Why is Wan2.2-S2V generation slow?

Lower the output resolution or reduce the number of sampler steps to speed up generation. Using the fp8 model variant instead of bf16 also reduces VRAM usage and generation time significantly.

Install Wan2.2-S2V in ComfyUI: Speaking Avatar Guide (2026)

Q: What models does Wan2.2-S2V require?

Wan2.2-S2V requires a diffusion model (wan2.2_s2v_14B), a VAE (wan_2.1_vae), a text encoder (umt5_xxl_fp8), a LoRA (lightx2v_I2V_14B), and an audio encoder (wav2vec2_large_english_fp16). All are available free on Hugging Face.

Q: How long should the audio clip be for Wan2.2-S2V?

Around 8 seconds works well for testing. Longer audio clips are supported but require more VRAM and generation time. Use clear, high-quality audio for the best lip-sync results.

Wan2.2-S2V is a powerful AI workflow that generates realistic talking avatar videos from audio and images. This guide explains how to install Wan2.2-S2V, set up all required models, and start creating natural lip-sync video avatars using ComfyUI on your local machine.

What is Wan2.2-S2V?

Quick answer: Wan2.2-S2V is an open-source ComfyUI workflow that turns an audio recording and a still image into a realistic talking avatar video — with natural lip-sync and expressive facial movements, running fully locally on your GPU.

Wan2.2-S2V is an open-source workflow designed for ComfyUI. It converts audio recordings and still images into talking video avatars with natural lip-sync and expressive facial movements. This tool is ideal for content creators, educators, marketers, and anyone who wants to create professional AI video presentations locally.

Example outputs — what Wan2.2-S2V can generate

Demo 1 — Professional Speaking Avatar

Wan2.2-S2V output: natural lip-sync and head movement from audio + image input.

Demo 2 — Song/Music Avatar

Wan2.2-S2V output: expressive avatar generated from a music audio track.

Download the Wan2.2-S2V Workflow

Download the ready-to-use workflow JSON file below. This file is pre-configured with all node connections — just load it into ComfyUI and select your model files.

💡 Tip: Load the workflow into ComfyUI first before placing model files. It will show you exactly which node needs which model file.

Drag and drop this JSON file into your ComfyUI canvas to load it instantly.

Download Required Wan2.2-S2V Models

Make sure you download all five required model files and place them in the correct folders inside your ComfyUI/models/ directory before running the workflow.

Folder Structure — Where to Place All Files

Place your downloaded models exactly like this:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── wan2.2_s2v_14B_fp8_scaled.safetensors
    │   └── wan2.2_s2v_14B_bf16.safetensors
    ├── vae/
    │   └── wan_2.1_vae.safetensors
    ├── text_encoders/
    │   └── umt5_xxl_fp8_e4m3fn_scaled.safetensors
    ├── audio_encoders/
    │   └── wav2vec2_large_english_fp16.safetensors
    └── loras/
        └── lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16.safetensors

⚠️ Warning: Note the audio_encoders folder — this is unique to S2V and different from standard ComfyUI setups. Create it if it doesn't exist inside ComfyUI/models/.

Step-by-Step Setup Guide to Install Wan2.2-S2V

Follow these seven steps to get Wan2.2-S2V running inside ComfyUI on your local machine.

Update ComfyUI

Make sure ComfyUI is updated to the latest version using the ComfyUI Manager or by downloading the newest release from GitHub.

Import the Workflow

Download the Wan2.2-S2V JSON workflow file from the button above, then drag and drop it into your ComfyUI canvas window to load it.

Place All Model Files

Download all five required model files listed above. Place each file in the exact correct subfolder under ComfyUI/models/ as shown in the folder structure.

Configure Nodes in ComfyUI

In the workflow, assign the correct file to each node:

Load Diffusion Model: wan2.2_s2v_14B_fp8_scaled.safetensors (or bf16)
Load Audio Encoder: wav2vec2_large_english_fp16.safetensors
Load CLIP: umt5_xxl_fp8_e4m3fn_scaled.safetensors
Load VAE: wan_2.1_vae.safetensors
Load LoRA: lightx2v_I2V_14B_480p_..._bf16.safetensors

Add Your Audio, Image, and Prompt

Attach an audio file (around 8 seconds is good for testing — longer is fine if you have enough VRAM). Attach the image you want to animate. Enter a prompt describing the speaking style, gestures, and expression.

Adjust Settings (Optional)

Change width and height as needed. Adjust CFG scale and sampler steps — higher steps give better quality but take longer. For first tests, use default settings.

Run the Workflow

Click Queue Prompt / Run in ComfyUI to generate your talking avatar video. Wait for the process to complete — generation time depends on video length and GPU.

Example Run

Here is a real example generated with Wan2.2-S2V in ComfyUI using the prompt below. Click the video to play it.

Prompt Used

In the video, a woman is speaking professionally to the audience. Her expression is confident, clear, and engaging, as if she is introducing new technology. She maintains natural eye contact with the camera, gestures slightly with her hands for emphasis, and her posture is upright and professional. Her delivery is calm, friendly, and polished, like a presenter on a stage.

Output Result

Wan2.2-S2V output: professional speaking avatar with confident posture, natural gestures, and clear lip-sync.

Tips & Troubleshooting

If something isn't working as expected, these are the most common causes and fixes:

Video generation is slow

Fix: Lower the output resolution or reduce the number of sampler steps. Use the fp8 diffusion model variant instead of bf16 — it significantly reduces VRAM usage and generation time.

Audio not syncing well with lip movements

Fix: Use clear, high-quality mono audio clips with minimal background noise. Noisy or compressed audio reduces lip-sync accuracy. Try trimming the clip to 8–15 seconds first.

Crashes or CUDA out-of-memory errors

Fix: Double-check all model files are in the correct folders. Reduce resolution and step count. Switch to the fp8 diffusion model if you're using bf16. Make sure the audio_encoders folder exists and contains the wav2vec2 file.

Output looks unnatural or stiff

Fix: Refine your prompt to describe gestures and expressions more precisely. Include phrases like 'natural eye contact', 'slight hand gestures', and 'upright posture' for more lifelike results.

Model files not appearing in node dropdowns

Fix: After placing new files in ComfyUI folders, press F5 or Ctrl+Shift+R to fully refresh ComfyUI in your browser. New files will not appear until a full refresh.

💡 Tip: For a complete list of ComfyUI errors and fixes across all workflow types, see the ComfyUI Troubleshooting Guide.

Conclusion

Wan2.2-S2V makes it easy to create realistic talking avatar videos from your audio and images using AI — entirely locally, with no subscription required. By following this guide, anyone can install Wan2.2-S2V, set up the necessary models, and start generating professional-quality video presentations in ComfyUI. The fp8 model keeps VRAM requirements manageable for mid-range GPUs.

Frequently Asked Questions

Wan2.2-S2V is an open-source ComfyUI workflow that generates realistic talking avatar videos from an audio file and a still image. It produces natural lip-sync and expressive facial movements using a 14B AI diffusion model.

Wan2.2-S2V requires five model files: a diffusion model (wan2.2_s2v_14B), a VAE (wan_2.1_vae), a text encoder (umt5_xxl_fp8), a LoRA (lightx2v_I2V_14B), and an audio encoder (wav2vec2_large_english_fp16). All are available free on Hugging Face.

Place the diffusion model in ComfyUI/models/diffusion_models, the VAE in ComfyUI/models/vae, the text encoder in ComfyUI/models/text_encoders, the LoRA in ComfyUI/models/loras, and the audio encoder in ComfyUI/models/audio_encoders (create this folder if it doesn't exist).

Lower the output resolution or reduce the number of sampler steps to speed things up. Using the fp8 diffusion model variant instead of bf16 also significantly reduces VRAM usage and generation time.

Around 8 seconds works well for testing. Longer clips are supported but require more VRAM and generation time. Always use clear, high-quality audio for the best lip-sync results.