musubi wan2.2 训练镜像

• 该镜像已经下载所有模型，并且准备了训练数据集
• musubi 官房地址：https://github.com/kohya-ss/musubi-tuner
• 官方训练文档 https://github.com/kohya-ss/musubi-tuner/blob/main/docs/wan.md

有趣的80后程序员-镜像作者交流群

关键步骤说明

1、打开终端

• 点击jupyterlab点开终端

2、进入musubiTunner主目录

cd /workspace/musubi-tuner/

3、模型下载

• 建议使用huggingface-cli 或 wget 明星进行模型下载
• ditmodel https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/blob/main/split_files/diffusion_models/qwen_image_bf16.safetensors
• textencoder https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/blob/main/split_files/text_encoders/qwen_2.5_vl_7b.safetensors
• vae https://huggingface.co/Qwen/Qwen-Image/blob/main/vae/diffusion_pytorch_model.safetensors

4、数据集前处理

• 数据配置文件格式

[general]
resolution = [1024, 1024]
caption_extension = ".txt"
batch_size = 1
enable_bucket = true
bucket_no_upscale = false

[[datasets]]
image_directory = "dataset/train"
cache_directory = "dataset/cache"
num_repeats = 1

• 图片编码

python src/musubi_tuner/qwen_image_cache_latents.py \
   --dataset_config dataset/data_comfy.toml \
   --vae models/vae.safetensors

• 提示词编码

 python src/musubi_tuner/qwen_image_cache_text_encoder_outputs.py \
    --dataset_config dataset/data_comfy.toml \
    --text_encoder models/qwen_2.5_vl_7b.safetensors \
    --batch_size 1

5、模型训练

accelerate launch --num_cpu_threads_per_process 1 --mixed_precision bf16 src/musubi_tuner/qwen_image_train_network.py \
  --dit models/qwen_image_bf16.safetensors \
  --vae models/vae.safetensors \
  --text_encoder models/qwen_2.5_vl_7b.safetensors \
  --dataset_config dataset/data_comfy.toml \
  --sdpa --mixed_precision bf16 --fp8_base \
  --timestep_sampling shift \
  --weighting_scheme none --discrete_flow_shift 2.2 \
  --optimizer_type adamw8bit --learning_rate 5e-5 --gradient_checkpointing \
  --max_data_loader_n_workers 2 --persistent_data_loader_workers \
  --network_module networks.lora_qwen_image \
  --network_dim 16 \
  --max_train_epochs 300 --save_every_n_epochs 25 --seed 42 \
  --output_dir output --output_name cendy-qwen-image-lora-v1 \
  --blocks_to_swap 20 

6、模型转化

python src/musubi_tuner/convert_lora.py --input output/cendy-qwen-image-lora-v1.safetensors --output output/cendy-qwen-image-lora-comfy-v1.safetensors --target other

注意：在具体使用时需要把文件名路径名按照实际情况修改