Previous local guiding methods used 3D data structures to model spatial radiance variations but struggled with additional dimensions in the path integral, such as temporal changes in dynamic scenes. Extending these structures to higher dimensions also proves inefficient due to the curse of dimensionality. In this study, we investigate the potential of compact neural representations to model additional scene dimensions efficiently, thereby enhancing the performance of path guiding in specialized rendering applications, such as distributed effects including motion blur. We present an approach that models a higher dimensional spatio-temporal distribution through neural feature decomposition. Additionally, we present a cost-effective approximate with lower-dimensional representation to model only subspace by progressive training strategy. We also investigate the benefits of modeling correlations with the additional dimensions on typical distributed ray tracing scenarios, including the motion blur effect in dynamic scenes, as well as spectral rendering. Experimental results demonstrate the effectiveness of our method in these applications.
Recent methods uses small MLPs to model the neural mixture fields, serving a fast alternative for neural path guiding. In this work, we target distribution rendering applications where the target distribution for path guiding in conditioned on additional domains (e.g., motion blur). We investigate multiple alternatives that condition the network prediction on the additional domains, including learnable feature-grid-based encoding, simple Fourier positional encoding, and a progressive training+rendering strategy, and discuss their effectiveness and performance on specific rendering applications.
In Sec. 6.2 we compared the performance of decomposed feature grids with two alternatives (frequency encoding and 4D hash grids). It turns out that they could perform better with a more reasonable configuration. We therefore updated a more fair experiment (Fig. 4) with refined parameters, the specific parameters we used are included in the revised author version of the paper in this page. See more details here.
@inproceedings{dong2024efficient,
author = {Dong, Honghao and Su, Rui and Wang, Guoping and Li, Sheng},
title = {Efficient Neural Path Guiding with 4D Modeling},
year = {2024},
isbn = {9798400711312},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3680528.3687687},
doi = {10.1145/3680528.3687687},
booktitle = {SIGGRAPH Asia 2024 Conference Papers},
articleno = {21},
numpages = {11},
keywords = {Ray Tracing, Global Illumination, Path Guiding, Neural Networks, Feature Decomposition},
series = {SA '24}
}