Xinran (Nicole) Han

I am a final-year Ph.D. candidate at Harvard University, advised by Prof. Todd Zickler. I also work closely with Prof. Ko Nishino and have spent two wonderful summers in Kyoto.

Email: xinranhan [at] g [dot] harvard [dot] edu

Previously, I graduated from the University of Pennsylvania majoring in Mathematics and Computer Science. During my undergrad, I was fortunate to work with Prof. Jianbo Shi and Prof. Dan Roth.

Google Scholar  /  Twitter  /  Github  /  CV

Research

I'm broadly interested in computer vision, generative models, multimodal learning.

Specifically, my research combines data-driven methods with mathematical modeling and insights from the human visual system. My goal is to build generative world models that are robust, ambiguity-aware, and capable of efficient, human-like generalization. Some papers are highlighted.

We introduce derivative representation alignment (dREPA) for image-to-video generation and show it improves subject consistency and leads to better generalization across artistic styles.

We show that a novel pixel-space video diffusion model trained from scratch estimates accurate shape and material from short videos, and also produces diverse shape and material samples for ambiguous input images.

We present a bottom-up, patch-based diffusion model for monocular shape from shading that produces multimodal outputs, similar to multistable perception in humans.

We present new theoretical insight on the equivalence of multi-task and single-task learning for stationary kernels and develop MPHD for model pre-training on heterogeneous domains.

We present a neural model for inferring a curvature field from shading images that is invariant under lighting and texture variations, drawing on perceptual insights and mathematical derivations.

Auxiliary objectives and instruction augmentation improve spatial reasoning in the 'blocks world' task, especially under limited data.

We introduce a task-agnostic image translation model ForkGAN that effectively disentangle domian-specific and domain-invariant information.

Miscellaneous

Outside of research, I enjoy visiting art museums, watching movies and reading about philosophy and psychology.

I'm also using Jon's website template.