前言（Preamble）：
Original Author：ゴミャ
X：https://x.com/gomya0_0/status/1938895459747004695
bilibili：https://www.bilibili.com/video/BV1p689zSEzX/
更新日志（Update Log）：
2025.08.08：
收集了最近在 QQ 群里传播的“刨手手”表情包。
Collected the paw pose meme pack that’s been circulating in QQ groups lately.
该 LoRA 用于生成描改参考图。
This LoRA is meant to generate reference images for tracing and redrawing.
主要触发词：
animalization，paw-pose，no-paw。次要触发词：
upper body，= =，blush stickers，:3，{colors}-white skin。Primary trigger word:
animalization,paw-pose,no-paw.Secondary trigger words:
upper body,= =,blush stickers,:3,{color}-white skin.
训练集（Training Set）：
number of images = 80
2_paw-pose = 70
6_no-paw = 10
epochs = 29
[Illustrious XL]Paw Pose (meme) - v1[LoRA].toml：
model_train_type = "sdxl-lora"
pretrained_model_name_or_path = "D:/AI Files/ComfyUI-aki/models/checkpoints/Illustrious XL 1.0.safetensors"
train_data_dir = "D:/AI Files/SD-Trainer/train/Paw Pose (meme)"
prior_loss_weight = 1
resolution = "1024,1024"
enable_bucket = false
min_bucket_reso = 256
max_bucket_reso = 1024
bucket_reso_steps = 64
bucket_no_upscale = true
output_name = "[Illustrious XL]Paw Pose (meme) - v1 EP"
output_dir = "D:/AI Files/SD-Trainer/output/LoRA/[Illustrious XL]Paw Pose (meme) - v1[LoRA]"
save_model_as = "safetensors"
save_precision = "bf16"
save_every_n_epochs = 1
save_state = false
max_train_epochs = 30
train_batch_size = 2
gradient_checkpointing = true
gradient_accumulation_steps = 40
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 = "networks.lora"
network_dim = 8
network_alpha = 8
randomly_choice_prompt = false
positive_prompts = "(masterpiece, extremely aesthetic, best quality, amazing quality), newest, highres, absurdres, animalization, paw-pose4, upper body, facing viewer, 1girl, formidable \\(azur lane\\), long hair, grey hair, hair ribbon, two-tone ribbon, white ribbon, black ribbon, blunt bangs, = =, blush stickers, :3, grey-white skin, white background"
negative_prompts = "(worst aesthetic, worst quality, low quality), low details, bad quality, poorly drawn, (logo, watermark, signature, artist name), multiple views, 4koma, anthro"
sample_width = 1024
sample_height = 1024
sample_cfg = 5.5
sample_seed = 2333
sample_steps = 28
sample_sampler = "euler"
sample_every_n_epochs = 1
log_with = "tensorboard"
log_prefix = "[Illustrious XL]Paw Pose (meme) v1 - "
logging_dir = "./logs/LoRA"
caption_extension = ".txt"
shuffle_caption = true
keep_tokens = 2
max_token_length = 255
seed = 1919810
mixed_precision = "bf16"
xformers = true
lowram = false
cache_latents = true
cache_latents_to_disk = true
persistent_data_loader_workers = true
optimizer_args = [
"decouple=True",
"weight_decay=0.01",
"use_bias_correction=True",
"d_coef=2.0"
]