I’ve been experimenting with global illumination techniques, and instead of the typical explosive movie intro, I decided to share some images for motivation.

Global illumination is the goal we’ve chased for years. We’ve evolved from using light maps (see image below) to techniques like ray tracing and screen-space ambient occlusion (SSAO). Each approach has its benefits and trade-offs. For example, SSAO became standard in games after its debut by Vladimir Kajalin of Crytek, yet it only provides localized, screen-space shading.

Light probes, however, offer a compact, efficient solution for capturing directional lighting in 3D space. By leveraging spherical harmonics—using as few as 9 numbers for a base-3 function—we can record light from all directions much more efficiently than with cubemaps.

Consider the difference in storage and compute complexity: a low-resolution cubemap might use 6144 pixels at 32×32 per face, which is far more data than a simple sh3 representation. At higher fidelities, while cubemaps can be sampled analytically, spherical harmonics remain constant in their cost, using a fixed number of operations per sample.

This post outlines my journey with tetrahedral light probe meshes. In modern engines, we use hundreds or even thousands of probes to capture the light data. For instance, in Infamous, around 468,000 probes were used worldwide, resulting in massive savings compared to light maps for UV data alone.

The challenge was generating the tetrahedral mesh—the process of connecting nearby probes into little pyramids. Despite searching for existing JavaScript libraries or trying to port solutions like Tetgen and CGAL, nothing fit well. In the end, I developed a state-of-the-art incremental generator in about 16Kb of uncompressed JS, capable of generating a 100,000 point mesh in around 2 seconds.

Below are several screenshots from my testing environment, showing different probe distributions and the performance of the mesh generation and baking process:

[Image: Light Probe Mesh Test 1]
[source: referenced link]

[Image: Light Probe Mesh Test 2]
[source: referenced link]

For example, a tetrahedral mesh with 100,000 points (669,340 tets) built in 2057.6ms, and baking 100,000 probes took 294668ms (roughly 2.95ms per probe). In another test on the Sponza scene, 10,080 probes were baked, averaging 5.87ms per probe.

The next goal is to move the baking process over to ray tracing to capture multiple light bounces and possibly perform baking within workers. I’m also exploring using the probe data at runtime more effectively—for instance, sampling across tetrahedral meshes to get smooth transitions and avoiding contributions from occluded probes. Automatic probe placement is another interesting challenge, allowing denser placement near light/color transitions and sparser elsewhere.

It’s been a long journey—from reading technical papers to tackling the challenges of tetrahedral mesh generation in JavaScript. I hope you find this exploration into efficient global illumination techniques as fascinating as I do!

Credits:
• Thanks to @donmccurdy for his excellent work that served as a base for my light probe experiments.
• Célestin Marot for his insights into tetrahedral mesh generation.
• Michał Iwanicki and Peter-Pike Sloan for their comprehensive presentations and supplementary material.