HunyuanVideo: A Systematic Framework For Large Video Generation Model Training
This repo contains the weights of HunyuanVideo-PromptRewrite model, which can be directly deployed and inferred using the Hunyuan-Large original code. Our project page is here.
HunyuanVideo: A Systematic Framework For Large Video Generation Model Training
Contents
Abstract
We present HunyuanVideo, a novel open-source video foundation model that exhibits performance in video generation that is comparable to, if not superior to, leading closed-source models. HunyuanVideo features a comprehensive framework that integrates several key contributions, including data curation, image-video joint model training, and an efficient infrastructure designed to facilitate large-scale model training and inference. Additionally, through an effective strategy for scaling model architecture and dataset, we successfully trained a video generative model with over 13 billion parameters, making it the largest among all open-source models.
We conducted extensive experiments and implemented a series of targeted designs to ensure high visual quality, motion diversity, text-video alignment, and generation stability. According to professional human evaluation results, HunyuanVideo outperforms previous state-of-the-art models, including Runway Gen-3, Luma 1.6, and 3 top performing Chinese video generative models. By releasing the code and weights of the foundation model and its applications, we aim to bridge the gap between closed-source and open-source video foundation models. This initiative will empower everyone in the community to experiment with their ideas, fostering a more dynamic and vibrant video generation ecosystem.
HunyuanVideo Overall Architechture
HunyuanVideo is trained on a spatial-temporally compressed latent space, which is compressed through Causal 3D VAE. Text prompts are encoded using a large language model, and used as the condition. Gaussian noise and condition are taken as input, our generate model generates an output latent, which is decoded to images or videos through the 3D VAE decoder.
π HunyuanVideo Key Features
Unified Image and Video Generative Architecture
HunyuanVideo introduces the Transformer design and employs a Full Attention mechanism for unified image and video generation. Specifically, we use a "Dual-stream to Single-stream" hybrid model design for video generation. In the dual-stream phase, video and text tokens are processed independently through multiple Transformer blocks, enabling each modality to learn its own appropriate modulation mechanisms without interference. In the single-stream phase, we concatenate the video and text tokens and feed them into subsequent Transformer blocks for effective multimodal information fusion. This design captures complex interactions between visual and semantic information, enhancing overall model performance.
MLLM Text Encoder
Some previous text-to-video model typically use pretrainednCLIP and T5-XXL as text encoders where CLIP uses Transformer Encoder and T5 uses a Encoder-Decoder structure. In constrast, we utilize a pretrained Multimodal Large Language Model (MLLM) with a Decoder-Only structure as our text encoder, which has following advantages: (i) Compared with T5, MLLM after visual instruction finetuning has better image-text alignment in the feature space, which alleviates the difficulty of instruction following in diffusion models; (ii) Compared with CLIP, MLLM has been demonstrated superior ability in image detail description and complex reasoning; (iii) MLLM can play as a zero-shot learner by following system instructions prepended to user prompts, helping text features pay more attention to key information. In addition, MLLM is based on causal attention while T5-XXL utilizes bidirectional attention that produces better text guidance for diffusion models. Therefore, we introduce an extra bidirectional token refiner for enhacing text features.
3D VAE
HunyuanVideo trains a 3D VAE with CausalConv3D to compress pixel-space videos and images into a compact latent space. We set the compression ratios of video length, space and channel to 4, 8 and 16 respectively. This can significantly reduce the number of tokens for the subsequent diffusion transformer model, allowing us to train videos at the original resolution and frame rate.
Prompt Rewrite
To address the variability in linguistic style and length of user-provided prompts, we fine-tune the Hunyuan-Large model as our prompt rewrite model to adapt the original user prompt to model-preferred prompt.
We provide two rewrite modes: Normal mode and Master mode, which can be called using different prompts. The Normal mode is designed to enhance the video generation model's comprehension of user intent, facilitating a more accurate interpretation of the instructions provided. The Master mode enhances the description of aspects such as composition, lighting, and camera movement, which leans towards generating videos with a higher visual quality. However, this emphasis may occasionally result in the loss of some semantic details.
The Prompt Rewrite Model can be directly deployed and inferred using the Hunyuan-Large original code. We release the weights of the Prompt Rewrite Model here.
π Comparisons
To evaluate the performance of HunyuanVideo, we selected five strong baselines from closed-source video generation models. In total, we utilized 1,533 text prompts, generating an equal number of video samples with HunyuanVideo in a single run. For a fair comparison, we conducted inference only once, avoiding any cherry-picking of results. When comparing with the baseline methods, we maintained the default settings for all selected models, ensuring consistent video resolution. Videos were assessed based on three criteria: Text Alignment, Motion Quality and Visual Quality. More than 60 professional evaluators performed the evaluation. Notably, HunyuanVideo demonstrated the best overall performance, particularly excelling in motion quality.
Model | Open Source | Duration | Text Alignment | Motion Quality | Visual Quality | Overall | Ranking |
---|---|---|---|---|---|---|---|
HunyuanVideo (Ours) | β | 5s | 61.8% | 66.5% | 95.7% | 41.3% | 1 |
CNTopA (API) | β | 5s | 62.6% | 61.7% | 95.6% | 37.7% | 2 |
CNTopB (Web) | β | 5s | 60.1% | 62.9% | 97.7% | 37.5% | 3 |
GEN-3 alpha (Web) | β | 6s | 47.7% | 54.7% | 97.5% | 27.4% | 4 |
Luma1.6 (API) | β | 5s | 57.6% | 44.2% | 94.1% | 24.8% | 6 |
CNTopC (Web) | β | 5s | 48.4% | 47.2% | 96.3% | 24.6% | 5 |
π BibTeX
If you find HunyuanVideo useful for your research and applications, please cite using this BibTeX:
@misc{kong2024hunyuanvideo,
title={HunyuanVideo: A Systematic Framework For Large Video Generative Models},
author={Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Junkun Yuan, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yanxin Long, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, and Jie Jiang, along with Caesar Zhong},
year={2024},
archivePrefix={arXiv preprint arXiv:2412.03603},
primaryClass={cs.CV}
}
Acknowledgements
We would like to thank the contributors to the SD3, FLUX, Llama, LLaVA, Xtuner, diffusers and HuggingFace repositories, for their open research and exploration. Additionally, we also thank the Tencent Hunyuan Multimodal team for their help with the text encoder.
- Downloads last month
- 303