Nanakusa Nazuna - Yofukashi no uta [Pony, SDXL]

SDXL v1 버전과 다른 데이터셋으로 학습되었습니다.

이것은 다양한 네트워크(LoRA, DoRA, LoKR)에 대한 작은 학습 비교 테스트의 일부이며, 비교 내용은 이 게시물을 참조하세요. 요청 시 LoKR 버전을 업로드할 수 있지만, LoRA/DoRA 버전이 더 나은 것으로 판단했습니다.

Kohya 학습 설정(toml):

bucket_no_upscale = true
bucket_reso_steps = 256
cache_latents = true
cache_latents_to_disk = true
caption_extension = ".txt"
clip_skip = 2
dynamo_backend = "no"
enable_bucket = true
epoch = 20
full_bf16 = true
gradient_accumulation_steps = 1
gradient_checkpointing = true
huber_c = 0.1
huber_schedule = "snr"
learning_rate = 1.0
logging_dir = "E:/work/LoRa_work/logging"
loss_type = "l2"
lr_scheduler = "cosine"
lr_scheduler_args = []
lr_scheduler_num_cycles = 1
lr_scheduler_power = 1
max_bucket_reso = 2048
max_data_loader_n_workers = 0
max_grad_norm = 1
max_timestep = 1000
max_token_length = 225
max_train_epochs = 20
max_train_steps = 153
metadata_author = "motimalu"
min_bucket_reso = 512
min_snr_gamma = 5
mixed_precision = "bf16"
multires_noise_discount = 0.3
multires_noise_iterations = 6
network_alpha = 32
network_args = []
network_dim = 32
network_module = "networks.lora"
no_half_vae = true
noise_offset_type = "Multires"
optimizer_args = []
optimizer_type = "Prodigy"
output_dir = "D:/model_output"
output_name = "pony_nanakusa_nazuna_lora_v1"
pretrained_model_name_or_path =
"C:/workspace/webui_forge_cu121_torch21/webui/models/Stable-diffusion/ponyDiffusionV6XL_v6Start
WithThisOne.safetensors"
prior_loss_weight = 1
resolution = "1024,1024"
sample_every_n_epochs = 10
sample_prompts = "D:/model_output\\prompt.txt"
sample_sampler = "euler_a"
save_every_n_epochs = 20
save_model_as = "safetensors"
save_precision = "bf16"
seed = 339491249
text_encoder_lr = 1
train_batch_size = 8
train_data_dir = "C:/data/nanakusa_nazuna"
unet_lr = 1
xformers = true