Xingchao Liu (刘星超)

Email: xcliu at utexas.edu

Welcome to my homepage! I am a machine learning researcher specializing in probabilistic inference and generative modeling, with a particular emphasis on their applications in multimodal intelligence.

Currently, I am a researcher in the multimodal group at DeepSeek AI. Prior to that, I received my Ph.D. from the University of Texas at Austin, where I had the privilege of being advised by Prof. Qiang Liu. During my undergraduate study at Beihang University, I also worked with Prof. Hao Su in UCSD.

Profile Photo

Research

For full publication list, please refer to [My Google Scholar].

Janus Series


  • Our Janus series of models aims to unify multimodal understanding and generation inside the same model. For now, we have [Janus] and [JanusFlow]. Janus generates images in an autoregressive way, while JanusFlow uses Rectified Flow. They are open-sourced on [github].

  • You may play with these models online at [HF space for Janus] and [HF space for JanusFlow]. Have fun!

    • Rectified Flow


  • Rectified Flow is a new framework for generative modeling. It is easy to understand mathematically. It can match two arbitrary distributions without constraints such as the initial distribution being Gaussian. This makes it suitable for general distribution matching problem, including generative modeling and unsupervised data transfer. Read its arXiv version [here]. The code and model are open-sourced on [its github] (there is a colab notebook that you can play with online). You may also enjoy [the introduction blog from Prof. Qiang Liu] and [our Chinese blog (中文版讲解)].

  • Rectified Flow incorporates a unique algorithm called reflow. By gradually straightening the probability flow trajectories, reflow simplifies the couplings between the two distributions. With the simplified couplings and the straightened flows, it transforms the infinite-step probability flow model into a few-step or even one-step model, with pure supervised learning. With my collaborators, we trained fast Stable Diffusion models including one-step InstaFlow ([paper], [model and code]) and few-step PeRFlow ([paper], [model and code]).

  • Some other interesting works with RF:
  • (1) FlowGrad ([paper], [model and code]), where we directly use gradient-based optimization to control the generated images from RF.
  • (2) AdaFlow ([paper], [model and code]), where we adaptively adjust the number of sampling steps depending on the different complexity of different states in flow-based imitation learning. It accelerates the action generation by a lot!
  • (3) Point Straight Flow ([paper], [code]), where we use RF to achieve fast point cloud generation.

    • Other Directions


    I also have publications on multi-task/multi-objective learning, sampling, quantization, etc. Please refer to the full publication list.
    Plain Academic