Scientific intelligence
Learning from telescope signals, medical scans, spectral measurements, and uncertainty instead of clean benchmark inputs.
AI for ScienceRobust learningUncertainty
A working grammar
01
PerceiveVision systems that read noisy, real-world signals.
02
ReconstructStructure, identity, pose, and 3D from incomplete views.
03
CreateGenerative pipelines that turn intent into controllable media.
01 / Research direction
One visual stack.
Many worlds to understand.
My work sits at the intersection of computer vision, scientific discovery, and generative systems. The common thread is simple: make difficult visual evidence legible, controllable, and useful.
3D and embodied vision
Reconstructing people, hands, poses, and objects from the messy viewpoints that real users actually produce.
Digital humansHand trackingVirtual try-on
Generative systems
Designing multi-stage creative workflows where agents, assets, and quality control reinforce each other.
Multi-agent systemsText-to-filmVisual continuity
The throughline
Build the bridge
between evidence and experience.
Whether the input is a radio telescope, an OCT volume, an unconstrained selfie, or a first-person hand video, the hard part is the same: recover the hidden structure without losing the conditions that produced it.
01Signalraw, noisy, incomplete
02Structurepose, identity, geometry
03Actionscience, design, creation
02 / Selected work
Systems I have
put into motion.
Projects are where the research becomes operational: a method, a pipeline, an artifact, and a result someone else can inspect.
In-house reproduction / comparison demo
Featured / 2026Embodied AI
ACE-Ego-Hand
Making egocentric hand motion measurable, trainable, and inspectable.
A reproducible research pipeline for first-person hand reconstruction, covering data conversion, training, evaluation, visualization, and error analysis.
3evaluation
tracks
tracks
20sampled
clips
clips
832width-aware
inference
inference
013D digital humans
Up2you
High-fidelity reconstruction from unconstrained photo collections. The work connects data rectification, pose-correlated feature aggregation, identity preservation, and multi-garment virtual try-on.
02Multi-agent creative systems
Cadence AI
A controllable text-to-film workflow that coordinates world-building, screenwriting, storyboarding, shot generation, quality gates, sound, narration, and post-production.
03AI for Science
FAST RFI suppression
A computer-vision view of radio-frequency interference in the FAST telescope, combining 24 TB scale pretraining, spatial array priors, generalized masking, and log-space regression.
04Medical imaging
Noise2Noise OCT
Unsupervised denoising for stochastic OCT noise through local patch domain transformation, designed for medical images where clean supervision is difficult to obtain.
03 / Publications
A record of
questions pursued.
Selected work across digital humans, scientific imaging, and intelligent sensing.
202620262025202520242024
Up2you: Fast Reconstruction of Yourself from Unconstrained Photo Collections
Digital humans · multi-view reconstruction
SelfieAvatar: High-Fidelity Animatable 3D Head Avatars from In-the-wild Selfies
3D heads · identity preservation
PSCFormer: A Lightweight Hybrid Network for Gas Identification in Electronic Nose System
Electronic nose · spectral learning
AdaNet: A Competitive Adaptive Convolutional Neural Network for Spectral Information Identification
Adaptive convolution · hyperspectral signals
Determination of the Quality of Tea from Different Picking Periods with an Electronic Nose
AI for agriculture · attention mechanisms
Lightweight Residual Convolutional Neural Network for Soybean Classification Combined with Electronic Nose
Edge-friendly sensing · classification
04 / Trajectory
Where the work
has taken me.
From robotics engineering to scientific AI, medical imaging, digital humans, and creative automation.
Cadence AI
FounderBuilding a multi-agent creative system for end-to-end text-to-film generation.
The Hong Kong Polytechnic University
Researcher · Prof. Mingguang He GroupDigital humans, virtual try-on, OCT registration, unsupervised denoising, and multi-view reconstruction.
Shanghai AI Laboratory
Assistant Researcher · AI for ScienceComputer vision and scientific intelligence for astronomy, with work spanning detection, generative models, uncertainty, and symbolic regression.
Northeast Electric Power University
B.Eng. · Robotics EngineeringBuilt a foundation across robotics, perception, deep learning, and applied sensing systems.
05 / Contact
Let's make the
next system legible.
I am interested in research collaborations, ambitious visual intelligence problems, and opportunities where a strong technical system needs a clear point of view.
China / Hong Kong · open to conversations