Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization

Yang Jin, Zhicheng Sun, Kun Xu, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, Kun Gai, Yadong Mu
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:22185-22209, 2024.

Abstract

In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-jin24f, title = {Video-{L}a{VIT}: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization}, author = {Jin, Yang and Sun, Zhicheng and Xu, Kun and Xu, Kun and Chen, Liwei and Jiang, Hao and Huang, Quzhe and Song, Chengru and Liu, Yuliang and Zhang, Di and Song, Yang and Gai, Kun and Mu, Yadong}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {22185--22209}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/jin24f/jin24f.pdf}, url = {https://proceedings.mlr.press/v235/jin24f.html}, abstract = {In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.} }
Endnote
%0 Conference Paper %T Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization %A Yang Jin %A Zhicheng Sun %A Kun Xu %A Kun Xu %A Liwei Chen %A Hao Jiang %A Quzhe Huang %A Chengru Song %A Yuliang Liu %A Di Zhang %A Yang Song %A Kun Gai %A Yadong Mu %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-jin24f %I PMLR %P 22185--22209 %U https://proceedings.mlr.press/v235/jin24f.html %V 235 %X In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.
APA
Jin, Y., Sun, Z., Xu, K., Xu, K., Chen, L., Jiang, H., Huang, Q., Song, C., Liu, Y., Zhang, D., Song, Y., Gai, K. & Mu, Y.. (2024). Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:22185-22209 Available from https://proceedings.mlr.press/v235/jin24f.html.

Related Material