HunyuanVideo

Architecture

• parameters: 13 billion
• context_length: N/A
• inference_type: Transformer + 3D VAE diffusion
• training_data: Large-scale text-video-image corpus (internal dataset)

Capabilities

• Chinese
• English
• Text-to-video
• Image-to-video generation (high-fidelity, coherent animations)

Benchmarks

• VideoGen: Surpasses existing open models; matches or exceeds leading proprietary models in video quality.

Safety

• Basic content filters; adheres to standard image/video content policies
• No special alignment; focuses on visual quality
• Visual output may reflect biases in training data

Deployment

• regions: Global
• hosting: Hugging Face, GitHub
• integrations: ComfyUI, RunwayGen community projects

Tags