HunyuanVideo
HunyuanVideo is an open-source large video generation model with over 13 billion parameters, designed to generate high-quality videos from text prompts. It aims to bridge the gap between closed-source and open-source video generation by providing code, weights, and inference tools. It is intended for developers and researchers who want to experiment with state-of-the-art video generation.
✨ Key features
- Unified image and video generation architecture with Full Attention
- Uses MLLM text encoder for better text-video alignment
- 3D VAE compresses video to latent space for efficient training
- Prompt rewrite modes (Normal and Master) to adapt user prompts
- Supports parallel inference on multiple GPUs via xDiT
- FP8 quantized weights available to reduce GPU memory
🎯 Use cases
- Generate videos from text prompts for creative projects
- Research on video generation models and techniques
- Build custom video generation applications with open-source code
- Use as a base for image-to-video or other video generation tasks
📦 Installation
🧰 Requirements: Requires Linux, Python, PyTorch, and CUDA-compatible GPU. Pretrained models are downloaded from Hugging Face.
git clone https://github.com/Tencent-Hunyuan/HunyuanVideo.git
cd HunyuanVideoconda create -n hunyuanvideo python=3.11
conda activate hunyuanvideopip install -r requirements.txt🚀 Usage
# Single GPU inference
python hyvideo/sample.py \
--model_dir ckpts \
--text_prompt "A cat playing with a ball" \
--output_dir outputs
❓ FAQ
What is the model size of HunyuanVideo?
HunyuanVideo has over 13 billion parameters, making it the largest among open-source video generation models.
How can I run inference on multiple GPUs?
You can use the parallel inference code powered by xDiT, which is released in the repository.
Are there quantized weights available?
Yes, FP8 quantized weights are released to save GPU memory.
Is there a prompt rewrite feature?
Yes, HunyuanVideo provides a prompt rewrite model with Normal and Master modes to adapt user prompts for better generation.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.