The problem
Traverse Research investigates machine learning for rendering, from material representation to denoising and compression. We evaluate the model in the context of the renderer, including data movement, memory and inference cost.
Model quality is only one part of a real-time implementation. Input features have to exist, inference has to be scheduled, and the result has to behave under motion and changing content. A promising offline result can hide costs elsewhere in the frame. We bring those costs into the experiment.
What we do
Our mixed-precision deep-learning framework supports GPU-based training and inference. Its integration with the render graph allows graphics-related techniques to be evaluated in the same environment as the rendering workload. Breda includes dedicated NRC, neural-material, inference and streaming components rather than treating a model as a disconnected offline prototype.
Our work covers neural radiance caching, learned global illumination, neural materials, path-space denoising, texture compression and GPU inference. Two public milestones show the progression: NeuBTF at SIGGRAPH Asia 2022, and Neural Radiance Cache Implementation on Mobile GPU at SIGGRAPH Asia 2025, co-authored with Samsung researchers. Our Arm Neural Graphics Rust bindings provide a further implementation entry point.
Research that reaches the engine
The full path matters: generate training data, choose a representation, train or update the model, package its assets, stream them, and evaluate the network inside the frame. Online techniques such as NRC also bring training into the runtime budget. We investigate those costs together with the image.