Wanhua Li

I am currently a postdoctoral fellow at Harvard University supervised by Prof. Hanspeter Pfister. Prior to that, I received my Ph.D. from the Department of Automation at Tsinghua University in 2022, advised by Prof. Jiwen Lu, Prof. Jianjiang Feng , and Prof. Jie Zhou. In 2017, I received my B.S. degree in computer science at Sun Yat-sen University, Guangzhou, China.

My research interests mainly include vision-language models, neural rendering, and 3D-aware synthesis.

I am on the academic job market! Please reach out if you're interested in my research and think I would be a good fit for your department!

Email: wanhua [AT] seas [DOT] harvard [DOT] edu

CV  /  Google Scholar  /  Twitter  /  GitHub

(*Equal Contribution, #Corresponding Author)

We propose Momentum-GS, a momentum-based self-distillation framework that significantly improves 3D Gaussian Splatting for large-scale scene reconstruction.

We present 4D LangSplat, an approach to constructing a dynamic 4D language field in evolving scenes, leveraging Multimodal Large Language Models.

We present SocialGPT, a modular framework with greedy segment prompt optimization for social relation reasoning, which attains competitive results while also providing interpretable explanations.

We ground CLIP features into a set of 3D language Gaussians, which attains precise 3D language fields while being 199 × faster than LERF.

To facilitate using pre-trained models in MMT, we propose CLIPTrans, which transfers the multimodal representations of M-CLIP into a multilingual mBART.

We propose an attribute hallucination framework named CLIP-Cluster to narrow the intraclass variance caused by different face attributes for face clustering.

We propose a language-powered paradigm for ordinal regression, which learns the rank concepts from the rich semantic CLIP latent space.

We propose a language modeling framework named Label2Label to model the complex instance-wise attribute relations, which regards each attribute label as a “word” and recovers the label “sentence” based on the masked one.

We propose dynamic facial radiance fields conditioned on the 3D aware reference image features. The facial field can rapidly generalize to novel identities with only 15s clip.

We propose a Frequency-Aware Spatiotemporal Transformer for video inpainting detection, which simultaneously mines the traces of video inpainting from spatial, temporal, and frequency domains.

We propose probabilistic ordinal embeddings to empower the present-day regression methods with the ability of uncertainty estimation.

A Discriminative Sample Meta-Mining strategy is proposed to mine discriminative information from limited positive pairs and sufficient negative samples for kinship verification.

It is the first face clustering method to train on very large-scale graph with 20M nodes, and achieve superior inference results on 12M testing data.

A simpler, faster, and more accurate method for social relation recognition.

We propose BridgeNet for age estimation, which aims to mine the continuous relation between age labels effectively.

Website Template