How to Run WAN 2.2 on RunPod:The Complete Visual Guide (2026)

Generate cinematic AI videos from text or images — step by step, with screenshots.

By Pek Pongpaet · Impekable · Updated May 2026 · 15 min read

What's in This Guide 

1. Introduction - Why WAN 2.2 + RunPod? 

2. Definition of Terms 

3. WAN 2.2 Model Variants & GPU Requirements 

4. Part 1 - Deploying the Pod (Steps 1–6) 

5. Part 2 - Loading the Workflow (Steps 7–8) 

6. Part 3 - Configuring the Workflow (Steps 9–12) 

7. Part 4 - Generating Video (Steps 13–14) 

8. Quick Reference Settings Table 

9. Pro Tips & Common Mistakes 

10. WAN 2.2 vs. Alternatives 

11. Official Resources & Links 

12. Call to Action 

13. FAQ (8 questions)

Introduction: Why Run WAN 2.2 on RunPod? 

AI video generation has had a breakout year. But for most creators, the biggest barrier isn't the software, it's the hardware. Running WAN 2.2, Alibaba's state-of-the-art open-source video model, demands at least 48 GB of GPU VRAM. That's data-center-grade hardware most people will never own. 

RunPod solves this. Instead of a $10,000+ workstation, you rent an enterprise GPU for a few dollars per hour spin it up, generate your video, and shut it down. No ongoing cost, no hardware maintenance, no compromise on quality. 

WAN 2.2 (released July 28, 2025) uses a Mixture-of-Experts (MoE) architecture to deliver cinematic-quality video at 720p/24fps. It's free to use (Apache 2.0 license), integrates with ComfyUI, and on benchmarks it matches or beats several leading commercial tools. This guide walks you through every step, screenshots included. 

WAN 2.2 Model Variants & GPU Requirements

■ Cost tip: A 5-second video on an H100 takes ~3–8 min, costing roughly $0.25–$0.60/clip. Always stop your pod when not generating.

Part 1 - Deploying the Pod

1.Navigate to RunPod

Click the link in the video description to go directly to the RunPod deploy screen.

2.Select your GPU

Choose a GPU with at least 48 GB VRAM. The H100 is recommended for best quality. Any GPU meeting the VRAM requirement will work, choose what fits your budget.

3.Open the template editor

Before deploying, click Edit Template and expand the Environment Variables tab.

4.Set the download flag

Find the environment variable download_WAN_2.2, set its value to true, then click Set Overrides.

⚠  Required:
This environment variable must be set to true before deploying. If skipped, the Wan 2.2 model will not be downloaded and the workflow will not function.

5.Deploy the pod

Scroll to the bottom of the deploy screen and click Deploy On Demand. Then navigate to My Pods.

6.Wait for the pod to be ready

Expand the pod logs. Wait until you see the message "config is up".
Once ready, click Connect → Connect to ComfyUI.

Part 2 - Loading the Workflow

7.Close the promotional screen

When ComfyUI first loads, you may see a promotional overlay. Click the X in the top right corner to dismiss it.

8.Open the Workflows folder

Click the folder icon in the left sidebar to open the Workflows panel.

9.Select the Wan 2.2 workflow

Inside the Workflows folder, open the Wan 2.2 subfolder. Choose either text_to_video or image_to_video - both work identically except for the input source.

ℹ  Note:
Workflows are linear and read left to right. If you have used previous Wan workflows, the layout will look familiar.

Part 3 - Configuring the Workflow

10.Verify models on the left side

Confirm that your CLIP, VAE, and Upscaler models are loaded correctly in the leftmost nodes.

11.Understand the dual-model architecture

Wan 2.2 uses two diffusion models: a high noise model and a low noise model. The high noise model must be used in Sampler 1 and the low noise model in Sampler 2. Do not swap them.

⚠  Critical:
The high noise model must always be assigned to the first sampler and the low noise model to the second. Swapping them will produce poor results.

12.Enter your prompts

Type your text prompt and negative prompt in the designated fields on the right side of the workflow.

13.Set video resolution and frame count

Select your desired resolution. For a 5-second video, use 121 frames (Wan 2.2 samples at 24 fps).

14.Set the number of steps

Enter a step count. This value must be an even number because the steps are divided equally between the two samplers.

15.Configure sampler settings

Recommended starting values: CFG 3.5, Euler sampler, Simple scheduler. Apply the same settings to both samplers.

ℹ  Note:
Best sampler settings for anime or realistic styles are still being explored by the community. Check the creator's Discord server for updated recommendations.

16.Leave upscale factor at default

Do not change the upscale factor setting. The quality improvement is negligible but the generation time increases significantly.

Part 4 - Generating Video

17.Queue the generation

Once all settings are configured, queue the workflow. The final nodes will automatically upscale and frame-interpolate the output to 60 fps.

18.Image-to-video workflow

The image-to-video workflow is identical to text-to-video with one difference: you must load a source image in the input node on the left side before queuing.

Quick Reference

Definition of Terms

WAN 2.2

Open-source AI video generation model by Alibaba. Converts text or images into video clips at up to 720p/24fps. Apache 2.0 licensed, free for commercial use. 

github.com/Wan-Video/Wan2.2
huggingface.co/Wan-A

RunPod 

Cloud GPU marketplace, rent high-performance GPUs by the minute. No contracts, no setup headaches. Pay only while running. 

runpod.io·
docs.runpod.io 

ComfyUI 

Free, open-source visual workflow tool for AI models. Connect blocks (nodes) in a canvas, no coding needed. github.com/comfyanonymous/ComfyUI 

GPU / VRAM

GPU = the chip that powers AI generation. VRAM = its onboard memory. WAN 2.2's 14B models need 24–80 GB VRAM. The H100 on RunPod has 80 GB. 

MoE (Mixture-of-Experts) 

The architecture powering WAN 2.2. Two specialist 14B models share the work: one handles rough early generation (high noise), one refines detail (low noise). 

Text-to-Video (T2V) 

Describe a scene in words → AI generates a matching video clip. 

Image-to-Video (I2V) 

Provide a still image → AI animates it into a video sequence. 

CFG (Classifier-Free Guidance) 

Controls prompt adherence. Higher = closer to your words but may over-process. Recommended start: 3.5. 

Network Volume 

Persistent storage on RunPod. Keeps models and outputs safe after pod shutdown, essential for regular users. 

Pro Tips & Common Mistakes 

Tips for Better Results 

• Write descriptive prompts. Include camera angle, lighting, motion type, and mood. "A slow dolly push through a misty forest at dawn" beats "forest video." 
  
  

• Use a Network Volume. Without one, all models and outputs vanish when the pod stops. Essential for regular use. 

  
  

• Reduce frames before reducing resolution when hitting memory limits frame count multiplies VRAM usage faster than resolution. 

  
  

• Frame interpolation is automatic in the pre-built workflows the 2nd Pass section handles upscaling and RIFE VFI to 60fps. 

  
  

• Save your workflow JSON after dialing in settings. Re-import it to restore the exact config instantly. 

  
  

• Check the Discord community for updated CFG / sampler recommendations — especially for anime or photorealistic styles. 

Common Mistakes to Avoid 

• Not setting download_wan22 = true. Most common error pod launches but WAN 2.2 never downloads. 

• Using an odd step count. Steps are split equally between two samplers. Odd number = unequal split = degraded quality. 

• Swapping sampler models. High noise → Sampler 1. Low noise → Sampler 2. Period. 

• Changing the upscale factor. Don't, the quality gain is negligible but generation time balloons. 

• Terminating the pod without a Network Volume. All outputs are lost permanently. • Leaving the pod idle. You're charged while it's running. Stop it between sessions. 

WAN 2.2 vs. Alternatives

Official Resources & Links 

WAN 2.2 - Official 

• GitHub - source code + model weights

  • github.com/Wan-Video/Wan2.2 
    
    

• Hugging Face - T2V-A14B 

  • huggingface.co/Wan-AI/Wan2.2-T2V-A14B 
    
    

• Hugging Face - I2V-A14B 

  • huggingface.co/Wan-AI/Wan2.2-I2V-A14B 

  
  

• Hugging Face - TI2V-5B (lightweight) 

  • huggingface.co/Wan-AI/Wan2.2-TI2V-5B 

  
  

• Try online (no setup) 

  • wan.video 

RunPod - Official 

• Main site + GPU pricing 

  • runpod.io 

  
  

• Documentation 

  • docs.runpod.io 

  
  

• Pricing page 

  • runpod.io/pricing 

    
    

• WAN 2.2 T2V serverless endpoint 

  • console.runpod.io/hub/playground/video/wan-2-2-t2v-720-lora 

  
  

• WAN 2.2 I2V serverless endpoint 

  • console.runpod.io/hub/playground/video/wan-2-2-i2v-720 

ComfyUI + Community

• ComfyUI GitHub 

  • github.com/comfyanonymous/ComfyUI 

  
  

• ComfyUI Manager 

  • github.com/ltdrdata/ComfyUI-Manager 

  
  

• Civitai — LoRAs & community models 

  • civitai.com 

  
  

• RunPod Discord 

  • discord.gg/runpod 

Ready to Build Something Remarkable? 

At Impekable, we help businesses design and ship AI-powered products people love. Whether you're integrating generative video into your platform, building a content creation tool, or prototyping the next AI-native experience, our team of designers and engineers can take you from idea to launch. 

Ready to Build Something Remarkable? 

At Impekable, we help businesses design and ship AI-powered products people love. Whether you're integrating generative video into your platform, building a content creation tool, or prototyping the next AI-native experience, our team of designers and engineers can take you from idea to launch. 

Ready to Build Something Remarkable? 

At Impekable, we help businesses design and ship AI-powered products people love. Whether you're integrating generative video into your platform, building a content creation tool, or prototyping the next AI-native experience, our team of designers and engineers can take you from idea to launch. 

Frequently Asked Questions 

Pek Pongpaet

Helping enterprises and startups achieve their goals through product strategy, world-class user experience design, software engineering and app development.

Let's talk

Pek Pongpaet

Helping enterprises and startups achieve their goals through product strategy, world-class user experience design, software engineering and app development.

Let's talk

Pek Pongpaet

Helping enterprises and startups achieve their goals through product strategy, world-class user experience design, software engineering and app development.

Let's talk

© 2026 Impekable · impekable.com · Written by Pek Pongpaet, CEO 

WAN 2.2 is open-source software by Alibaba (Apache 2.0). RunPod is an independent platform. Impekable is not affiliated with Alibaba or RunPod. Pricing estimates are approximate and subject to change, verify current GPU pricing at runpod.io/pricing.