AnimaLoraToolkit

git clone https://github.com/Moeblack/AnimaLoraToolkit.git
cd AnimaLoraToolkit

# 创建虚拟环境
python -m venv .venv

# Windows
.\.venv\Scripts\activate

# 安装依赖
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt

models/
├── transformers/
│   └── anima-preview.safetensors      # Anima 主模型
├── vae/
│   └── qwen_image_vae.safetensors     # VAE 解码器
└── text_encoders/
    ├── config.json                     # 已包含（小文件）
    ├── tokenizer_config.json           # 已包含（小文件）
    ├── merges.txt                      # 已包含（小文件）
    ├── tokenizer.json                  # 需下载（小文件，可用 download_tokenizers.py）
    ├── vocab.json                      # 需下载（小文件，可用 download_tokenizers.py）
    ├── special_tokens_map.json         # 需下载（小文件，可用 download_tokenizers.py）
    └── model.safetensors               # 需下载（大文件：Qwen3-0.6B 权重）

python download_tokenizers.py

dataset/
├── image001.jpg
├── image001.txt    # Danbooru 风格标签
└── ...

dataset/
├── image001.jpg
├── image001.json   # 结构化标签
└── ...

{
  "quality": "newest, safe",
  "count": "1girl",
  "character": "hatsune miku",
  "series": "vocaloid",
  "artist": "@wlop",
  "appearance": ["long hair", "blue hair", "twintails", "blue eyes"],
  "tags": ["singing", "microphone", "dynamic pose"],
  "environment": ["concert stage", "spotlight", "crowd"],
  "nl": "Miku performs energetically on stage."
}

cp config/train_template.yaml config/my_training.yaml
# 编辑 my_training.yaml

python anima_train.py --config ./config/my_training.yaml

python anima_train.py --config ./config/my_training.yaml --lr 5e-5 --epochs 20

# 模型路径
transformer_path: "models/transformers/anima-preview.safetensors"
vae_path: "models/vae/qwen_image_vae.safetensors"
text_encoder_path: "models/text_encoders"
t5_tokenizer_path: "models/t5_tokenizer"

# 数据集
data_dir: "./dataset"
resolution: 1024
repeats: 10              # 数据重复次数

# Caption 处理
shuffle_caption: true    # 打乱标签顺序
keep_tokens: 0           # 保护前 N 个标签不打乱
prefer_json: false       # 优先使用 JSON 标签文件
flip_augment: false      # 水平翻转增强
tag_dropout: 0.0         # 标签随机丢弃概率
cache_latents: true      # 缓存 VAE latent

lora_type: "lokr"        # lora 或 lokr
lora_rank: 32            # LoRA rank（建议 16-64）
lora_alpha: 32           # 通常与 rank 相同
lokr_factor: 8           # LoKr 专用参数

epochs: 50
max_steps: 0             # 0 = 不限制
batch_size: 1
grad_accum: 4            # 有效 batch = batch_size × grad_accum
learning_rate: 1e-4
mixed_precision: "bf16"  # fp16, bf16, 或 no
grad_checkpoint: true    # 梯度检查点（省显存）
xformers: false          # Windows 5090 用 SDPA 更好
num_workers: 0           # Windows 必须为 0

output_dir: "./output"
output_name: "my_lora"

# === 保存配置（重要！） ===
save_every: 0              # 每 N epoch 保存 LoRA (0=禁用)
save_every_steps: 500      # 每 N step 保存 LoRA (推荐)
save_state_every: 1000     # 每 N step 保存完整训练状态（可断点续训）

# === 继续训练 ===
resume_lora: ""            # 从已有 LoRA 继续训练
resume_state: ""           # 从训练状态恢复（断点续训）

seed: 42

sample_every: 5          # 每 N 个 epoch 采样
sample_steps: 0          # 或每 N step 采样
sample_infer_steps: 25
sample_cfg_scale: 4.0
sample_sampler_name: "er_sde"
sample_scheduler: "simple"

# 多提示词轮换
sample_prompts:
  - "newest, safe, 1girl, ..."
  - "newest, safe, 1boy, ..."

# 在配置文件中指定
resume_lora: "./output/my_lora_step1000.safetensors"

python anima_train.py --config config.yaml --resume-lora ./output/my_lora_step1000.safetensors

# 在配置文件中指定
resume_state: "./output/cosmic_kaguya/training_state_step1000.pt"

python anima_train.py --config config.yaml --resume-state ./output/training_state_step1000.pt

检测到 Ctrl+C，正在保存训练状态...
已保存！下次使用 --resume-state "xxx/training_state_step1234.pt" 继续训练

# 长时间训练推荐配置
save_every_steps: 500      # 每 500 step 保存 LoRA（方便选择最佳版本）
save_state_every: 2000     # 每 2000 step 保存训练状态（断点恢复用）

# 转换单个文件
python convert_lokr_for_comfyui.py ./my_lokr.safetensors

# 指定输出路径
python convert_lokr_for_comfyui.py ./my_lokr.safetensors --output ./converted.safetensors