Bingliang Zhang

I am an undergraduate student in Yao Class (a special computer science pivot class found by Andrew Chi-Chih Yao) at Tsinghua University. I was visiting Carnegie Mellon University in spring semester of my junior year, where I worked with Jun-Yan Zhu, Eli Shechtman and Richard Zhang.

I am an incoming computer science PhD student at Caltech in CMS, advised by Yang Song. My research interests lie in computer vision especifically generative modeling, neural 3D shape representations and few-shot learning.


profile photo
Ablating Concepts in Text-to-Image Diffusion Models
Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, Jun-Yan Zhu,
arXiv  /  code

An efficient method for ablating previously learnt concepts by providing the natural language description. The concept can be a type of artistic style, an object instance or a specific memorized image.

Multi-concept Customization of Text-to-Image Diffusion
Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, Jun-Yan Zhu,
CVPR, 2023
arXiv  /  code

An efficient method for augmenting existing text-to-image models to learn new concepts with few examples without forgeting learned knowledge. Also we propose a close-form optimization strategy for combining learned weights of multiple concepts.

Continuously Discovering Novel Strategies via Reward-Switching Policy Optimization
Zihan Zhou, Wei Fu, Bingliang Zhang, Yi Wu,
ICLR, 2022

A paradigm to discover diverse strategies in complex RL environments. Our method is able to discover a wide spectrum of strategies in a variety of domains.

Selected Course Project
Cryo2SSNet: 2-Stage Search Supervised Network for Cryo-electron Microscopy Single-Particle Reconstruction
Tsinghua University, Fall 2021

We propose a two-stage Search Supervised Network (Cryo2SSNet), an auto-encoder architecture to reconstruct a 3D density map in a two-stage pipeline. Our method can achieve a high reconstruction accuracy and efficiency.

Open Domain Question Answering with Generative Augmentation
Tsinghua University, Fall 2021

Improved the previous Dense Passage Retriever with Generation-Augmented Retrieval and propose a modified contrastive loss. Our proposed method can improve around 13.1% on top 100 retrieval accuracy.

Intelligent Gomuko: Reinforcement Learning with Advisor
Tsinghua University, Fall 2020

A simplified version of “AlphaGo” on Gomuko, using min-max search data as supervision. It can achieve intelligent human-machine competition and adjustable difficulty.

Website credits: Jon Barron. Thanks to him!