Captured in 2021 with our MonoStage algorithm.
Qing Shuai 帅青
I am a researcher in computer vision, with a focus on the multimodal understanding and generation of 3D content. My work spans reconstruction, animation, and generation of humans and scenes, with an emphasis on bringing these capabilities from research prototypes to real applications.
Since 2024, I have been a Senior Researcher at Tencent Hunyuan, where I work on production-grade 3D human motion generation, conditioned on a range of modalities including text, audio, and video. One recent output from this line of work is HY-Motion.
I received my BSc from Zhejiang University in 2019 and my PhD from the same institution in 2024, advised by Prof. Xiaowei Zhou. During my doctorate I led the development of EasyMocap, an open-source system for markerless motion capture from multi-view video, and published on neural rendering and free-viewpoint video of interacting people.
I have been particularly excited by recent progress in LLMs and agents — AI is evolving from a chatbot into a system that can invoke tools and interact with a computer, and increasingly observe, reason, and act the way a person does. I am interested in equipping such agents with multimodal understanding and generation capabilities, and, more broadly, in how AI can genuinely engage with the physical world.
Experience
- 2024 —
Senior Researcher
Tencent Hunyuan · Shenzhen, China
Main driver of 3D human motion generation from multimodal control inputs, taking the technology from research prototype to production. Along the way, built out a full pipeline spanning data curation, model training, evaluation, and deployment. Also open-sourced part of this work as HY-Motion, which reached the Hugging Face weekly trending list.
- 2019 — 2024
Ph.D. in Computer Science
Zhejiang University · State Key Lab of CAD & CG · Hangzhou, China
Advised by Prof. Xiaowei Zhou. Focused on markerless motion capture and neural scene representations. Designed and maintained EasyMocap, a widely used open-source toolkit and currently the most starred motion capture system on GitHub. Published at CVPR, ICCV, ECCV, and SIGGRAPH — building a solid foundation in computer vision and computer graphics.
- 2015 — 2019
B.Eng. in Mechanical Engineering
Zhejiang University · Chu Kochen Honors College · Hangzhou, China
Coursework centered on modern control theory and the theory of wheeled and legged robots. The engineering mindset — sensing, actuation, and feedback — later shaped how I approach vision and 3D.
Featured projects
A thread runs through my work: how do we turn the physical world and human behavior into something a computer can perceive, model, and act within? The three projects below attack three sides of that question — perceiving people, modeling the world they inhabit, and acting inside it. The long-term goal is an AI that understands and moves through the real world the way we do.
-
EasyMocap
Open-source infrastructure for capturing human data.
Understanding people starts with data. EasyMocap turns a handful of ordinary cameras into a full pipeline for markerless motion capture — calibration, keypoint estimation, and SMPL fitting — bringing 3D human data within reach of anyone with a few cameras.
4.7k stars -
LoG
Real-time interactive photorealistic 3D scenes.
Behavior only makes sense inside a world. LoG (Level-of-Gaussians) proposes an adaptive hierarchical Gaussian representation whose level of detail follows the viewpoint, delivering real-time interaction across scales — from a single object to an entire city.
770 stars -
HY-Motion 1.0
Understanding and generating human motion.
The last step is turning human intent into motion. We curated a large-scale human motion dataset and trained an MMDiT generative model on it, producing 3D human motion from language, audio, and control signals — with strong instruction understanding and fine-grained controllability.
2.5k stars
Publications
Selected papers, most recent first. See Google Scholar for the complete list.
-
![]()
PRISM: Streaming Human Motion Generation with Per-Joint Latent Decomposition
-
![]()
CoMoVi: Co-Generation of 3D Human Motions and Realistic Videos
-
![]()
AnyAct: Towards Human Reenactment of Character Motion From Video
-
![]()
AnchorCrafter: Animate Cyber-Anchors Selling Your Products via Human-Object Interacting Video Generation
-
![]()
HY-Motion 1.0: Scaling Flow Matching Models for Text-To-Motion Generation
-
![]()
Motion-2-to-3: Leveraging 2D Motion Data for 3D Motion Generations
-
![]()
IDOL: Instant Photorealistic 3D Human Creation from a Single Image
-
![]()
Dyn-e: Local Appearance Editing of Dynamic Neural Radiance Fields
-
![]()
Ready-to-React: Online Reaction Policy for Two-Character Interaction Generation
-
![]()
Animatable Implicit Neural Representations for Creating Realistic Avatars from Videos
-
![]()
AniDress: Animatable Loose-Dressed Avatar from Sparse Views Using Garment Rigging Model
-
Reconstructing Close Human Interactions from Multiple Views
-
![]()
Implicit Neural Representations with Structured Latent Codes for Human Body Modeling
-
![]()
Representing Volumetric Videos As Dynamic MLP Maps
-
Learning Analytical Posterior Probability for Human Mesh Recovery
-
![]()
Learning Human Mesh Recovery in 3D Scenes
-
Novel View Synthesis of Human Interactions from Sparse Multi-view Videos
-
![]()
QuickPose: Real-Time Multi-View Multi-Person Pose Estimation in Crowded Scenes
-
![]()
Efficient Neural Radiance Fields for Interactive Free-viewpoint Video
-
![]()
Reconstructing Hand-Held Objects from Monocular Video
-
![]()
Shape Prior Guided Instance Disparity Estimation for 3D Object Detection
-
Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans
-
Reconstructing 3D Human Pose by Watching Humans in the Mirror
-
Animatable Neural Radiance Fields for Modeling Dynamic Human Bodies
-
Motion Capture from Internet Videos