EASY Offline Stable Diffusion Model Training, make your own custom LoRA with Kohya, simple process!
Vložit
- čas přidán 17. 06. 2024
- In Part 1, I covered how to install Kohya to create your own LoRA.
This is part 2, describing the process of preparing your images, creating the description files, and the steps to create your own LoRA. I also dig a little into what a Low-Rank Adaptation (LoRA) is, and how it differs from a base model. Hope you find this useful!
Other items discussed in this video:
Low-Rank Adaptation of Large Language Models:
arxiv.org/pdf/2106.09685.pdf
What Every User Should Know About Mixed Precision Training in PyTorch:
pytorch.org/blog/what-every-u...
Base model that I used to train my LoRA (A-Zovya RPG, V3):
civitai.com/models/8124/a-zov... - Věda a technologie
Easily the best tutorial on getting started with LORA while cutting out 90% of the crap most of us don't really care about.
Thanks!
Great and quick explanation of the workflow! Thx a lot!
Thank you for the kind comment!
Yo Thanks so much for your videos, out of all the ones on CZcams only your tutorials have been helpful to me so thank you.
Glad you like them! Thanks for watching!
Very clear explanation ! Love the Render of Mojo it came out great! Sorry to hear he passed on, Who the hell do you think you are making a super clear tutorial , with a screen where all the text is clear and readable!!!! Didn't you know all the text is supposed to be out of focus and fuzzy!!!! I don't impress easy but I am looking forward to more of your tutorials! Maybe you could make one about how to make CZcams videos that have readable screengrabs of text screens!! Subscribed!!!!
Thanks Randy! I appreciate the condolences, he was my friend for over 17 years and a real comfort. Glad the video was helpful (and text was visible!) I'm trying to put some helpful information out there as I discover things.
that's dope
Good tutorial on latest kohya .. most tutorials are waste of time
Heya, I got a question. Can you make a single LORA multifunctional? Like focusing on multiple different similar or related topics? But still different enough that I don't want them confused with each other by the AI.
You can, however the idea of a lora is to "fine tune" to something specific. Once you get the tools set up, it would be better to make separate loras so you can adjust the weights of each topic individually. I think it would be a better result. Thanks for watching!
If I try to make a cloth lora, should I do the same as in the video? Is there any difference?
Very good instruction. So no point of generating 2500 reference images? An older tutorial recommended 1 reference per Training image per repeat with 100 repeats recommended but the current system doesn't even have an assignable repeat slider for the LoRA tab. Also, is there a way to assign keywords specific to the LoRA (that some would say are essential to activating the LoRA) or are keywords obsolete now?
There are so many options but I was trying to put out the most basic steps. I haven't found a way to add keywords, but I believe they should naturally be added with the words in the text file used to describe each image. Thanks for watching!
How did you set up Stable Diffusion on your system? I can't find a link for it and I didn't see you start it up (if it's included as part of Kohya). I'm talking specifically about 14:13.
Hello, I have a vid on installing Stable Diffusion that I posed a little while ago, this should get you going on a Windows system. Thanks for watching.
studio.czcams.com/users/videoEl8JusKOkIY/edit
Hi, thanks for this amazing tutorial. Too bad I got this error "No data found. Please verify arguments (train_data_dir must be the parent of folders with images)" do you have any idea how to solve this?
Hello. I suspect there could be something with the training image folder and sub-folder structure. It's very specific. The folders should be set up like this (it would be more evident if the video I uploaded wasn't so hideously low quality! ;) :
image (
Hi, I would like to ask. I have a character sheet which I want to make a LoRA of. How can I make a LoRA in a way that the style of the character sheet is retained. I do not want to have other style from the model. Is there any neutral model on civit ai which I can use to make a LoRA in the style of the character sheet? Or is there any other way to do it? Thank you
It's not making any models, I get is a json file, but no safesensor file.....:(
Too bad this doesn't work with 1080 TI GPU.
It's extremely frustrating and time consuming. Here's what I can recommend (works on my 1070). I'm going to assume you installed with fp16 and not bfp16, then try these things in steps rather than all at once: 1) change Optimizer “AdamW8bit” to “Lion”; 2) uncheck Xformers and check Gradient checkpointing and Memory efficient attention; 3) uninstall CUDA and install older version CUDA 11.8 dlls; 4) I haven't tried this but some say use Mixed precision 0, save precision fp16