Rendering is a complex process. Its differentiation cannot be uniquely defined; thus, a straightforward integration into neural networks is impossible.
Differentiable rendering (DR) constitutes a family of techniques that tackle such integration for end-to-end optimization by obtaining useful gradients of the rendering process.
Nvidia and Aalto University introduce a modular primitive to provide high-performance primitive operations for rasterization-based differentiable rendering. The proposed modular primitive uses highly optimized hardware graphics pipelines to deliver better performance than previous differentiable rendering systems.