Research

What we work on

Four research threads, woven together by physically grounded modeling and learning from the data those models produce.

Computer Vision

We work at the intersection of geometry, perception, and learning — designing models that reconstruct the world from images and generate it back out. Architectures we build respect the structure of their inputs, with spatial priors drawn from simulation.

  • Reconstruction from sparse observations, including Gaussian-splat flames and plant geometry from images
  • Generative models — diffusion and transformers — conditioned on simulated or structured priors

Synthetic Data

We use our simulators as data engines. The lab studies how synthetic imagery can be made useful — and trustworthy — for downstream systems, with a focus on closing the synthetic-to-real gap in domains where collecting real data is expensive or unsafe.

  • Synthetic agricultural imagery for vision-based farming
  • Synthetic medical imaging, including hand radiograph generation with diffusion models
  • Datasets and benchmarks for AI training and evaluation
  • Methods that quantify and reduce domain gap

Natural Phenomena Modeling

The lab's most visible thread. We design simulators for the structure and behavior of the natural world — both as scientific tools and as production-quality content. Our work spans biology of plants, large-scale environmental physics, and the dynamics of failure and damage.

  • Tree architecture, root–shoot coordination, and stress-induced breaking
  • Wildfire propagation, combustion in wooden structures, flame reconstruction
  • Hurricane and tornado generation, weather and climate dynamics
  • Forest- and ecosystem-scale modeling

Formal Simulation

The methodological backbone connecting the rest. We are interested in mathematical structures for complex systems and in writing simulators we can later learn against — bridging classical procedural modeling and modern differentiable methods.

  • L-systems and procedural geometry for biology and architecture
  • Cosserat rod physics for branches, hair, and other slender structures
  • Differentiable simulation as a learning signal
  • Numerical methods for combustion, fluids, and granular media

Interested in joining?

We welcome motivated MSc and PhD candidates with interests in graphics, simulation, or machine learning.

Get in touch →