Tencent-Hunyuan/HunyuanVideo

HunyuanVideo: A Systematic Framework For Large Video Generation Model

What it solves

HunyuanVideo is a large-scale open-source video foundation model designed to bridge the gap between open-source and closed-source video generation. It addresses the challenges of achieving high visual quality, diverse motion, and strong text-to-video alignment in generated content, aiming to compete with leading proprietary models.

How it works

The model operates on a spatial-temporally compressed latent space using a Causal 3D VAE. It employs a "Dual-stream to Single-stream" hybrid Transformer architecture: first processing video and text tokens independently (dual-stream) to learn modality-specific modulations, then concatenating them (single-stream) for multimodal fusion.

Key technical components include:

  • MLLM Text Encoder: Uses a decoder-only Multimodal Large Language Model instead of standard CLIP/T5 encoders for better image-text alignment and complex reasoning.
  • 3D VAE: Compresses pixel-space videos and images into a compact latent space to reduce token count and maintain original resolution/frame rate.
  • Prompt Rewrite: A fine-tuned Hunyuan-Large model that converts user prompts into model-preferred descriptions to improve intent comprehension or visual quality.

Who it’s for

Researchers and developers in the AI community who want to experiment with high-fidelity video generation, as well as users looking for an open-source alternative to closed-source video generators.

Highlights

  • Massive Scale: A video generative model with over 13 billion parameters.
  • High Performance: Outperforms several closed-source baselines in professional human evaluations for motion quality and overall performance.
  • Flexible Inference: Supports single-GPU inference, multi-GPU sequence parallel inference via xDiT, and FP8 quantized weights to reduce memory usage.
  • Ecosystem Integration: Integrated into the Hugging Face Diffusers library and supported by ComfyUI.

Related

  • Project
    Tencent-Hunyuan/HunyuanVideo-I2V

    An image-to-video generation framework that transforms static images into high-quality videos using a multimodal LLM for semantic consistency.

  • Project
    Tencent-Hunyuan/HunyuanVideo-1.5

    A lightweight 8.3B parameter video generation model that enables high-quality text-to-video and image-to-video synthesis on consumer-grade GPUs.

  • Dispatch
    State of Open Video Generation Models in Diffusers

    Hugging Face provides a comprehensive overview of open video generation models and introduces a suite of Diffusers optimizations that can reduce VRAM requirements for models like HunyuanVideo from 60GB to approximately 6.5GB.

  • Project
    madebyollin/taehv

    TAEHV is a tiny auto‑encoder that speeds up and memory‑optimizes video VAE decoding for many diffusion models (Hunyuan, MiniMax‑H3, Wan‑Video, CogVideoX, Open‑Sora, LTX‑2, etc.). It reduces decoding of a 61‑frame 512×320 clip from ~2‑3 s / 6‑9 GB to ~0.5 s / <0.5 GB, with a small quality trade‑off. The repo ships model checkpoints, Python API (`TAEHV`, `StreamingTAEHV`), example notebooks, and integration notes for ComfyUI, stable‑diffusion.cpp, SDNext, and Diffusers.

  • Project
    OpenGVLab/InternVideo

    A series of video foundation models and large-scale datasets designed for multimodal video understanding, long-context modeling, and contextual reasoning.