Over the past few years, I have delved in several fascinating research topics, including generative models, adversarial examiners, and 3D computer vision. My journey began with robotic manipulation using reinforcement learning and 3D computer vision. Recognizing robustness as a critical challenge for real-world applications of computer vision and other AI systems, I turned my focus to integrating human knowledge, graph-based models, and generative models to enhance both performance and robustness of vision systems. These efforts led to the development of adversarial examiner, which I further applied to generative models to systematically uncover and analyze their failure modes.

Recognizing that generative and discriminative models can enhance each other—through methods including adversarial examiners—and the relatively underexplored potential of generative models, I also focused on improving generative models for image, 3D, and video generation. For example, DiMR progressively refines details from low to high resolution, significantly reducing distortions and enhancing visual fidelity in image generation. CrossFlow introduces a novel approach for T2I generation by directly mapping between modalities with flow matching, bypassing the traditional reliance on Gaussian noise. DIRECT-3D establishes a new training paradigm for 3D generative models, enabling effective learning from large-scale, noisy, and unaligned 3D data sourced from the Internet. ReVision utilizes explicit 3D knowledge to improve video generation, particularly for generating complex motions.

I have also worked on topics including segmentation, video understanding, foundational object models, and 3D datasets during my internships and collaborations.

In short, during my PhD studies, I have been the core contributor to several projects that align closely with my research interests: