Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

Thomas Müller

doi:10.1145/3528223.3530127

Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

2022-07-22

1.1K

OA

AI

DOI：10.1145/3528223.3530127

Original

Original request for help

Save

Abstract

En 中文

Neural graphics primitives, parameterized by fully connected neural networks, can be costly to train and evaluate. We reduce this cost with a versatile new input encoding that permits the use of a smaller network without sacrificing quality, thus significantly reducing the number of floating point and memory access operations: a small neural network is augmented by a multiresolution hash table of trainable feature vectors whose values are optimized through stochastic gradient descent. The multiresolution structure allows the network to disambiguate hash collisions, making for a simple architecture that is trivial to parallelize on modern GPUs. We leverage this parallelism by implementing the whole system using fully-fused CUDA kernels with a focus on minimizing wasted bandwidth and compute operations. We achieve a combined speedup of several orders of magnitude, enabling training of high-quality neural graphics primitives in a matter of seconds, and rendering in tens of milliseconds at a resolution of 1920x1080.

Keywords:

Image Synthesis

Neural Networks

Encodings

Hashing

GPUs

Parallel Computation

Function Approximation

Instant Neural Graphics Primitives with a Multiresolution Hash Encoding

Abstract

Journal

Organization