[sdxl_arguments]
cache_text_encoder_outputs = true
no_half_vae = true
min_timestep = 0
max_timestep = 1000
shuffle_caption = false

[model_arguments]
pretrained_model_name_or_path = "/workspace/pretrained_model/animagine-xl.safetensors"
vae = "/workspace/vae/sdxl_vae.safetensors"

[dataset_arguments]
debug_dataset = false
in_json = "/workspace/LoRA/meta_lat.json"
train_data_dir = "/workspace/LoRA/train_data"
dataset_repeats = 1
keep_tokens = 0
resolution = "1024,1024"
color_aug = false
token_warmup_min = 1
token_warmup_step = 0

[training_arguments]
output_dir = "/workspace/LoRA/outputs"
output_name = "jingliu_sdxl_lora"
save_precision = "fp16"
save_every_n_epochs = 4
train_batch_size = 1
max_token_length = 225
mem_eff_attn = false
sdpa = true
xformers = false
max_train_epochs = 20
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 = "wandb"
log_tracker_name = "sdxl_lora"
logging_dir = "/workspace/LoRA/logs"

[sample_prompt_arguments]
sample_every_n_epochs = 4
sample_sampler = "euler_a"

[saving_arguments]
save_model_as = "safetensors"

[optimizer_arguments]
optimizer_type = "AdaFactor"
learning_rate = 4e-7
max_grad_norm = 1.0
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_weights: network_weight,
network_module = "networks.lora"
network_dim = 64
network_alpha = 32
network_args = []
network_train_unet_only = true

[advanced_training_config]