Anime Sketch (Style) SDXL & SD1.5

Anime Sketch Style!

I made a more flexible and powerful PonyXL version over here: https://civitai.com/models/362745/modelVersionId=418302

I intentionally used some "sketchy" art so that this Lora generates not professional tier art but rather something a bit more rough.

Check out all the Art styles I’ve made: https://civitai.com/user/CitronLegacy/models?tag=style

Update 12/10/2023 - Added SDXL version.

SDXL doesn't work the exact same way as SD1.5 version so please review my example images for the prompts.

Also generation data is not being populated correctly so I manually edited the generation data for many of the posts -- if something looks wrong thats why lol

Training data is in the "About this version" menu in the side bar.

SD1.5 Version

This is a little unstable but it produces nice results with the right prompts. Please refer to my examples to see how to use it.

I intended to use the 10 Epochs version but Google Colab crashed after 75 minutes. I figured I'd test out the 5 Epoch version while I waited another 120 minutes for the 10 Epochs. I want to refine this to make it more sketchy and more like lineart.

Training Data

Trained on: 218 images

Training Model: SD-v1-5-pruned-noema-fp16

flip_aug: True

Num of Repeats: 10

Unit is Epochs or Steps: Epochs

Number of Epochs or Steps: 10

Training Batch Size: 2

Network Dim: Dim 64 and Alpha 32

Lora Creation Process took: about 75 minutes

[sdxl_arguments]

cache_text_encoder_outputs = false

no_half_vae = true

min_timestep = 0

max_timestep = 1000

shuffle_caption = true

lowram = true

[model_arguments]

pretrained_model_name_or_path = "stabilityai/stable-diffusion-xl-base-1.0"

vae = "/content/vae/sdxl_vae.safetensors"

[dataset_arguments]

debug_dataset = false

in_json = "/content/LoRA/meta_lat.json"

train_data_dir = "/content/drive/MyDrive/lora_training/datasets/Anime_Sketch_SDXL"

dataset_repeats = 2

keep_tokens = 1

resolution = "1024,1024"

color_aug = false

token_warmup_min = 1

token_warmup_step = 0

[training_arguments]

output_dir = "/content/drive/MyDrive/kohya-trainer/output/Anime_Sketch_SDXL"

output_name = "Anime_Sketch_SDXL"

save_precision = "fp16"

save_every_n_epochs = 1

train_batch_size = 4

max_token_length = 225

mem_eff_attn = false

sdpa = true

xformers = false

max_train_epochs = 10

max_data_loader_n_workers = 8

persistent_data_loader_workers = true

gradient_checkpointing = true

gradient_accumulation_steps = 1

mixed_precision = "fp16"

[logging_arguments]

log_with = "tensorboard"

logging_dir = "/content/LoRA/logs"

log_prefix = "Anime_Sketch_SDXL"

[sample_prompt_arguments]

sample_sampler = "euler_a"

[saving_arguments]

save_model_as = "safetensors"

[optimizer_arguments]

optimizer_type = "AdaFactor"

learning_rate = 0.0001

max_grad_norm = 0

optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False",]

lr_scheduler = "constant_with_warmup"

lr_warmup_steps = 100

[additional_network_arguments]

no_metadata = false

network_module = "networks.lora"

network_dim = 32

network_alpha = 16

network_args = []

network_train_unet_only = true