Candi Zheng | Hong Kong University of Science and Technology

About Me

I am currently a post doc researcher at Hong Kong University of Science and Technology (HKUST). I received my Ph.D. in a joint program between the Department of Mathematics at Hong Kong University of Science and Technology (HKUST) and the Department of Mechanics and Aerospace Engineering at the Southern University of Science and Technology (SUSTech), superivsed by Professor Yang Wang and Professor Shiyi Chen. My academic background includes a B.S. in Physics from SUSTech and a pure math Ph.D. qualification from HKUST.

Research Interests

My research centers on generative diffusion models, with a focus on developing model-agnostic and controllable sampling methods grounded in robust mathematical frameworks and physics-inspired insights. My key areas of interest include:

Model-Agnostic Editing: Developing universal image editing techniques applicable to diffusion models with varying architectures, eliminating the need for retraining. (Ongoing)
Training-Free Fine-Tuning: Creating methods to enable diffusion models to generate new content based on a small set of reference images without requiring training or fine-tuning. (Ongoing)
Model-Agnostic Inpainting: Designing model-agnostic inpainting and conditional sampling methods universally applicable to various diffusion architectures. See our work at LanPaint.
Guidance Methods: Improving generation quality and prompt adherence by integrating multiple diffusion models in a mathematically solid and natural manner. Explore our work at Characteristic Guidance Web UI.

Additionally, I specialize in non-equilibrium physics, with expertise in kinetic theory and rarefied gas dynamics.

Blog Posts

Check my Blog Posts about diffusion model theory.

Publications

Arxiv

Lanpaint: Training-Free Diffusion Inpainting with Exact and Fast Conditional Inference (2024)

Candi Zheng*^, Yuan Lan^, Yang Wang (*Corresponding authors, ^Equal contribution)

Arxiv

PDF Project Page

ICML

Characteristic Guidance: Non-linear Correction for Diffusion Model at Large Guidance Scale (2023)

Candi Zheng*^, Yuan Lan^ (*Corresponding authors, ^Equal contribution)

ICML 2024

PDF Project Page

Arxiv

Phase Transition in Extended Thermodynamics Triggers Sub-shocks (2023)

Candi Zheng*, Yang Wang, Shiyi Chen (*Corresponding authors)

Arxiv Preprint

PDF

Arxiv

Stabilizing the Maximal Entropy Moment Method for Rarefied Gas Dynamics at Single-Precision (2023)

Candi Zheng*, Yang Wang, Shiyi Chen (*Corresponding authors)

Arxiv Preprint

PDF Code

PRE

Data-driven constitutive relation reveals scaling law for hydrodynamic transport coefficients (2023)

Candi Zheng*, Yang Wang, Shiyi Chen (*Corresponding authors)

Physical Review E

PDF

Nanoscale

The prediction of a family group of two-dimensional node-line semimetals (2017)

Jin, Yuan-Jun, Rui Wang, Jin-Zhu Zhao, Yong-Ping Du, Can-Di Zheng, Li-Yong Gan, Jun-Feng Liu, Hu Xu*, and S. Y. Tong* (*Corresponding authors)

Nanoscale

PDF

News

LanPaint

We are excited to share our publicly available extension, the LanPaint, which provides high quality inpaint ability that is universally applicable to Stable Diffusion, FLUX, and HiDream models.

Characteristic Guidance Web UI

We are excited to share our publicly available extension, the Characteristic Guidance Web UI, which provides large CFG (Cassifier-Free Guidance) scale correction for the Stable Diffusion web UI (AUTOMATIC1111).

Moment Gauge

Try our python & JAX library Moment Gauge, designed to facilitate the implementation of numerical solvers using the maximal entropy moment method. Built on the JAX framework, Moment Gauge aims to provide reusable code for researchers and developers working with rarefied gas dynamics and other applications of the maximal entropy moment method.