HunyuanVideo is Tencent's open-source large-scale video generation foundation model project. The repository contains PyTorch model definitions, pretrained weights, and inference/sampling code, and can serve as a video generation base for research and applications.
In terms of use, HunyuanVideo targets text-to-video generation and can be extended to tasks such as image-to-video, audio-driven character animation, and customized video generation. The official team positions it as an open-source video foundation model, with the goal of approaching or surpassing leading closed-source models in video generation performance.
Main capabilities and architectural features: the model adopts a unified image and video generation architecture, is based on a Transformer design and uses a full attention mechanism, and adopts a "dual-stream to single-stream" hybrid model design: the dual-stream stage processes video tokens and text tokens separately, while the single-stream stage concatenates the two for multimodal information fusion. Text encoding uses a pretrained multimodal large language model (Decoder-Only structure), which has advantages over CLIP and T5-XXL in image-text alignment, detailed description, and complex reasoning, and introduces an additional bidirectional token refiner to enhance text features. Video and images are compressed into a compact latent space through a 3D VAE trained with CausalConv3D, with video length, spatial, and channel compression ratios of 4, 8, and 16, respectively. The model also provides a prompt rewriting model, fine-tuned based on Hunyuan-Large, supporting two rewriting modes, Normal and Master, which can improve understanding of user intent or enhance descriptions of composition, lighting, camera movement, and more.
Usage: the repository provides command-line single-card inference, a Gradio Web demo, multi-GPU parallel inference based on xDiT (Unified Sequence Parallelism), and FP8 quantized weight inference, with FP8 saving about 10GB of VRAM. Installation requires cloning the code from the repository, and manual installation using CUDA 12.6 or 11.8 is recommended; prebuilt Docker images are also provided. Pretrained model weights need to be downloaded separately.
Applicable scenarios: suitable for video generation research, model fine-tuning and secondary development, multimodal generation experiments, and inference in tools such as ComfyUI through community integrations. Community contributions include ComfyUI-Kijai, ComfyUI-Native, FastVideo, HunyuanVideo-gguf, TeaCache, RIFLEx, FramePack, and more.
Limitations: running requiresan NVIDIA GPU that supports CUDA, and the official team tested on a single 80G GPU; the minimum VRAM requirements are 60GB for 720px×1280px×129f and 45GB for 544px×960px×129f, with 80GB VRAM recommended for better generation quality, and the tested operating system is Linux. Evaluation comparisons are based on the high-quality version, which differs from the currently released fast version. Some specific command-line parameters, supported width/height/frame settings, and other details are not fully shown in the materials and are not yet provided.
