Instant neural graphics primitives with a multiresolution hash encoding

This work speeds up neural graphics primitives, which are scene or signal representations parameterized by fully connected neural networks that are otherwise costly to train and evaluate. The authors introduce a versatile new input encoding that allows the use of a smaller network without sacrificing quality, significantly reducing the number of floating-point and memory-access operations. Concretely, the small neural network is augmented by a multiresolution hash table of trainable feature vectors whose values are optimized through stochastic gradient descent, and the multiresolution structure allows the network to disambiguate hash collisions, producing a simple architecture that is trivial to parallelize on modern GPUs.

The authors exploit this parallelism by implementing the entire system with fully-fused CUDA kernels, focusing on minimizing wasted bandwidth and compute. Together these choices deliver a combined speedup of several orders of magnitude, enabling high-quality neural graphics primitives to be trained in a matter of seconds and rendered in tens of milliseconds at a resolution of 1920x1080. This dramatic acceleration made neural representations far more practical for interactive graphics applications.