Awesome

MentisOculi Pytorch Path Tracer

• A very simple and small path tracer written in pytorch meant to be run on the GPU
• Why use pytorch and not some other cuda library or shaders? To enable arbitrary automatic differentiation. And because I can.

Features

• Can trace reflective spheres and open cylinders with flat and checkered materials.
• Energy Redistribution Path Tracing with transitions limited to rays intersecting same initial diffuse material.
• Backwards (camera to light) path tracing
• Ray permutations in the hypercube of random numbers
• Nearly everything done in large batches on the GPU.

Future Directions

• Langevin Metropolis and HMC both use gradients to increase the efficiency of sampling. This paper outlines how to do these for ray tracing.
• This paper demonstrates how to make ray tracing even more differentiable for the benefit of inverse rendering
• Triangles and a GPU tailored ray acceleration datastructure.
• Metropolis SPPM? With HMC?

Credits

• While the code has been significantly morphed, it was originally a fork James Bowmans' python raytracer
• This was inspired by my ongoing work on secure differentiable programming, specifically adversarial examples in neural networks, at the ETH SRI Lab.