We present a novel Markov Chain Monte Carlo (MCMC) rendering algorithm that incorporates the second-order derivatives or Hessian information of the path throughput function, by using Hamiltonian dynamics. We show that the algorithm results in traditional Metropolis-Hastings sampling with anisotropic Gaussian mutations.

Tzu-Mao Li, Massachusetts Institute of Technology (MIT)
Jaakko Lehtinen, Aalto University, NVIDIA
Ravi Ramamoorthi, University of California, San Diego
Wenzel Jakob, ETH Zurich
Fredo Durand, Massachusetts Institute of Technology (MIT)