Physics · Generative Models · 3D/4D Vision

Chen Mu

Physics undergraduate at Sun Yat-sen University interested in generative models and 3D/4D vision, especially reliable, low-latency reconstruction of dynamic physical states from limited visual evidence.

Institution
Sun Yat-sen University
Program
Physics undergraduate
Location
Shenzhen, China
Period
Sept. 2024–Present
Expected June 2028

Reliable, Low-Latency 3D/4D Reconstruction with Generative Models

I want to build visual models that recover a usable 3D/4D state from sparse observations, generate only where evidence is missing, expose what remains uncertain, and spend additional computation where it produces a measurable correction. My four connected research directions are:

  • Low-latency perception and correction for 4D models: training-time mechanisms that teach models to recognize and repair inconsistencies, together with selective inference-time correction for uncertain regions or events.
  • Time-varying field reconstruction from low-dimensional 2D observations: jointly recovering geometry and motion while distinguishing what the observations determine from what a learned prior supplies.
  • Simulation under explicit accuracy-latency constraints: adapting fidelity, update frequency, and correction depth to uncertainty and the available response time.
  • Plausible generation with an accompanying validation mechanism: checking generated objects or sequences against the input, geometry, and dynamics, then refining only the parts that need correction.

Independent research · 2026

Validation Geometry for Reward Optimization under Proxy Misspecification

A finite validation protocol can miss precisely the proxy errors that an optimizer amplifies. This project develops an optimizer-conditioned geometry for certifying a specified reward-optimization endpoint.

Validation geometry diagram showing optimizer sensitivity decomposed into measured and validation-invisible components, with the invisible projection determining sharp worst-case proxy bias.
Finite validation controls only the component of optimizer sensitivity inside its measured subspace; the orthogonal component determines the sharp undetectable bias.
Core result
For any fixed optimizer endpoint, the centered density ratio relative to the reference distribution is the exact sensitivity direction for proxy overstatement.
Design insight
Average risk, restricted worst-case coverage, and feasible measurement selection require distinct validation-design objectives.
Evidence and scope
Experiments verify sharp kernel witnesses, identifiability transitions, noisy certificates, and transfer failures under shifted or unseen optimizer directions. The result certifies a declared endpoint rather than universal safety.

Independent coursework project · 2026

Diffusion, Flow Matching, and CRH-CFG

I independently completed the publicly available coursework and assignments from CMU 10-799, Diffusion and Flow Matching.

CRH-CFG architecture: a reliability gate chooses between bypass and a constrained search over classifier-free guidance scales, followed by Heun predictor-corrector execution and receding-horizon replanning.
CRH-CFG performs training-free, inference-time search and control over classifier-free guidance scales.
Implemented
DDPM, straight-path Flow Matching, DDIM, Euler and Heun samplers, and conditional classifier-free guidance.
Experiment setting
CelebA 64×64; Modal L40S; independent ResNet-18 evaluator.
Abstract
CRH-CFG adapts classifier-free guidance online through constrained receding-horizon search, reusing frozen conditional and unconditional fields without extra generator evaluations. On CelebA, it modestly improves target adherence at low overhead; preservation and conditional-quality gains are not yet supported.

Engineering disclosure: Codex assisted engineering; I retained research design, interpretation, and validation.

Physics and research engineering projects

Darcy-coupled tumor drug transport

A Mathematica finite-element framework coupling Darcy flow with convection–diffusion–reaction transport for intratumoral drug simulation.

Read project paper (PDF)

Bird-vocalization acoustic features

An automated acoustic feature-extraction workflow developed around bird-vocalization data from the Shenzhen–Hong Kong border.

Read project paper (PDF)

Mathematica-based agent harness

Research engineering for AI-assisted scientific workflows, connecting symbolic computation, experiment execution, and auditable outputs.

Private repository

Latest Blog

View All

Education and skills

Education

Major in Physics
Sun Yat-sen University
Sept. 2024–Present
Expected June 2028 · Shenzhen

Technical skills

  • Python, C++, C, Wolfram Language
  • PyTorch and scientific ML
  • Server administration and Unix-like systems
  • Docker, LaTeX, and BibTeX
  • TOEFL iBT 104; CET-6 596

Selected recognition

First Prize, 2nd “Illuminating the Future” Optics Science Communication Writing Competition

UCAS Education Foundation / Wiley · “Chirped Pulse Amplification: Born from Playing with Fire”

USACO Gold

Spring 2026