About Me

Hi — people from the Internet. I am Shikun Liu (刘诗昆 in Chinese characters). I am currently an AI Research Scientist at Meta GenAI, based in London, England. My current research focuses on video generation and world models. Previously, I earned my Ph.D. at Imperial College London, where I was part of the Dyson Robotics Lab, working with Andrew J. Davison and Edward Johns.

I am a pixel researcher. I believe that the path to more general and capable machine intelligence lies in a deep and fundamental understanding of vision and perception. My research pursues this direction through two interconnected themes: 1. teaching machines to understand the world from pixels; and 2. enabling them to reason, create and reconstruct worlds with them.

• Visual Representation Learning: I study self-supervised, multi-task, and multi-modal learning techniques that improve the generalisation and robustness of visual systems, which emphasis on neural architecture designs, training objectives, learning efficiency and interpretability. My notable work includes multi-task neural architectures: MTAN; multi-task and auxiliary learning optimisation strategies: MAXL, Auto-Lambda; and multi-modal reasoning frameworks: Prismer.
• Visual Synthesis and Reconstruction: I develop efficient and scalable visual systems that can model and create complex 3D/4D worlds. I have contributed to the development of one of the first real-time neural SLAM systems using neural fields: iMAP; and I designed the foundational components which enable large-scale generative neural rendering: Kaleido and EscherNet.

Additionally, I am a strong advocate for reproducible and open science, for which I have open-sourced and actively maintained my research projects in my personal GitHub.

Apart from research, I especially enjoy arts and design, particularly in architecture, graphics, typography and generative arts, endorsing unobtrusive and minimal design principles. I also occasionally spend time building and collecting bespoke mechanical keyboards.