Generative and Multimodal AI for the Physical World
Can generative and multimodal models produce and reason about the physical world with geometric and physical consistency — enough to drive perception, simulation, and robotics?
We develop generative and multimodal models that respect the structure of the physical world: 3D-aware and geometry-consistent generation, controllable synthesis, and vision-language reasoning that generalizes across domains. These models produce synthetic data for perception systems, power simulation for autonomous systems and robotics, and increasingly run efficiently at the edge.
What we work on
- 3D-aware and geometry-consistent generative models
- Controllable human and scene generation
- Synthetic data for perception systems
- Vision-language models and multimodal reasoning
- Domain generalization and causal representation learning
- Physical consistency in generative AI
- Simulation for autonomous systems and robotics
- Efficient and edge-deployable generative AI
Selected work
-
TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process
CVPR 2024 -
2D GANs Meet Unsupervised Single-View 3D Reconstruction
ECCV 2022 -
Learning Implicit Functions for Topology-Varying Dense 3D Shape Correspondence
NeurIPS 2020 (Oral) / TPAMI 2023 -
Bridge: Basis-Driven Causal Inference Marries VFMs for Domain Generalization
CVPR 2026 -
OrgFlow: Generative Modeling of Organic Crystal Structures from Molecular Graphs
2026 -
Vision Language Models Cannot Plan, but Can They Formalize?
2025