I think it's interesting how graphics is an easy to grasp application of many topics, such as machine learning, probablility, and graphics.
This neural network is surprisingly simple for a transformation that seems to convert from complex elements to simpler ones. How much more efficient is this method of rendering fur as clouds than rendering fur normally?
I also wonder if there are other conversions between physical systems in a similar vein using neural networks. For example could we make a neural network to render cloth or granular material as bumpy textures?
This is the most interesting part for me. Neural network is a tool to explore inherent relationship between things that is hard to explicitly discover. And this is the first time I see it plays out in realistic rendering.