Non-Exponential Light Transport for Differentiable Rendering of Participating Media

Zhou et al. (2024) highlight the power of non-exponential transport in unified Gaussian representations, but differentiable rendering in participating media is still challenging. Here, the goal is to create a renderer that supports arbitrary transmittance profiles (not just exponential attenuation), allowing for more realistic simulation and gradient-based optimization in inverse rendering tasks. This could be particularly impactful for medical imaging, atmospheric visualization, or any scenario where light interacts with complex, non-homogeneous volumes. The novelty is in marrying the latest physically-based neural representations with efficient, accurate differentiable solvers for real-world inverse and forward rendering problems—potentially enabling new applications in graphics, vision, and scientific computing.

References:

1. Unified Gaussian Primitives for Scene Representation and Rendering. Yang Zhou, Songyin Wu, Ling-Qi Yan (2024). arXiv.org.