BLADE: Block-Sparse Attention Meets Step Distillation for Efficient Video Generation

📖 Paper | 🚀 Homepage | 💾 Models | 📖 中文阅读

BLADE is a data-free framework for efficient video generation. By jointly training an adaptive sparse attention mechanism with a step distillation technique, it achieves a significant acceleration in video generation models. This project combines a block-sparse attention mechanism with step distillation, reducing the number of inference steps from 50 to just 8 while maintaining high-quality generation.

📢 News

• [Aug 2025] 🎉 The code and pre-trained models for BLADE have been released!
• [Aug 2025] 📝 Support for two mainstream video generation models, CogVideoX-5B and WanX-1.3B, is now available.
• [Aug 2025] ⚡ Achieved high-quality video generation in just 8 steps, a significant speedup compared to the 50-step baseline.

✨ Key Features

• 🚀 Efficient Inference: Reduces the number of inference steps from 50 to 8 while preserving generation quality.
• 🎯 Adaptive Sparse Attention: Employs a block-sparse attention mechanism to significantly reduce computational complexity.
• 📈 Step Distillation: Utilizes the Trajectory Distillation Method (TDM), enabling training without the need for video data.
• 🎮 Plug-and-Play: Supports CogVideoX-5B and WanX-1.3B models without requiring modifications to their original architectures.

🛠️ Environment Setup

System Requirements

• Python >= 3.11 (Recommended)
• CUDA >= 11.6 (Recommended)
• GPU Memory >= 24GB (for Inference)
• GPU Memory >= 80GB (for Training)

Installation Steps

1. Clone the repository

   git clone https://github.com/ziplab/BLADE.git
   cd BLADE

2. Install dependencies

   # Install using uv (Recommended)
   uv pip install -r requirements.txt
   
   # Or use pip
   pip install -r requirements.txt

3. Compile the Block-Sparse-Attention library

   git clone https://github.com/mit-han-lab/Block-Sparse-Attention.git
   cd Block-Sparse-Attention
   pip install packaging
   pip install ninja
   python setup.py install
   cd ..

📥 Model Weights Download

Base Model Weights

Please download the following base model weights and place them in the specified directories:

1. CogVideoX-5B Model

   # Download from Hugging Face
   git lfs install
   git clone https://huggingface.co/zai-org/CogVideoX-5b cogvideox/CogVideoX-5b

2. WanX-1.3B Model

   # Download from Hugging Face
   git clone https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers wanx/wan1.3b

Pre-trained BLADE Weights

We provide pre-trained weights for BLADE:

# Download pre-trained weights
git clone https://huggingface.co/GYP666/BLADE pretrained_weights

Weight Directory Structure

Ensure your directory structure for weights is as follows:

BLADE/
├── cogvideox/
│   └── CogVideoX-5b/           # Base model weights for CogVideoX
├── wanx/
│   └── wan1.3b/               # Base model weights for WanX
└── pretrained_weights/         # Pre-trained weights for BLADE
    ├── BLADE_cogvideox_weight/
    └── BLADE_wanx_weight/

🚀 Quick Start - Inference

CogVideoX Inference

cd cogvideox
python train/inference.py \
    --lora_path ../pretrained_weights/cogvideox_checkpoints/your_checkpoint \
    --gpu 0

Argument Descriptions:

• --lora_path: Path to the LoRA weights file.
• --gpu: The ID of the GPU device to use (Default: 0).

Output: The generated videos will be saved in the cogvideox/outputs/inference/ directory.

WanX Inference

cd wanx
python train/inference.py \
    --lora_path ../pretrained_weights/wanx_checkpoints/your_checkpoint \
    --gpu 0

Output: The generated videos will be saved in the wanx/outputs/ directory.

🔧 Training Process

Step 1: Prompt Preprocessing

Before training, you need to preprocess the text prompts to generate embeddings.

CogVideoX Preprocessing

cd utils
python process_prompts_cogvideox.py \
    --input_file your_prompts.txt \
    --output_dir ../cogvideox/prompts \
    --model_path ../cogvideox/CogVideoX-5b \
    --batch_size 32 \
    --save_separate

Argument Descriptions:

• --input_file: A .txt file containing prompts, with one prompt per line.
• --output_dir: The directory to save the output embeddings.
• --model_path: Path to the CogVideoX model.
• --batch_size: The batch size for processing.
• --save_separate: Whether to save each embedding as a separate file.

WanX Preprocessing

cd utils
python process_prompts_wanx.py

This script will automatically process the prompts in utils/all_dimension_aug_wanx.txt and generate the corresponding embeddings.

Step 2: Start Training

CogVideoX Training

cd cogvideox
bash train_tdm_1.sh

Core Training Parameters:

# If not training with 8 GPUs, you must modify CUDA_VISIBLE_DEVICES and the num_processes in config.yaml
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch \
    --config_file train/config.yaml \
    train/train_cogvideo_tdm.py \
    --pretrained_model_name_or_path CogVideoX-5b \        # Path to the base model
    --mixed_precision bf16 \                              # Use mixed-precision for reduced memory usage
    --train_batch_size 5 \                                # Training batch size
    --gradient_accumulation_steps 4 \                     # Number of gradient accumulation steps
    --learning_rate 1e-4 \                                # Learning rate for the student model
    --learning_rate_g 1e-4 \                              
    --learning_rate_fake 5e-4 \                           # Learning rate for the fake model
    --lambda_reg 0.5 \                                    # Regularization weight
    --k_step 8 \                                          # Target number of steps for distillation
    --cfg 3.5 \                                           # Classifier-Free Guidance scale
    --eta 0.9 \                                           # ETA parameter for DDIM
    --use_sparsity true \                                 # Enable sparse attention
    --rank 64 \
    --lora_alpha 64 \                                     # LoRA configuration
    --max_train_steps 300 \                               # Maximum number of training steps
    --checkpointing_steps 15 \                            # Interval for saving checkpoints
    --gradient_checkpointing \                            # Use gradient checkpointing to save memory
    --enable_slicing \
    --enable_tiling                                       # VAE memory optimization

WanX Training

cd wanx
bash train_wanx_tdm.sh

📊 Project Structure

BLADE/
├── README.md                   # Project documentation
├── requirements.txt           # List of Python dependencies
│
├── cogvideox/                 # Code related to CogVideoX
│   ├── CogVideoX-5b/         # Directory for base model weights
│   ├── train/                # Training scripts
│   │   ├── inference.py      # Inference script
│   │   ├── train_cogvideo_tdm.py  # Training script
│   │   ├── train_tdm_1.sh    # Script to launch training
│   │   ├── modify_cogvideo.py # Model modification script
│   │   └── config.yaml       # Training configuration file
│   ├── prompts/              # Preprocessed prompts and embeddings
│   └── outputs/              # Output from training and inference
│
├── wanx/                     # Code related to WanX
│   ├── wan1.3b/             # Directory for base model weights
│   ├── train/               # Training scripts
│   │   ├── inference.py     # Inference script
│   │   ├── train_wanx_tdm.py # Training script
│   │   ├── train_wanx_tdm.sh # Script to launch training
│   │   └── modify_wan.py    # Model modification script
│   ├── prompts/             # Preprocessed prompts and embeddings
│   └── outputs/             # Output from training and inference
│
├── utils/                   # Utility scripts
│   ├── process_prompts_cogvideox.py  # Data preprocessing for CogVideoX
│   ├── process_prompts_wanx.py       # Data preprocessing for WanX
│   └── all_dimension_aug_wanx.txt    # Training prompts for WanX
│
├── Block-Sparse-Attention/  # Sparse attention library
│   ├── setup.py            # Compilation and installation script
│   ├── block_sparse_attn/  # Core library code
│   └── README.md           # Library usage instructions
│
└── ds_config.json          # DeepSpeed configuration file

🤝 Acknowledgements

• FlashAttention, Block-Sparse-Attention: For the foundational work on sparse attention.
• CogVideoX, Wan2.1: For the supported models.
• TDM: For the foundational work on distillation implementation.
• Diffusers: For the invaluable diffusion models library.

📄 Citation

If you use BLADE in your research, please cite our work:

@misc{gu2025bladeblocksparseattentionmeets,
      title={BLADE: Block-Sparse Attention Meets Step Distillation for Efficient Video Generation}, 
      author={Youping Gu and Xiaolong Li and Yuhao Hu and Minqi Chen and Bohan Zhuang},
      year={2025},
      eprint={2508.10774},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.10774}, 
}

📧 Contact

For any questions or suggestions, feel free to:

• Contact Youping Gu at youpgu71@gmail.com.
• Submit an issue on our Github page.