ali-vilab/MS-Vid2Vid-XL

Video-to-Video

MS-Vid2Vid-XL旨在提升视频生成的时空连续性和分辨率，其作为I2VGen-XL的第二阶段以生成720P的视频，同时还可以用于文生视频、高清视频转换等任务。其训练数据包含了精选的海量的高清视频、图像数据（最短边>=720），可以将低分辨率的视频提升到更高分辨率（1280 * 720），且其可以处理几乎任意分辨率的视频(建议16:9的宽视频)。

MS-Vid2Vid-XL aims to improve the spatiotemporal continuity and resolution of video generation. It serves as the second stage of I2VGen-XL to generate 720P videos, and can also be used for various tasks such as text-to-video synthesis and high-quality video transfer. The training data includes a large collection of high-definition videos and images (with the shortest side >=720), allowing for the enhancement of low-resolution videos to higher resolutions (1280 * 720). It can handle videos of almost any resolution (preferably 16:9 aspect ratio).

Fig.1 MS-Vid2Vid-XL

体验地址(Project experience address): https://modelscope.cn/studios/damo/I2VGen-XL-Demo/summary

模型介绍 (Introduction)

MS-Vid2Vid-XL和I2VGen-XL第一阶段相同，都是基于隐空间的视频扩散模型(VLDM)，且其共享相同结构的时空UNet(ST-UNet)，其设计细节延续我们自研VideoComposer，具体可以参考其技术报告。

MS-Vid2Vid-XL and the first stage of I2VGen-XL share the same underlying video latent diffusion model (VLDM). They both utilize a spatiotemporal UNet (ST-UNet) with the same structure, which is designed based on our in-house VideoComposer. For more specific details, please refer to its technical report.

代码范例 (Code example)

from modelscope.pipelines import pipeline
from modelscope.outputs import OutputKeys

# VID_PATH: your video path
# TEXT : your text description
pipe = pipeline(task="video-to-video", model='damo/Video-to-Video')
p_input = {
            'video_path': VID_PATH,
            'text': TEXT
        }

output_video_path = pipe(p_input, output_video='./output.mp4')[OutputKeys.OUTPUT_VIDEO]

模型局限 (Limitation)

本MS-Vid2Vid-XL可能存在如下可能局限性：

• 目标距离较远时可能会存在一定的模糊，该问题可以通过输入文本来解决或缓解；
• 计算时耗大，因为需要生成720P的视频，隐空间的尺寸为(160 * 90)，单个视频计算时长>2分钟
• 目前仅支持英文，因为训练数据的原因目前仅支持英文输入

This MS-Vid2Vid-XL may have the following limitations:

• There may be some blurriness when the target is far away. This issue can be addressed by providing input text.
• Computation time is high due to the need to generate 720P videos. The latent space size is (160 * 90), and the computation time for a single video is more than 2 minutes.
• Currently, it only supports English. This is due to the training data, which is limited to English inputs at the moment.

相关论文以及引用信息 (Reference)

@article{videocomposer2023,
  title={VideoComposer: Compositional Video Synthesis with Motion Controllability},
  author={Wang, Xiang* and Yuan, Hangjie* and Zhang, Shiwei* and Chen, Dayou* and Wang, Jiuniu and Zhang, Yingya and Shen, Yujun and Zhao, Deli and Zhou, Jingren},
  journal={arXiv preprint arXiv:2306.02018},
  year={2023}
}

@inproceedings{videofusion2023,   
  title={VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation},   
  author={Luo, Zhengxiong and Chen, Dayou and Zhang, Yingya and Huang, Yan and Wang, Liang and Shen, Yujun and Zhao, Deli and Zhou, Jingren and Tan, Tieniu},   
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},   
  year={2023}   
}

使用协议 (License Agreement)

我们的代码和模型权重仅可用于个人/学术研究，暂不支持商用。

Our code and model weights are only available for personal/academic research use and are currently not supported for commercial use.