Quoted By:
So the run_popup.bat works for me, which is nice I guess.
It just asks so many questions and I dunno what to say to 80% of them ¯\_(ツ)_/¯
I'll just go through what it says for fun. Some of you might find the struggles of stoneagechama entertaining.
>Do you want to load a JSON config file?
EZ
>Select your base model
Uh... I usually use HoloCreamsafe but lets go with HLLCounterfeit nuhard-fusion-pruned.
>Select your image folder
wat. "sololive" I guess.
>Select your output folder
...
"sololive" again?!
>Do you want to save a json of your configuration?
Fuck no.
>Which optimizer do you want? The original optimizer was 8bit_adam, or admw for AMD. Lion is a new one that was recently introduced
Lion?!
>Are you training on a SD2 based model?
MAYBE?!
>Are you training on an realistic model?
No...?
>Do you watn tot use regularization images?
nani the fuck
>Do you want to continue from an earlier version?
What earlier version?!
>Do you want to flip all your iamges? It is supposed to reduce biases within your dataset but it can also ruin learning an asymmetrical element.
No because this outfit has a lot of asymmetrical aspects to its design.
>Do you want to change the name of output checkpoints?
... No?
>Do you want to eset a comment that gets put into the metadata?
No comment.
>Do you want to train only one of unet and text encoder?
NANI THE FUCK
>Do you want to save a txt file that contains a list of all tags that you have used in your training data?
Sure I guess.
>Do you want to use caption dropouts?
N A N I T H E F U C K
>Do you want to use noise offset? Noise offset seems to allow for SD to better generate darker or lighter images using this than normal
Sure! Why not?
TL;DR: Here are the results
>lora_train_popup.py: error: ambiguous optoin: --lr_scheduler=cosine_with_restarts could match --lr_shceduler_num_cycles, --lr_scheduler_power
