Skip to content
ComfyUI Wiki•
Help Build a Better ComfyUI Knowledge Base Become a Patron
NewsWAN2.2 Open Source Version Released and ComfyUI Native Support in Day 0

WAN2.2 Open Source Version Released and ComfyUI Native Support in Day 0

07/28/2025

WAN2.2 Open Source Version Released and ComfyUI Native Support in Day 0

ComfyUI Wan2.2 Open Source Version Support Release

WAN team has officially released the Wan2.2 open source version! This is a brand new multimodal video generation model that adopts an innovative MoE (Mixture of Experts) architecture, bringing significant quality improvements to video generation. The model is completely open source under the Apache 2.0 license and supports commercial use.

ComfyUI has now achieved native support for Wan2.2 in the first instance! You can now directly experience the video generation technology brought by Wan2.2 in ComfyUI. The model consists of high-noise expert models and low-noise expert models, which can perform expert model division based on denoising time steps, thereby generating higher quality video content.

I have completed the official native version tutorial in the ComfyUI official documentation. For the WanVideoWrapper tutorial, I will update it in the near future.

Wan2.2 Model Features

  • MoE Expert Model Architecture: High-noise expert models handle overall layout, low-noise expert models refine details
  • Cinematic Aesthetic Control: Professional lens language, supporting multi-dimensional visual control including lighting, color, and composition
  • Large-scale Complex Motion: Smoothly restore various complex motions, enhancing motion controllability and naturalness
  • Precise Semantic Adherence: Complex scene understanding, multi-object generation, better restoration of creative intent
  • Efficient Compression Technology: Significant data upgrade compared to version 2.1, 5B version high compression ratio VAE, optimized memory usage

Wan2.2 Technical Breakthroughs

Innovative MoE Architecture Design The Wan2.2 model is the first to successfully apply MoE architecture to video generation diffusion models. The 27B version of this architecture consists of high-noise expert models and low-noise expert models, performing expert model division based on the stage differences in the diffusion model denoising process. The high-noise stage focuses on generating the overall layout of videos, while the low-noise stage pays more attention to detail refinement. This division of labor significantly improves generation quality.

Major Data Training Upgrade Compared to the Wan2.1 model, Wan2.2 has achieved significant expansion in training data, with image data increasing by 65.6% and video data increasing by 83.2%. Data expansion not only improves the model’s generalization ability but also enhances creative diversity, making the model perform more excellently in complex scenes, aesthetic expression, and motion generation.

Aesthetic Fine-tuning and Reinforcement Learning Wan2.2 introduces a dedicated aesthetic fine-tuning stage, integrating film industry standards for lighting design, lens composition principles, and color psychology systems. Through reinforcement learning (RL) technology for further fine-tuning, it effectively aligns with human aesthetic preferences, making generated videos more consistent with professional film standards.

Wan2.2 Model Versions

Original Versions

ComfyUI Repackaged Version

https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged

📚 Available Versions

Wan2.2-TI2V-5B: FP16 Wan2.2-I2V-14B: FP16/FP8 Wan2.2-T2V-14B: FP16/FP8

ComfyUI Support for Wan2.2

ComfyUI, as a leading AI image generation workflow platform, has achieved complete native support for Wan2.2. Users can directly use various Wan2.2 functions in ComfyUI, including:

  • Text to Video Generation: Generate high-quality videos through simple text descriptions
  • Image to Video Conversion: Convert static images into dynamic video content
  • Mixed Mode: Support for combined text and image input modes

ComfyUI’s node-based workflow design makes Wan2.2 usage more flexible and efficient, allowing users to easily combine different parameters and settings to achieve optimal generation results.

Tongyi Wanxiang Web Platform Upgrade

In addition to the open source model, the Tongyi Wanxiang Web platform has also undergone comprehensive upgrades:

Creative Function Renewal

  • Wanxiang Box: Unified creative entry point, supporting unified creation of image and video content
  • Aggregated View: New aggregated view function, supporting aggregation display of tasks with same input and continuous editing

Project Collection Function

  • Project-based Asset Management: Video creation management by project unit
  • Timeline Editing: Provides timeline function, supporting asset editing and arrangement
  • Video Editing and Processing: Supports local editing, redrawing, extension and other operations

Related Links