A year of strand-rendering work
2024–2025 · Rendering research. We worked on hair-strand rendering for a platform holder inside O3DE, the Open 3D Engine, investigating the representation and execution strategies that make large strand workloads practical.
Working in an open engine meant the research had somewhere concrete to live. Strand rendering had to fit O3DE’s existing render pipeline and asset flow rather than sit in a standalone harness, so every result came with the integration cost attached.
Why strands are difficult
Hair combines very fine geometry, overlapping contributions and irregular visibility. How strands are grouped, and how their work is distributed across the GPU, can matter more than the shading calculation itself.
Capsule clustering and execution strategies
The May 2025 experiments compared global-pixel and groupshared approaches to capsule clustering. A test with 16,000 clusters and 100 lines per cluster recorded 17.7 ms for the global-pixel path and 19.5 ms for the groupshared path.
That gap is the useful part: the two paths compute the same image, so the difference is entirely in how the work was organised.
What it led to
On the strength of the strand research, the client introduced us to its device benchmarking team, turning one rendering engagement into a second conversation about hardware evaluation.