Gaussian Compression for Precomputed Indirect Illumination

Chao Li

Abstract:

经查阅提供的论文文档，该论文**未包含单独标注的“Abstract”（摘要）章节**。以下是从论文核心部分（引言、方法、结论）提取的、能完整反映论文核心内容的**原文字句拼接**，完全保留原文表述，可直接复制使用：

Precomputed global illumination (GI) techniques, such as light probes, particularly focus on capturing indirect illumination and have gained widespread adoption. However, as the scale of the scenes continues to expand, the demand for storage space and runtime memory for light probes also increases substantially. To address this issue, we propose a novel Gaussian fitting compression technique specifically designed for light field probes, which enables the use of denser samples to describe illumination in complex scenes. The core idea of our method is utilizing low-bit adaptive Gaussian functions to store the latent representation of light probes, enabling parallel and highspeed decompression on the GPU. Additionally, we implement a custom gradient propagation process to replace conventional inference frameworks, like PyTorch, ensuring an exceptional compression speed. At the same time, by constructing a cascaded light field texture in realtime, we avoid the need for baking and storing a large number of redundant light field probes arranged in the form of 3D textures. This approach allows us to achieve further compression of the memory while maintaining high visual quality and rendering speed. Compared to traditional methods based on Principal Component Analysis (PCA), our approach consistently yields superb results across various test scenarios, achieving compression ratios of up to 1:50.

To address the issue of increasing storage and memory demands for light probes, we propose an adaptive compression method based on Gaussian fitting to compress high-dimensional signals in the space. This approach utilizes freely distributed Gaussian functions to fit the probe information within the scene. The parameters of the Gaussian functions can be updated using gradient descent, allowing for the automatic identification of the most suitable arrangement for the current scene. For a given probe to be compressed, we compute the weighted sum of all Gaussian functions influencing that probe as a predicted value for its latent feature, which is then decoded by a lightweight multi-layer perceptron (MLP) to obtain the decompressed high-dimensional lighting data for the probe. Furthermore, to accelerate the entire compression process, we implement a custom Gaussian function inference and backpropagation process using CUDA. We also propose Cascaded Lighting Volume (CLV), a method that is able to interpolate relatively sparse probe data into 3D textures and only compute detailed volumetric lightmap surrounding the current viewport, achieving a balance between shading quality and memory usage.

Our experiments show that compared to traditional PCA-based methods (including global PCA, blockwise PCA (BPCA), windowed blockwise PCA (WBPCA), clustered PCA (CPCA)), our method demonstrates significant improvements in both the compression ratio and decompression quality. We achieve rapid probe data compression and fast real-time decompression, substantially reducing both disk and memory usage for light probe data. We hope our work will assist more performance-constrained devices in rendering high-quality global illumination, and inspire further exploration of neural representations for illumination.