Wan2.2-Distill-Loras / README.md

lightx2v

add pipeline tag for better discoverability :) (#8)

b2cf448 verified 8 days ago

preview code

raw

history blame contribute delete

12.1 kB

metadata

license: apache-2.0
tags:
  - diffusion-single-file
  - comfyui
  - distillation
  - LoRA
  - video
  - video genration
pipeline_tags:
  - image-to-video
  - text-to-video
base_model:
  - Wan-AI/Wan2.2-I2V-A14B
library_name: diffusers
pipeline_tag: image-to-video

🎬 Wan2.2 Distilled LoRA Models

⚡ High-Performance Video Generation with 4-Step Inference Using LoRA

LoRA weights extracted from Wan2.2 distilled models - Flexible deployment with excellent generation quality

🌟 What's Special?

⚡ Flexible Deployment

Base Model + LoRA: Can be combined with base models
Offline Merging: Pre-merge LoRA into models
Online Loading: Dynamically load LoRA during inference
Multiple Frameworks: Supports LightX2V and ComfyUI

🎯 Dual Noise Control

High Noise: More creative, diverse outputs
Low Noise: More faithful to input, stable outputs
Rank 64 LoRA, compact size

💾 Storage Efficient

Small LoRA Size: Significantly smaller than full models
Flexible Combination: Can be combined with quantization
Easy Sharing: Convenient for model weight distribution

🚀 4-Step Inference

Ultra-Fast Generation: Generate high-quality videos in just 4 steps
Distillation Acceleration: Inherits advantages of distilled models
Quality 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.safetensors`	64	More creative image-to-video
I2V	Low Noise	`wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_xxx.safetensors`	64	More stable image-to-video

💡 Note:

xxx in filenames represents version number or timestamp, please check HuggingFace repository for the latest version

These 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-Models

Method 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_file

1.3 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_file

Merge 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_distil_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

Using I2V as example:

cd scripts
bash wan22/run_wan22_moe_i2v_distill.sh

Method 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 standalone
Other 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]

📚 Related Resources

Documentation Links

LightX2V Quick Start: Quick Start Documentation
Model Conversion Tool: Conversion Tool Documentation
Online LoRA Loading: Configuration File Example
Quantization Guide: Quantization Documentation
Model Structure: Model Structure Documentation

Related Models

Distilled Full Models: Wan2.2-Distill-Models
Wan2.2 Official Models: Wan2.2-Official-Models - Contains high noise and low noise base models
Base Model (Wan-AI): Wan2.2-I2V-A14B

🤝 Community & Support

GitHub Issues: https://github.com/ModelTC/LightX2V/issues
HuggingFace: https://huggingface.co/lightx2v/Wan2.2-Distill-Loras
LightX2V Homepage: https://github.com/ModelTC/LightX2V

If you find this project helpful, please give us a ⭐ on GitHub