Hunyuan Video Introduction

HunyuanVideo utilizes a comprehensive framework that integrates data curation, image-video joint model training, and efficient infrastructure for large-scale model training and inference. It employs a Transformer design with Full Attention mechanism and a unique 'Dual-stream to Single-stream' hybrid model architecture for video generation. Unlike traditional models using CLIP or T5 encoders, it leverages a Multimodal Large Language Model (MLLM) with a Decoder-Only structure to improve image-text alignment. The model also features a Prompt Rewrite Model that fine-tunes user prompts to better match model preferences. Additionally, it implements FP8 quantization for 50% reduced memory usage while maintaining precision.