🎬 Wan2.2 Distilled LoRA Models#
LoRA weights extracted from Wan2.2 distilled models - Flexible deployment with excellent generation quality
Image: img_lightx2v
🤗 HuggingFace GitHub License
🌟 What's Special?#
|
⚡ Flexible Deployment#
Base Model + LoRA : Can be combined with base modelsOffline Merging : Pre-merge LoRA into modelsOnline Loading : Dynamically load LoRA during inferenceMultiple Frameworks : Supports LightX2V and ComfyUI |
🎯 Dual Noise Control#
High Noise : More creative, diverse outputsLow Noise : More faithful to input, stable outputsRank 64 LoRA, compact size |
💾 Storage Efficient#
Small LoRA Size : Significantly smaller than full modelsFlexible Combination : Can be combined with quantizationEasy Sharing : Convenient for model weight distribution |
🚀 4-Step Inference#
Ultra-Fast Generation : Generate high-quality videos in just 4 stepsDistillation Acceleration : Inherits advantages of distilled modelsQuality Assurance : Maintains excellent generation quality📦 LoRA Model Catalog#
🎥 Available LoRA Models#
Task Type Noise Level Model File Rank Purpose I2V High Noise wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors64 More creative image-to-video I2V Low Noise wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors64 More stable image-to-video
💡 Note :
xxx in filenames represents version number or timestamp, please check HuggingFace repository for the latest versionThese LoRAs must be used with Wan2.2 base models 🚀 Usage#
Prerequisites#
Base Model : You need to prepare Wan2.2 I2V base model (original model without distillation)
Download base model (choose one):
Method 1: From LightX2V Official Repository (Recommended)
# Download high noise base model
huggingface-cli download lightx2v/Wan2.2-Official-Models \
wan2.2_i2v_A14b_high_noise_lightx2v.safetensors \
--local-dir ./models/Wan2.2-Official-Models
# Download low noise base model
huggingface-cli download lightx2v/Wan2.2-Official-Models \
wan2.2_i2v_A14b_low_noise_lightx2v.safetensors \
--local-dir ./models/Wan2.2-Official-ModelsMethod 2: From Wan-AI Official Repository
huggingface-cli download Wan-AI/Wan2.2-I2V-A14B \
--local-dir ./models/Wan2.2-I2V-A14B💡 Note : lightx2v/Wan2.2-Official-Models provides separate high noise and low noise base models, download as needed
Method 1: LightX2V - Offline LoRA Merging (Recommended ⭐)#
Offline LoRA merging provides best performance and supports quantization simultaneously.
1.1 Download LoRA Models#
# Download both LoRAs (high noise and low noise)
# Note: xxx represents version number, please check HuggingFace for actual filename
huggingface-cli download lightx2v/Wan2.2-Distill-Loras \
wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--local-dir ./loras/1.2 Merge LoRA (Basic Merging)#
Merge LoRA:
cd LightX2V/tools/convert
# For directory-based base model: --source /path/to/Wan2.2-I2V-A14B/high_noise_model/
python converter.py \
--source ./models/Wan2.2-Official-Models/wan2.2_i2v_A14b_high_noise_lightx2v.safetensors \
--output /path/to/output/ \
--output_ext .safetensors \
--output_name wan2.2_i2v_A14b_high_noise_lightx2v_4step \
--model_type wan_dit \
--lora_path /path/to/loras/wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--lora_strength 1.0 \
--single_file
# For directory-based base model: --source /path/to/Wan2.2-I2V-A14B/low_noise_model/
python converter.py \
--source ./models/Wan2.2-Official-Models/wan2.2_i2v_A14b_low_noise_lightx2v.safetensors \
--output /path/to/output/ \
--output_ext .safetensors \
--output_name wan2.2_i2v_A14b_low_noise_lightx2v_4step \
--model_type wan_dit \
--lora_path /path/to/loras/wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--lora_strength 1.0 \
--single_file1.3 Merge LoRA + Quantization (Recommended)#
Merge LoRA + FP8 Quantization:
cd LightX2V/tools/convert
# For directory-based base model: --source /path/to/Wan2.2-I2V-A14B/high_noise_model/
python converter.py \
--source ./models/Wan2.2-Official-Models/wan2.2_i2v_A14b_high_noise_lightx2v.safetensors \
--output /path/to/output/ \
--output_ext .safetensors \
--output_name wan2.2_i2v_A14b_high_noise_scaled_fp8_e4m3_lightx2v_4step \
--model_type wan_dit \
--lora_path /path/to/loras/wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--lora_strength 1.0 \
--quantized \
--linear_dtype torch.float8_e4m3fn \
--non_linear_dtype torch.bfloat16 \
--single_file
# For directory-based base model: --source /path/to/Wan2.2-I2V-A14B/low_noise_model/
python converter.py \
--source ./models/Wan2.2-Official-Models/wan2.2_i2v_A14b_low_noise_lightx2v.safetensors \
--output /path/to/output/ \
--output_ext .safetensors \
--output_name wan2.2_i2v_A14b_low_noise_scaled_fp8_e4m3_lightx2v_4step \
--model_type wan_dit \
--lora_path /path/to/loras/wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--lora_strength 1.0 \
--quantized \
--linear_dtype torch.float8_e4m3fn \
--non_linear_dtype torch.bfloat16 \
--single_fileMerge LoRA + ComfyUI FP8 Format:
cd LightX2V/tools/convert
# For directory-based base model: --source /path/to/Wan2.2-I2V-A14B/high_noise_model/
python converter.py \
--source ./models/Wan2.2-Official-Models/wan2.2_i2v_A14b_high_noise_lightx2v.safetensors \
--output /path/to/output/ \
--output_ext .safetensors \
--output_name wan2.2_i2v_A14b_high_noise_scaled_fp8_e4m3_lightx2v_4step_comfyui \
--model_type wan_dit \
--lora_path /path/to/loras/wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--lora_strength 1.0 \
--quantized \
--linear_dtype torch.float8_e4m3fn \
--non_linear_dtype torch.bfloat16 \
--single_file \
--comfyui_mode
# For directory-based base model: --source /path/to/Wan2.2-I2V-A14B/low_noise_model/
python converter.py \
--source ./models/Wan2.2-Official-Models/wan2.2_i2v_A14b_low_noise_lightx2v.safetensors \
--output /path/to/output/ \
--output_ext .safetensors \
--output_name wan2.2_i2v_A14b_low_noise_scaled_fp8_e4m3_lightx2v_4step_comfyui \
--model_type wan_dit \
--lora_path /path/to/loras/wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--lora_strength 1.0 \
--quantized \
--linear_dtype torch.float8_e4m3fn \
--non_linear_dtype torch.bfloat16 \
--single_file \
--comfyui_mode📝 Reference Documentation : For more merging options, see LightX2V Model Conversion Documentation
Method 2: LightX2V - Online LoRA Loading#
Online LoRA loading requires no pre-merging, loads dynamically during inference, more flexible.
2.1 Download LoRA Models#
# Download both LoRAs (high noise and low noise)
# Note: xxx represents version number, please check HuggingFace for actual filename
huggingface-cli download lightx2v/Wan2.2-Distill-Loras \
wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors \
--local-dir ./loras/2.2 Use Configuration File#
Reference configuration file: wan_moe_i2v_distill_with_lora.json
LoRA configuration example in config file:
{
"lora_configs": [
{
"name": "high_noise_model",
"path": "/path/to/loras/wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_xxx.safetensors",
"strength": 1.0
},
{
"name": "low_noise_model",
"path": "/path/to/loras/wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors",
"strength": 1.0
}
]
}💡 Tip : Replace xxx with actual version number (e.g., 1022). Check HuggingFace repository for the latest version
2.3 Run Inference#
Use the I2V launch script .
In the script, set lightx2v_path, model_path (the Wan2.2-I2V-A14B base model directory), CUDA_VISIBLE_DEVICES, and the input/output paths. Set --config_json to ${lightx2v_path}/configs/distill/wan22/wan_moe_i2v_distill_with_lora.json for online LoRA loading.
In that JSON file, update both paths in lora_configs to your downloaded high-noise and low-noise LoRA files. Then run from the LightX2V repository root:
bash scripts/wan22/inference/distill/run_wan22_moe_i2v_distill.shMethod 3: ComfyUI#
Please refer to workflow
⚠️ Important Notes#
Base Model Requirement : These LoRAs must be used with Wan2.2-I2V-A14B base model, cannot be used standaloneOther Components : In addition to DIT model and LoRA, the following are also required at runtime:T5 text encoder CLIP vision encoder VAE encoder/decoder Tokenizer Please refer to LightX2V Documentation for how to organize complete model directory Inference Configuration : When using 4-step inference, configure correct denoising_step_list, recommended: [1000, 750, 500, 250]Documentation Links#
If you find this project helpful, please give us a ⭐ on GitHub
