A hybrid image representation that mixes 2D Gaussians with feature grids, pitched as GPU-native dataset storage.
The idea: Gaussian-Grid splits computation between 2D Gaussian distributions and their attached multi-resolution feature grids, based on local frequency content, so the architecture adapts to the complexity of each image.
Storage use: The paper shows it as GPU-native dataset storage, with 1.24-2.14x faster decoding and 23x-176x less PCIe data traffic than JPEG + ZIP archiving.
Claims: It keeps memory footprints small while reaching high reconstruction fidelity across a variety of high-resolution natural image datasets.
Sign in to comment.
TL;DR of the paper, generated automatically.
A hybrid image representation that mixes 2D Gaussians with feature grids, pitched as GPU-native dataset storage.