Combining Resampled Importance and Projected Solid Angle Samplings for Many Area Light Rendering

During my time at Meta, one junior from the lab by the name of Ishaan Shah got in touch with me. He was working on vison-y problems and was not very much into them, and wanted to explore computer graphics and specifically rendering.

We decided to work together on a project - extending LTCs to layered and multi-lobed BRDFs. This was pretty fun, but also challenging since neither of us had any experience with layered materials. We lingered around this problem for about 4 months, before finally deciding to shelf it and work on another extension of LTCs – efficient rendering with LTC and many area lights.

For this exploration, we did a lot of experiments with Light BVH and other data structures – it turned out finally that ray tracing with RTX hardware is fast enough, that its not worth the extra cost of querying such data structures.
These explorations led us to create the “best” baseline for direct lighting with ray tracing – combining RIS from ReSTIR (NVIDIA’s work) with projected solid angle sampling (Christoph Peter’s work from his time at KIT).

Turns out, we can very easily and directly combine the two approaches and improve convergence with each sample – however, the convergence with time sufffers. Essentially, the compute increases by a lot.
This began a quest to pinpoint the exact problem and source of high compute - the solution is what our method in this paper is!

In a curious turn of events, Ishaan has now shifted to our group with Prof. PJN. He has now abandoned working with me and is working on efficient glint rendering with Adrien Gruson and Luis Gamboa. Jokes aside, glad to have worked with good students like him.