This guide provides a comprehensive breakdown of Wan2.2, the latest evolution in open-source video generation from Alibaba’s Tongyi Lab. Whether you are a professional motion designer or a hobbyist, integrating Wan2.2 into ComfyUI offers unparalleled control over cinematic AI video production.

1. What is Wan2.2? The Architecture Revolution

Released in early 2026, Wan2.2 isn't just an incremental update to Wan2.1; it introduces a Mixture-of-Experts (MoE) architecture to the world of video diffusion.

Key Technical Highlights:

• MoE Architecture: Unlike traditional "dense" models that activate all parameters for every calculation, Wan2.2 uses a dual-expert system. It features a High-Noise Expert for initial scene layout/motion and a Low-Noise Expert for fine-textured details.
• Efficiency: Despite having 27B total parameters, only 14B are active at any given time. This allows for high-tier quality with the VRAM footprint of a much smaller model.
• Cinematic Aesthetics: The model was trained on a dataset with over 80% more video content than its predecessor, specifically labeled for lighting, contrast, and professional camera movement.
• Performance: Capable of native 720p and 1280p outputs at 24fps.

2. Model Variants: 14B vs. 5B

Choosing the right version depends entirely on your hardware:

3. Installation Guide for ComfyUI

To run Wan2.2, you need an up-to-date ComfyUI installation and the specific custom nodes designed for Wan's video wrappers.

Step 1: Install Custom Nodes

Open your ComfyUI Manager and search for:

1. ComfyUI-WanVideoWrapper (by Kijai): The most stable implementation for Wan2.2.
2. ComfyUI-VideoHelperSuite: Essential for loading images and saving MP4/GIF outputs.
3. ComfyUI-KJNodes: Provides specialized mask and noise tools.

Step 2: Download the Weights

Place your model files in the following directories:

• Diffusion Model: ComfyUI/models/checkpoints/ (or models/diffusion_models/ depending on your node version).
• VAE: ComfyUI/models/vae/ (Ensure you use the dedicated Wan2.2 VAE for proper 16x16x4 compression).
• Text Encoders: Usually requires T5-v1.1-xxl and UMT5, placed in models/clip/.

4. Text-to-Video (T2V) Workflow Tutorial

Generating a video from scratch requires a structured prompt and proper sampler settings.

The Node Setup

1. WanVideo Loader: Select the Wan2.2-T2V-14B checkpoint.
2. Empty Wan Latent: Set your resolution. For 14B, 1280x720 is the sweet spot. Set frames to 81 (approx. 5 seconds at 24fps).
3. CLIP Text Encode: Wan2.2 understands natural language better than tags.
   • Good Prompt: "A cinematic tracking shot of a futuristic cyberpunk city during a rainstorm, neon lights reflecting on puddles, hyper-realistic, 8k, high contrast."
4. KSampler (Advanced):
   • Steps: 30–50.
   • CFG: 5.0 to 7.0 (Wan2.2 is sensitive; don't go too high).
   • Sampler: uni_pc or euler.
   • Scheduler: simple.

5. Image-to-Video (I2V) Workflow Tutorial

The I2V model is the "gold standard" for 2026, allowing you to animate static AI-generated art with incredible temporal consistency.

Step-by-Step Implementation

1. Load Image: Upload a high-quality source image (e.g., from Midjourney or Flux).
2. WanVideo I2V Loader: Select the Wan2.2-I2V-14B model.
3. Image-to-Latent: Connect your image to the WanVideo I2V Encoder. This converts the pixels into the latent space the model understands.
4. Prompting: Describe the action only.
   • Example: "The character turns their head and smiles at the camera, wind blowing through hair."
5. Motion Bucket: Adjust the "Motion" parameter. Higher values (80+) create more aggressive movement; lower values (30-50) are better for subtle portraits.

6. Advanced Optimization Techniques

Using Lightx2v V2 LoRA

If your generation takes too long, use the Lightx2v V2 distillation LoRA. This allows you to drop your sampling steps from 40 down to 8–12 without losing significant quality. This is a game-changer for those running on a single RTX 3080 or 4070.

VRAM Management

If you run out of memory (OOM error):

• Enable fp8 or bf16 precision in the loader.
• Use the "VAE Tile Encode" node to process the video in chunks during the final decoding stage.
• Reduce the resolution to 832x480 for a faster "preview" workflow.

Conclusion

Wan2.2 represents a massive leap in accessibility for high-end video generation. By leveraging the MoE architecture in ComfyUI, you can produce professional-grade clips that rival commercial closed-source tools.