[Arxiv 2024] Pyramidal Flow Matching for Efficient Video Generative Modeling

Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, Zhouchen Lin

[Paper] [Project] [Code]

• We design a unified flow matching objective that jointly generates and decompresses visual content across pyramidal representations.

[ICML 2024 Oral] 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

[Paper] [Project] [Code]

• We present a multimodal LLM capable of both comprehending and generating videos, based on an efficient decomposed video representation.

[ICLR 2024] Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization

Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, Di Zhang, Wenwu Ou, Kun Gai, Yadong Mu

[Paper] [Code]

• We present an effective dynamic discrete visual tokenizer that represents an image as the foreign language in Large Language Models, which supports both multi-modal understanding and generation.

[Arxiv 2024] RectifID: Personalizing Rectified Flow with Anchored Classifier Guidance

Zhicheng Sun, Zhenhao Yang, Yang Jin, Haozhe Chi, Kun Xu, Kun Xu, Liwei Chen, Hao Jiang, Di Zhang, Yang Song, Kun Gai, Yadong Mu

[Paper] [Code]

• We introduce a training-free approach to personalizing rectified flow, based on a fixed-point formulation of classifier guidance.

[CVPR 2023] Learning Instance-Level Representation for Large-Scale Multi-Modal Pretraining in E-commerce

Yang Jin, Yongzhi Li, Zehuan Yuan, Yadong Mu

[Paper]

• We propose a generic multi-modal foundation model in E-commerce that learns the instance-level representation of products and achieves superior performance on massive downstream E-commerce applications.

[NeurIPS 2022 Spotlight] Embracing Consistency: A One-Stage Approach for Spatio-Temporal Video Grounding

Yang Jin, Yongzhi Li, Zehuan Yuan, Yadong Mu

[Paper] [Code]

• We propose STCAT, a new one-stage spatio-temporal video grounding model that enjoys more consistent cross-modal feature alignment and tube prediction. It also achieved state-of-the-art performance on VidSTG and HC-STVG benchmarks.

[CVPR 2022] Complex Video Action Reasoning via Learnable Markov Logic Network

Yang Jin, Linchao Zhu, Yadong Mu

[Paper]

• We devise an video action reasoning framework that performs Markov Logic Network (MLN) based probabilistic logical inference. The proposed framework enjoys remarkable interpretability through the learned logical rules.

[TMM 2021] Zero-Shot Video Event Detection with High-Order Semantic Concept Discovery and Matching

Yang Jin, Wenhao Jiang, Yi Yang, Yadong Mu

[Paper]

• We design an effective zero-shot video event retrieval framework based on the utilization of high-order concepts (such as subject-predicate-object triplets or adjective-object). The constructed high-order concept library and proposed query-expanding scheme significantly improved the perfromance.