Complete Wan 2.2 ComfyUI Workflow Guide: 5B TI2V Hybrid, 14B T2V, I2V & FLF2V Setup

Wan 2.2 from WAN AI represents one of the most powerful open-source video generation models available. Featuring an innovative Mixture of Experts (MoE) architecture with specialized high-noise and low-noise diffusion models, Wan 2.2 delivers exceptional cinematic composition, large-scale fluid motion, and tight prompt compliance.

With native support integrated directly into ComfyUI using official Comfy-Org repackaged weights, running Wan 2.2 locally is fast and flexible. Whether you are operating on an 8GB GPU using the lightweight 5B TI2V Hybrid model or building master 4K production pipelines with the flagship 14B MoE checkpoints, this guide walks you through every step of setup and workflow execution.

1. Introduction & Core Strengths

Before configuring your nodes, let’s understand why Wan 2.2 has become a staple for local AI video creators:

Key Strengths of Wan 2.2

• MoE Dual-Expert Architecture (14B Models): Splitting sampling between a high-noise expert (for structural composition and camera movement) and a low-noise expert (for fine facial detail and texture refinement) maximizes rendering quality.
• 5B TI2V Hybrid Versatility: A unified 5-billion parameter checkpoint that handles both Text-to-Video and Image-to-Video in a single compact node.
• Exceptional Motion Physics: Generates natural fluid dynamics, realistic cloth physics, and complex character actions without melting or warping.
• Comfy-Org Single-File Repackaging: Official single-file .safetensors distribution simplifies VAE and text encoder routing inside ComfyUI.

2. Model Overview – Which Version Should You Use?

Wan 2.2 is available in several distinct variants depending on your hardware VRAM budget and generation goals:

Hardware & VRAM Recommendation Summary

3. Getting Started – Loading Official Templates

ComfyUI provides built-in native template workflows designed around Comfy-Org/Wan_2.2_ComfyUI_Repackaged models.

4. Model Files & Directory Structure

To ensure all nodes locate model weights correctly, organize your files into standard ComfyUI directories. Using Comfy-Org repackaged versions is strongly recommended for seamless single-file loading.

5. Step-by-Step Workflow Walkthroughs

Workflow A: Wan 2.2 5B TI2V Hybrid (Text + Image to Video)

The 5B hybrid model is the ultimate starting point for creators on 8GB to 12GB VRAM GPUs.

Switching T2V vs. I2V in the 5B Hybrid

• Pure Text-to-Video (T2V): Bypass or mute the Load Image node (Ctrl+B in ComfyUI). Leave the image socket disconnected.
• Image-to-Video (I2V): Enable the Load Image node (Ctrl+B to un-bypass) and connect its image output into the image socket of Wan2_2_5B_TI2V.

Checklist & Key Settings

1. Model Loader: Select wan2.2_5B_ti2v_fp8.safetensors.
2. Sampling Steps: Set to 20–30 steps.
3. CFG Scale: Set between 5.0 and 6.0.
4. Resolution: Recommended 832 x 480 or 1024 x 576.

Workflow B: Wan 2.2 14B Text-to-Video (T2V Dual-Expert MoE)

The flagship 14B T2V model uses a two-stage MoE sampler pass for uncompromised cinematic quality.

Checklist & Key Settings

1. Load Dual Models:
   • High-Noise Model: wan2.2_14B_t2v_high_noise_fp8.safetensors
   • Low-Noise Model: wan2.2_14B_t2v_low_noise_fp8.safetensors
2. Configuring Advanced Sampler Split:
   • Total Steps: 30 steps.
   • High-Noise Sampler: Runs from step 0 to step 15 (50% denoise boundary).
   • Low-Noise Sampler: Takes intermediate latents and runs from step 15 to step 30.
3. Resolution: 1280 x 720 or 1024 x 576.

Workflow C: Wan 2.2 14B Image-to-Video (I2V)

The 14B I2V model applies dual-expert MoE sampling to reference images, enabling action sequences while keeping the subject’s face perfectly intact.

1. Load I2V Checkpoints: Connect wan2.2_14B_i2v_high_noise_fp8 and wan2.2_14B_i2v_low_noise_fp8.
2. Upload Reference Image: Connect Load Image into the image latent pre-processor.
3. Prompt Motion: Describe desired camera motion and action (e.g. "The character turns quickly toward the camera, cinematic lighting, wind blowing hair").
4. Execute MoE Sampler: Run dual-pass sampling.

Workflow D: Wan 2.2 14B First & Last Frame (FLF2V)

FLF2V interpolates motion between start and end keyframes using the 14B model architecture.

1. Upload Matching Keyframes: Ensure Image 1 and Image 2 have matching resolutions.
2. Describe Transition: Focus prompt on physical motion and camera trajectory bridging the two images.
3. Execute Interpolation.

6. Prompting Tips for Wan 2.2

Wan 2.2 is exceptionally responsive to cinematic vocabulary, lighting specs, and camera directives.

Recommended Prompt Blueprint

Good vs. Weak Prompt Comparison

7. Performance Optimization & Ecosystem Tools

Community Acceleration Tools

• LightX2V Turbo LoRA: Speeds up Wan 2.2 sampling from 30 steps down to 4–8 steps for near real-time previewing.
• ComfyUI-GGUF Quantization: Enables running the 14B MoE models on 12GB to 16GB GPUs via GGUF Q4_K_M and Q5_K_S quantizers.
• ComfyUI-WanVideoWrapper: Alternative custom node suite offering fine-grained control over temporal attention masking and offloading.

8. Conclusion & Next Steps

Wan 2.2 is a major advancement in open-weights video generation. By starting with the lightweight 5B TI2V Hybrid model before moving up to the 14B MoE dual-expert workflows, you can dial in the ideal balance between generation speed and cinematic rendering quality on your workstation.

Next Steps to Explore

• Try running Wan2.2 5B TI2V with LightX2V Turbo LoRA for rapid 4-step generation.
• Combine Wan 2.2 with spatial upscalers for 4K video renders inside ComfyUI.
• Explore related guides on our ComfyHub Tutorials Directory.