The problem
Traverse Research develops neural material representations and the engineering needed to use them in a renderer. We work across compression, appearance quality, inference and runtime asset handling.
A learned material has to do more than reproduce a reference in isolation. Its feature data and weights must be packaged, loaded and evaluated. Quantization changes both the resource footprint and the reconstruction. The useful comparison includes those costs and their effect on appearance.
What we did in NeuBTF
Our SIGGRAPH Asia 2022 poster investigates two changes to a NeuMIP-based representation: sine activations and concatenation of multi-level samples. The publication evaluates them on captured and synthetic material datasets, and demonstrates integration in Mitsuba 2.
The publication presents material-quality and compression comparisons. Breda’s engine integration extends that work into the rendering pipeline.
Beyond the paper
Breda contains neural-material assets, streaming support, render-graph evaluation and quantized material variants. This lets us investigate the representation as part of an engine’s data flow, not just as an offline model.