Ren, Xiaowei
Training Video Foundation Models with NVIDIA NeMo
Patel, Zeeshan, He, Ethan, Mannan, Parth, Ren, Xiaowei, Wolf, Ryan, Agarwal, Niket, Huffman, Jacob, Wang, Zhuoyao, Wang, Carl, Chang, Jack, Bai, Yan, Huang, Tommy, Wang, Linnan, Jain, Sahil, Ramasamy, Shanmugam, Jennings, Joseph, Sirazitdinova, Ekaterina, Sudakov, Oleg, Ma, Mingyuan, Chen, Bobby, Lin, Forrest, Wang, Hao, Sabavat, Vasanth Rao Naik, Niverty, Sriharsha, Ou, Rong, Bhattacharya, Pallab, Page, David, Tajbakhsh, Nima, Aithal, Ashwath
Video Foundation Models (VFMs) have recently been used to simulate the real world to train physical AI systems and develop creative visual experiences. However, there are significant challenges in training large-scale, high quality VFMs that can generate high-quality videos. We present a scalable, open-source VFM training pipeline with NVIDIA NeMo, providing accelerated video dataset curation, multimodal data loading, and parallelized video diffusion model training and inference. We also provide a comprehensive performance analysis highlighting best practices for efficient VFM training and inference.
Cosmos World Foundation Model Platform for Physical AI
NVIDIA, null, :, null, Agarwal, Niket, Ali, Arslan, Bala, Maciej, Balaji, Yogesh, Barker, Erik, Cai, Tiffany, Chattopadhyay, Prithvijit, Chen, Yongxin, Cui, Yin, Ding, Yifan, Dworakowski, Daniel, Fan, Jiaojiao, Fenzi, Michele, Ferroni, Francesco, Fidler, Sanja, Fox, Dieter, Ge, Songwei, Ge, Yunhao, Gu, Jinwei, Gururani, Siddharth, He, Ethan, Huang, Jiahui, Huffman, Jacob, Jannaty, Pooya, Jin, Jingyi, Kim, Seung Wook, Klár, Gergely, Lam, Grace, Lan, Shiyi, Leal-Taixe, Laura, Li, Anqi, Li, Zhaoshuo, Lin, Chen-Hsuan, Lin, Tsung-Yi, Ling, Huan, Liu, Ming-Yu, Liu, Xian, Luo, Alice, Ma, Qianli, Mao, Hanzi, Mo, Kaichun, Mousavian, Arsalan, Nah, Seungjun, Niverty, Sriharsha, Page, David, Paschalidou, Despoina, Patel, Zeeshan, Pavao, Lindsey, Ramezanali, Morteza, Reda, Fitsum, Ren, Xiaowei, Sabavat, Vasanth Rao Naik, Schmerling, Ed, Shi, Stella, Stefaniak, Bartosz, Tang, Shitao, Tchapmi, Lyne, Tredak, Przemek, Tseng, Wei-Cheng, Varghese, Jibin, Wang, Hao, Wang, Haoxiang, Wang, Heng, Wang, Ting-Chun, Wei, Fangyin, Wei, Xinyue, Wu, Jay Zhangjie, Xu, Jiashu, Yang, Wei, Yen-Chen, Lin, Zeng, Xiaohui, Zeng, Yu, Zhang, Jing, Zhang, Qinsheng, Zhang, Yuxuan, Zhao, Qingqing, Zolkowski, Artur
Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make our platform open-source and our models open-weight with permissive licenses available via https://github.com/NVIDIA/Cosmos.