we know what to expect from our training results, and what to expect when we try to generate new images in Stable diffusion... and hopefully we can do better training based on such understanding.
the video may be over simplifying the concepts, and may also contain some inaccuracies, which i tried to reduce to best of my ability.
Note: iN • stable diffusion simpl... for LoRA/LyCoris please note that Convolution rank is better 4-8 and Alpha is 1 (i mistakenly said convolution rank once again, but the settings in the image is correct)
#stablediffusionsimplified #howitworks

[@MrSongib]

8:00 Hmm im not sure this part of visualisation is correct You could explain it like "This abbreviation mean this and that, and how it works is simmply this way and that way." etc so people can stop asking themself "Why?" Nonetheless it's good walkthrough.

[@schmeat850]

you are amazing, now i understand a few things i didn't before, now i can train my stuff better, i really appreciate it

[@AI-HowTo]

Happy to hear that!

[@lefourbe5596]

damn...great job ! you managed to comfirm ALL of my own research on the topic and even added a bit more ! i'm helping someone making a tutorial to train SD models for two weeks and learned a lot. and now here is your video X). understanding that is key to producte best in class result in Dreambooth but it's too hard to fit in a single tutorial. a course like yours is far beter before hand.

[@AI-HowTo]

Glad to hear, hopefully the video contains something useful for some.

[@Xungames]

Thank you very much for all these explanations !

[@AI-HowTo]

Thank you, hopefully the video contains something useful for some.

[@paveloconnor7759]

Hell yeah brother, thank you 😎

[@AI-HowTo]

you are welcome

[@paveloconnor7759]

I am trying to create a model that will generate photos of interiors for me. I've got thousands of good quality pictures (some are 3000x4000, some smaller) of a particular interior style with and I need SD to be able to generate images in that exact style. My GPU is weak tho, its a 2060 RTX with 6GB of VRAM. What would you recommend for the best possible results? LoRA, LyCoris, Dreambuth? Maybe traiin SDXL model somehow?

[@lefourbe5596]

what you need first is a 12Gb RTX GPU minimum. 6Gb just WON'T do it, you can barely do a small Lora if you're lucky with the best optimizers possible. sorry. best possible result is dreambooth and most likely on SDXL. (even more heavy on VRAM) but more power will be found on SD1.5 training thank to controlnet extension being compatible.

[@paveloconnor7759]

@@lefourbe5596 Shit. How about Colab tho?

[@AI-HowTo]

Dreambooth can produce best results in General but unlikely it will work with your GPU, for Style LyCORIS is more recommended than Standard LoRA while they both work and may (or add ) run on your GPU (but there should be options for low VRAM), currently SDXL is still not yet well supported, i think soon it will be the best model to be used, right now, we stick with what works best ...LoRA is really useful, light weight in general... you may need to look up options for low VRAM too or use google colab which could be much faster for you.... also you might want to consider a stable diffusion check point that works better for interior styles ... for the pictures... either ways you need to scale them down to around 768 or less for your GPU to handle them... with powerful GPU however, using 1024 pictures can produce the best results for style or anything else.... so basically, you need to lots of attempts and repeats to get what you want

[@paveloconnor7759]

@@AI-HowTo already trying the Colab, thanks 😎