Onineko Style | 鬼猫

서론（Preamble）：

이 모델은 학습 및 연구 목적에만 사용 가능하며, 상업적 용도로는 절대 사용할 수 없습니다.

This model is for learning and research purposes only.

It is strictly prohibited for commercial use.

업데이트 로그（Update Log）：

2025.05.20：

• 훈련 데이터셋을 업데이트하고 최적화했으며, 최신 데이터는 2025년 5월 기준입니다.

• The training set has been updated and optimized, with the latest data being current as of May 2025.

• Illustrious XL 및 NoobAI XL V-Pred에서 각각 훈련된 두 개의 LoKr 모델을 공개했습니다.

• Two LoKr models have been released, trained on Illustrious XL and NoobAI XL V-Pred respectively.

• 트리거 단어: onineko。

• Trigger Words: onineko.

2023.07.17：

• LoHa 모델에 일부 과적합 현상이 있습니다. 권장 가중치: 0.6 ~ 0.8。

• This model has some overfitting. Recommended weights: 0.6 ~ 0.8.

• 1024px 이미지를 생성할 때는 "고화질 복원" 기능을 사용하는 것을 권장합니다。

• It is recommended to use "Hires. fix" to generate 1024px images.

• 트리거 단어: style_onineko, quality_best。

• Trigger Words: style_onineko, quality_best.

훈련 데이터셋（Training Set）：

이미지 수 = 124

• 1_좋은 품질 = 21

• 3_최고 품질 = 15

• 4_치비 = 9

• 5_마스터피스 = 79

에포크 수 = 15

[NoobAI XL]onineko-v1[LoKr][V-Pred].toml：

model_train_type = "sdxl-lora"
pretrained_model_name_or_path = "D:/AIGC/ComfyUI-AKI/models/checkpoints/NoobAI-XL (NAI-XL) V-Pred-1.0-Version.safetensors"
train_data_dir = "D:/AIGC/lora-scripts/train/Onineko XL"
prior_loss_weight = 1
resolution = "1024,1024"
enable_bucket = true
min_bucket_reso = 512
max_bucket_reso = 2048
bucket_reso_steps = 64
bucket_no_upscale = true
output_name = "[NoobAI XL]onineko-v1 EP"
output_dir = "D:/AIGC/lora-scripts/output/LoKr/[NoobAI XL]onineko-v1[LoKr][V-Pred]"
save_model_as = "safetensors"
save_precision = "bf16"
save_every_n_epochs = 1
save_state = false
max_train_epochs = 20
train_batch_size = 2
gradient_checkpointing = true
gradient_accumulation_steps = 62
network_train_unet_only = true
network_train_text_encoder_only = false
learning_rate = 1
unet_lr = 1
text_encoder_lr = 1
lr_scheduler = "constant"
lr_warmup_steps = 0
optimizer_type = "Prodigy"
network_module = "lycoris.kohya"
network_dim = 114514
network_alpha = 1
randomly_choice_prompt = false
positive_prompts = "(masterpiece, extremely aesthetic, best quality, amazing quality), newest, highres, absurdres, onineko, sitting, yokozuwari, arm support, depth of field, looking at viewer, 1girl, short hair, grey hair, panda ears, hair between eyes, grey eyes, light blush, expressionless, closed mouth, china dress, black dress, medium breasts, bare arms, hand on own thigh, barefoot, nature, on grass, flower, butterfly, dappled sunlight, blurry background"
negative_prompts = "(worst aesthetic, worst quality, low quality), low details, bad quality, poorly drawn, (logo, watermark, signature, artist name), (bad hands, mutated hands), extra digits, fewer digits, bad feet, multiple views, 4koma, furry, anthro"
sample_width = 1024
sample_height = 1536
sample_cfg = 5.5
sample_seed = 2333
sample_steps = 28
sample_sampler = "euler"
sample_every_n_epochs = 1
log_with = "tensorboard"
log_prefix = "[NoobAI XL]onineko-v1-"
logging_dir = "./logs/LoKr"
caption_extension = ".txt"
shuffle_caption = true
keep_tokens = 1
max_token_length = 255
flip_aug = true
seed = 1919810
mixed_precision = "bf16"
xformers = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
network_args = [
  "conv_dim=4",
  "conv_alpha=1",
  "dropout=0",
  "algo=lokr",
  "factor=4"
]
optimizer_args = [
  "decouple=True",
  "weight_decay=0.01",
  "use_bias_correction=True",
  "d_coef=2.0"
]
v_parameterization = true
zero_terminal_snr = true
scale_v_pred_loss_like_noise_pred = true
debiased_estimation_loss = true