Wan2.2-Animate-14B-Quant-Compression

Mixed-precision quantization scheme (16/8/4bit mixed quantization) for the Wan2.2-Animate-14B model. Compresses the original 35GB base model to 17GB, balancing inference performance and model size.

Prepare

This project is the quantization compression for the Wan-AI/Wan2.2-Animate-14B model. You need to download the code and pre-trained model in advance:

• Download the code: Wan2.2 repository
• Download the pre-trained model: Wan2.2-Animate-14B on Hugging Face

Quantization Schemes

Quantization Strategy Comparison

Quantization Category	Specific Targets	16/8/4bit Mixed Quantization (17.3G)	16/8bit Mixed Quantization (18.3G)
1. Image Branch Weights	cross_attn.k_img.weight cross_attn.v_img.weight	4bit Quantization (with compensation + packing)	8bit Quantization
2. FFN Weights	blocks.*.ffn.0.weight blocks.*.ffn.2.weight	8bit Quantization Biases: 16bit	8bit Quantization Biases: 16bit
3. KQV Attention Weights	All self_attn and cross_attn weights (except image branch)	cross_attn.q, cross_attn.o → 16bit Others → 8bit	cross_attn.q, cross_attn.o → 16bit Others → 8bit
4. Face-Related Weights	face_encoder.*.weight face_adapter.*.weight	8bit Quantization (except sensitive layers)	8bit Quantization (except sensitive layers)
5. All Other Weights	All weights not in above categories (modulation, norm, embedding, etc.)	Keep 16bit	Keep 16bit

Key Differences Summary

• Image Branch Weights (only 2 layers):
  • 16/8/4bit Scheme: 4bit quantization with statistical compensation and packing
  • 16/8bit Scheme: Standard 8bit quantization
• Other Layers: Identical quantization strategies
• Compression Ratio: 34.5G → 17.3G (50% reduction) vs 34.5G → 18.3G (47% reduction)

Shared Quantization Rules

1. FFN Weights: 8bit for weights, 16bit for biases
2. Shallow Q/O Layers: cross_attn.q and cross_attn.o weights remain 16bit
3. Face Sensitive Layers: Specific sensitive weights remain 16bit
4. All Biases: All bias terms remain 16bit
5. Other Modules: All other components remain 16bit

Comparison Results

Animation

Original Model (34.5G) | Mixed-Precision Quantized Model (17.3G)

Click the image to play the video

Replacement

Original Model (34.5G) | Mixed-Precision Quantized Model (17.3G)

Click the image to play the video

Usage Guide

Step 1: Prepare Original Model

1. Download the original Wan2.2 repository:

git clone https://github.com/Wan-Video/Wan2.2.git

Download the pre-trained Wan2.2-Animate-14B model from Hugging Face.

Step 2: Apply Quantization Patches

Replace the original files with our quantized versions:

# Replace animate.py with our quantized version
cp /animate.py Wan2.2/wan/animate.py

# Replace generate.py with our quantized version
cp /generate.py Wan2.2/generate.py

Step 3: Choose Quantization Scheme

Option 1: 16/8/4bit Mixed Quantization

Run the 16/8/4bit mixed quantization script:

python quantization_with_4bit.py

This will create a quantized/ folder in Wan2.2-Animate-14B/ containing:

quantized/
├── diffusion_pytorch_model-00001-of-00004.safetensors
├── diffusion_pytorch_model-00002-of-00004.safetensors
├── diffusion_pytorch_model-00003-of-00004.safetensors
├── diffusion_pytorch_model-00004-of-00004.safetensors
├── diffusion_pytorch_model.safetensors.index.json
├── advanced_quantization_info.json
└── quantization_info.json

Option 2: 16/8bit Mixed Quantization

Run the 16/8bit mixed quantization script:

python quantization_with_8bit.py

This will also create a quantized/ folder with the quantized model files.

Step 4: Generate Videos with Quantized Model

After quantization, follow the original Wan2.2 instructions to generate videos. The quantized model will be automatically loaded.