Methods for computing global illumination using Markov Chain Monte Carlo techniques generate a new light path by locally mutating an already generated light path. Such methods can generate light paths carrying large energies in high probabilities, aiming at higher computation efficiency. However, using the mutation strategies proposed in previous methods, when the energy distribution of the light paths is highly localized (i.e., there are spiky peaks), the generated light paths tend to get stuck in such localized regions, thus the computation could be non-robust. To resolve this problem, we propose a new framework for global illumination computation based on a replica exchange method. In our method, we introduce a set of auxiliary path spaces whose localities are gradually relaxed, and apply a replica exchange method to this set of spaces. Using our method, we can mutate light paths through the relaxed path spaces, even if the energy distribution in the original path space is highly localized. Hence, we can robustly generate light paths over the whole path space.
