A Fresh Approach to Neural Radiance Field Modeling

In this work we present a novel approach to Neural Radiance Field modeling using Julia programming language. Current research is stuck in a cycle of quick prototyping in high-level dynamic languages followed by offloading of performance-critical parts to a lower-level statically typed languages. This creates an entry barried since such systems use several programming languages with complex configuraion systems to interact between them, disallowing transparent usage of types and creating opaque abstractions which increases the complexity and hinders research progress. With Nerf.jl we demonstrate how modelling Neural Radiance Fields can be done using a single high-level stack of technologies in Julia programming language. Nerf.jl implements popular Instant Neural Graphics primitives [1] for Neural Radiance Fields, contains direct kernel programming, integration of GPU kernels with Automatic Differentiation systems and targeting multiple compute backends. Implementation of Nerf.jl is significantly simpler than with other systems and is done in a backend-agnostic manner while delivering real-time performance.