The problem
Traverse Research develops neural radiance caching as part of its rendering and machine-learning work. We investigate how learned radiance prediction interacts with path tracing, online training and the frame budget.
A cache is not just a network dispatch. Training samples have to be gathered, inputs encoded, predictions combined with the path, and the model updated as the scene changes. Image quality and stability depend on that full loop. We work on the implementation around the model as well as the model itself.
Running in Evolve
The capture above is neural radiance caching in the Evolve jungle scene, from an internal debug build. It shows the cache queried at the primary vertex and updating as the camera and animation move, which is the part a still image cannot convey: the model is trained and refreshed while the frame is being rendered, not baked ahead of it.
What we do
Breda has a dedicated NRC implementation, with radiance and loss operations, configurable input encoding, latent-grid support and integration into its path-tracing systems. Training and inference are scheduled through the render graph so their resource use can be understood in context.
We investigate learning behaviour, representation, temporal updates and rendering integration against a partner’s workload. Results are reported with the build, hardware and conditions needed to interpret them.
Published with Samsung
Neural Radiance Cache Implementation on Mobile GPU was published in SIGGRAPH Asia 2025 Technical Communications. The authors are Collin Allen of Samsung Advanced Computing Laboratory, Luca Quartesan, Jasper de Winther and Jasper Bekkers. The publication connects mobile NRC implementation with hardware/software performance modelling.