NVIDIA Research Toronto AI Lab

Compact Neural Graphics Primitives
with Learned Hash Probing

2University of Toronto
3Adobe Research
SIGGRAPH Asia 2023

Compact neural graphics primitives (Ours) have an inherently small size across a variety of use cases with automatically chosen hyperparameters. In contrast to similarly compressed representations like JPEG for images (top) and masked wavelet representations [Rho et al. 2023] for NeRFs [Mildenhall et al. 2020] (bottom), our representation neither uses quantization nor coding, and hence can be queried without a dedicated decompression step. This is essential for level of detail streaming and working-memory-constrained environments such as video game texture compression. The compression artifacts of our method are easy on the eye: there is less ringing than in JPEG and less blur than in Rho et al. (though more noise). Compact neural graphics primitives are also fast: training is only 1.2-2.6x slower (depending on compression settings) and inference is faster than Instant NGP because our significantly reduced file size fits better into caches.

Size vs. PSNR Pareto curves on the NeRF scene from Figure 1. Our work is able to outperform Instant NGP across the board and performs competitively with masked wavelet representations [Rho et al. 2023].


Up to 4x smaller than Instant NGP at 1.2-2.6x training cost and no inference speed penalty

40-second fast-forward video.

Neural graphics primitives are faster and achieve higher quality when their neural networks are augmented by spatial data structures that hold trainable features arranged in a grid. However, existing feature grids either come with a large memory footprint (dense or factorized grids, trees, and hash tables) or slow performance (index learning and vector quantization). In this paper, we show that a hash table with learned probes has neither disadvantage, resulting in a favorable combination of size and speed. Inference is faster than unprobed hash tables at equal quality while training is only 1.2-2.6x slower, significantly outperforming prior index learning approaches. We arrive at this formulation by casting all feature grids into a common framework: they each correspond to a lookup function that indexes into a table of feature vectors. In this framework, the lookup functions of existing data structures can be combined by simple arithmetic combinations of their indices, resulting in Pareto optimal compression and speed.

Overview of Compact NGP.


Compact Neural Graphics Primitves with Learned Hash Probing

Towaki Takikawa, Thomas Müller, Merlin Nimier-David, Alex Evans, Sanja Fidler, Alec Jacobson, Alexander Keller

description Paper (PDF, 6.2 MB)
description arXiv version
insert_comment BibTeX

Please send feedback and questions to Towaki Takikawa


  title = {Compact Neural Graphics Primitives with Learned Hash Probing}, 
  author = {Takikawa, Towaki and 
            M\"{u}ller, Thomas and 
            Nimier-David, Merlin and 
            Evans, Alex and 
            Fidler, Sanja and 
            Jacobson, Alec and 
            Keller, Alexander}, 
  booktitle = {SIGGRAPH Asia 2023 Conference Papers}, 
  year = {2023}, 


PSNR vs. file size for varying hyperparameters in compressing the Kodak image dataset. We sweep three parameters: the feature codebook size N_f, the index codebook size N_c (curves ranging from 2^12 to 2^20), and the probing range N_p (dashing and dotting). A value of N_p = 1 corresponds to Instant NGP (shown as *) and has no curve because it is invariant under N_c. We see that the optimal curve at a given file size N has a feature codebook size (same-colored *) of roughly N_f = 1/3 N and index codebook size N_c = 2/3 N. Small probing ranges (solid curves) are sufficient for good compression—in-fact optimal for small values of N_c (left side of curves)—but larger probing ranges (dashed and dotted curves) yield further small improvements for large values of N_c (right side of curves) at the cost of increased training time.

PSNR vs. file size for varying hyperparameters in compressing the NeRF Lego digger.

We fit Compact NGP to the 8000x8000px Pluto image. We show that we are able to outperform JPEG on a wide range of quality levels. The qualitative comparisons at equal size (insets) show the visual artifacts exhibited by different methods: while JPEG has color quantization arfitacts, ours appears slightly blurred.

Training and inference time overheads of Compact NGP. The relative training overhead (denoted with nx) is measured with respect Instant NGP.


We would like to thank David Luebke, Karthik Vaidyanathan, and Marco Salvi for useful discussions throughout the project.

The Lego Bulldozer scene of Figure 6 was created by Blendswap user Heinzelnisse. The Pluto image of Figure 8 was created by NASA/Johns Hopkins University Applied Physics Laboratory/Southwest Research Institute/Alex Parker.