https://gitgud.io/gayshit/makesomefuckingporn
https://mega.nz/folder/OppA0QDI#w0_KOY0CyTpPtG6xopKwOw

:)

@animesenpai69 class ArgStore: def __init__(self): # Important, these are the most likely things you will modify self.base_model: str = r"G:\\stable-diffusion\\stable-diffusion-webui\\stable-diffusion-webui\\models\\Stable-diffusion\\nai.ckpt" self.img_folder: str = r"G:\\stable-diffusion\\LoRA Training setup\\img_folder" self.output_folder: str = r"G:\\stable-diffusion\\LoRA Training setup\\output_folder" self.change_output_name: Union[str, None] = None # OPTIONAL, changes how the output files are named self.save_json_folder: Union[str, None] = "G:\\stable-diffusion\\LoRA Training Configs" # OPTIONAL, saves a json folder of your config to whatever location you set here. self.load_json_path: Union[str, None] = None # OPTIONAL, loads a json file partially changes the config to match. things like folder paths do not get modified. self.net_dim: int = 128 # network dimension, 128 seems to work best, change if you want self.alpha: float = 128 # setting it equal to net_dim makes it work equally to how it used to work. # list of schedulers: linear, cosine, cosine_with_restarts, polynomial, constant, constant_with_warmup self.scheduler: str = "cosine_with_restarts" self.warmup_lr_ratio: Union[float, None] = None # OPTIONAL, make sure to set this if you are using constant_with_warmup, None to ignore self.learning_rate: Union[float, None] = 5e-5 # OPTIONAL, None to ignore, seems like people have started not setting this value, so I updated the script to allow for that. self.text_encoder_lr: Union[float, None] = 1.5e-5 # OPTIONAL, None to ignore self.unet_lr: Union[float, None] = 1.5e-4 # OPTIONAL, None to ignore self.batch_size: int = 5 self.num_epochs: int = 5 self.save_at_n_epochs: Union[int, None] = 1 # OPTIONAL, how often to save epochs, None to ignore self.shuffle_captions: bool = True # OPTIONAL, False to ignore self.keep_tokens: Union[int, None] = 1 # OPTIONAL, None to ignore self.max_steps: Union[int, None] = None # OPTIONAL, None to ignore, if you want, you can define an exact step count, the script will do the rest. # These are the second most likely things you will modify self.train_resolution: int = 832 self.min_bucket_resolution: int = 400 self.max_bucket_resolution: int = 1360 self.lora_model_for_resume: Union[str, None] = None # OPTIONAL, takes an input lora to continue training from, not exactly the way it *should* be, but it works, None to ignore self.save_state: bool = False # OPTIONAL, is the intended way to save a training state to use for continuing training, False to ignore self.load_previous_save_state: Union[str, None] = None # OPTIONAL, is the intended way to load a training state to use for continuing training, None to ignore self.training_comment: Union[str, None] = None # OPTIONAL, great way to put in things like activation tokens right into the metadata.

(This is everything, rest of the script was not touched. Trained on my friend's RTX 3090)

@animesenpai69 Less than an hour for a character. The hardest part is collecting data, cropping out unnecessary things, removing watermarks, and pruning the tags. Took me about 6 hours for FDG's v2. 8 hours for St. Louis luxurious wheels (preparing).

@animesenpai69 https://github.com/starik222/BooruDatasetTagManager/releases after you tagged the dataset with wd1.4 tagger, use this to prune the tags from ALL images, and make sure to save at the end. Pruning/deleting the tag means you want the AI to learn about that trait of the thing you're training, for example St. Louis I would prune blue hair, side ponytail, large breasts, pink eyes etc. Try not to prune outfit tags or it might be hard to see different outfits using the LoRA. St. Louis is looking pretty good at the moment, I'm just generating images to show off at the end :D

That "\" symbol is called an escape bracket. On automatic1111's webui, (tag) puts emphasis on the tag, we don't want that when we want something like "friedrich der grosse (azur lane)" because the webui will interpret it as "friedrich der grosse" with weight of 1 and "azur lane" with weight of 1.05, two separated tags when it should be one. "\" tells Stable Diffusion to interpret the parentheses literally.

I told someone else before check the comments

Unfortunately I think my v1 was probably the best, newer doesn't mean better :/

@Trauson Sorry there’s so many notifications on my civitai feed. I can try to do one just for default skin though, but I don’t think it’s worthwhile anyways. Maybe I should just retrain but without pruning.