Final Report


This is my final report detailing the work I've done for the "Many Lights Sampling in Cycles X" project as part of GSoC 2022. There are two main ways that others can access my work:

  1. I've uploaded patches on the developer portal as differential D14664. This also contains some discussion on the subject.
  2. All my work is committed to the branch soc-2022-many-lights-sampling with relatively detailed commit messages.

At a higher level, the following links may also be of interest:

  1. A list of weekly reports detailing the progress I made that week.
  2. A feedback thread where other people interested in the project could ask questions and report issues. For the sake of posterity, I tried to respond to the bug reports with a description of how I resolved it, or if it was still an ongoing issue.
  3. A set of technical posts describing the motivation of the algorithm, as well as why certain decisions were made. Some parts still need to be updated and/or edited, but it may be a useful resource for beginners to get started contributing to Cycles.


The rewritten Cycles X meant that past GSoC work on the Many Lights Sampling project became outdated and unable to merge with the master branch. As such, the purpose of this project was to update, or reimplement, the code so that it could work with Cycles X. The original goal was to gradually support sampling for the following lights in both surfaces and volumes:

  • Point Lights
  • Spot Lights
  • Area Lights
  • Emissive Triangles (and later, instanced lights)
  • Background Lights

Originally, my plan was to use the original work as a reference, complete the initial implementation, and then debug from there. However, Brecht and Lukas suggested implementing support for features one at a time. I actually liked this method a lot more because it forced me to understand the paper at a deeper level, and in retrospect, it probably saved me a lot of extra debugging time.

These features were first implemented without adaptive splitting in mind because the original idea was not immediately compatible with Cycles' architecture. However, as more bugs were reported, it became apparent that some sort of adaptive splitting would be useful (e.g. when the importance heuristic wasn't accurate enough). Thus, the latter half of my project was related to implementing our version of adaptive splitting. Brecht first suggested the use of resampled importance sampling, which I then implemented using weighted reservoir sampling to save on memory usage. This seems to be effective in the problematic cases mentioned.

So far, the implementation seems to be effective for certain scenes, such as this sample attic scene provided by NVIDIA, slightly adjusted for Blender by Alaska.

After 2 minutes of render time, these are the results:

NVIDIA Attic (Original - 2 Minutes)
NVIDIA Attic (MLS - 2 Minutes)

After 30 minutes of render time, these are the results:

NVIDIA Attic (Original - 30 Minutes)
NVIDIA Attic (MLS - 30 Minutes)

There are still some scenes where the original implementation performs better. However, I generally observed that the MLS implementation would produce better results sample-wise, but the samples themselves were taking too long to complete. I suspect that the MLS implementation will perform much better once the algorithm has been optimized. Furthermore, these implementations should all be GPU compatible, so optimizations should be the only work needed on that end as well.

Future Tasks

Unfortunately, some of the debugging took me longer than expected (and there were also new issues to resolve), so the project is not fully ready to merge into master yet. There are also a few outstanding issues that need some further discussion in order to be resolved. Many of the to-do tasks are listed here, but the main ones are the following:

  • Support for volumes
  • Better heuristics for textured lights
  • More debugging and benchmarking

Although I won't be able to dedicate much time to this project during my fall/spring semesters, I plan to continue working on these tasks in the winter.