ComfyUI HunyuanVideo Guide: Text-to-Video Generation Setup

What is HunyuanVideo?

HunyuanVideo is Tencent's open-source video generation model. It produces high-quality videos from text descriptions with good motion coherence and visual fidelity. It's one of the first major video models to receive native ComfyUI support.

Hardware Requirements

Model Download

Download these files and place them in the corresponding ComfyUI folders:

Diffusion Model

Text Encoders

VAE

Directory Structure

ComfyUI/
├── models/
│   ├── diffusion_models/
│   │   └── hunyuan_video_t2v_720p_bf16.safetensors
│   ├── text_encoders/
│   │   ├── clip_l.safetensors
│   │   └── llava_llama3_fp8_scaled.safetensors
│   └── vae/
│       └── hunyuan_video_vae_bf16.safetensors

Supported Resolutions

HunyuanVideo supports multiple aspect ratios:

Frame count is typically 73 or 129 frames.

Workflow Setup

Key Nodes

1. UNET Loader — loads hunyuan_video_t2v_720p_bf16.safetensors. Set weight type to default (or fp8 for lower VRAM)
2. DualCLIPLoader — loads both clip_l.safetensors and llava_llama3_fp8_scaled.safetensors, type set to hunyuan_video
3. VAE Loader — loads the VAE model
4. EmptyHunyuanLatentVideo — sets video dimensions and frame count
5. CLIP Text Encode — your video description prompt
6. FluxGuidance — controls prompt adherence (default: 6.0)
7. KSampler — sampler: euler, scheduler: simple, steps: 20–30
8. VAEDecodeTiled — decodes video (use tiled version for memory efficiency)
9. Save Animated WEBP — saves the output

Running the Workflow

1. Verify all models are loaded in the correct nodes
2. Set video dimensions and frame count in EmptyHunyuanLatentVideo
3. Write a detailed prompt describing the scene, motion, and style
4. Click Run (Ctrl+Enter)

Prompt Tips

HunyuanVideo responds well to structured prompts:

[Subject], [Action], [Scene], [Style], [Quality]

Example:

a girl with long black hair wearing a white dress walking through
a field of sunflowers, golden hour lighting, cinematic composition,
high quality, detailed

Use detailed English descriptions. Include motion words (walking, running, flowing) for better animation.

Memory Optimization

If you're running into VRAM limits:

1. Switch to FP8 — In the UNET Loader, change weight type to fp8_e4m3fn
2. Reduce VAEDecodeTiled settings — Lower tile_size to 128, overlap to 32
3. Lower resolution and frame count — Use 540p instead of 720p, fewer frames
4. Close other applications — HunyuanVideo is memory-intensive

Common Issues and Fixes

Out of memory

• Use FP8 weight type in UNET Loader
• Reduce tile_size and overlap in VAEDecodeTiled
• Use 540p resolution with 73 frames instead of 129

Slow generation

• Video generation is inherently slow — a 5-second clip can take 10–30 minutes depending on hardware
• Reduce steps to 20 and use fewer frames
• Use lower resolution for test runs, then increase for final output

Poor video quality

• Increase sampling steps to 25–30
• Adjust FluxGuidance value (try 5.0–8.0)
• Write more detailed, specific prompts
• Try different samplers (euler_ancestral, dpmpp_2m)

EmptyHunyuanLatentVideo node not found

• Update ComfyUI to the latest version — this node was added in a recent update

Related Guides

• Wan Video Guide — Alibaba's video generation models
• FramePack Guide — Low-VRAM video generation
• Text to Image — Basic image generation concepts