A$^2$TG: Adaptive Anisotropic Textured Gaussians for Efficient 3D Scene Representation

TL;DR

A2TG addresses memory inefficiency in textured Gaussian Splatting by assigning adaptive anisotropic textures per Gaussian, guided by a gradient-based control strategy. The approach jointly optimizes per-Gaussian appearance and per-Gaussian texture resolution, enabling non-uniform, detail-aware allocation that aligns with anisotropic footprints and measured image gradients, with texture caps up to $2\times2$–$4\times4$. Key contributions include a gradient-driven Gaussian selection mechanism and an anisotropic texture upscaling rule that significantly reduces memory usage while maintaining rendering fidelity across multiple benchmarks. This yields scalable textured Gaussian representations suitable for real-time 3D scene rendering, with potential extensions to richer primitives and dynamic scenes.

Abstract

Gaussian Splatting has emerged as a powerful representation for high-quality, real-time 3D scene rendering. While recent works extend Gaussians with learnable textures to enrich visual appearance, existing approaches allocate a fixed square texture per primitive, leading to inefficient memory usage and limited adaptability to scene variability. In this paper, we introduce adaptive anisotropic textured Gaussians (A$^2$TG), a novel representation that generalizes textured Gaussians by equipping each primitive with an anisotropic texture. Our method employs a gradient-guided adaptive rule to jointly determine texture resolution and aspect ratio, enabling non-uniform, detail-aware allocation that aligns with the anisotropic nature of Gaussian splats. This design significantly improves texture efficiency, reducing memory consumption while enhancing image quality. Experiments on multiple benchmark datasets demonstrate that A TG consistently outperforms fixed-texture Gaussian Splatting methods, achieving comparable rendering fidelity with substantially lower memory requirements.

Figures (6)

• Figure 1: Overview of gradient-based adaptive texture control. Given an initial 2DGS model, (a) our system first optimize the parameters of the 2D Gassians and their textures. (b) Next, we compute the positional gradient of the textured 2D Gassians (as depicted in gray blocks) and select the 2D Gassians that need to increase the resolution of the texture to gain more details. (c) Finally, we adaptively upscale the textures according to the anisotropy of the Gaussians.
• Figure 2: Comparison of 2DGS, A2TG and Textured Gaussians* on the DeepBlending datasets. Left: PSNR versus memory size (MB). Right: memory size (MB) versus point count. A2TG achieves higher reconstruction quality under the same memory budget and requires less memory than Textured Gaussians for the same number of Gaussians.
• Figure 3: Qualitative comparisons. We show the qualitative comparisons of 2DGS* and Textured Gaussians* with A2TG under fixed memory constraint from the Mip-NeRF360 datasets and the DeepBlending datasets. With textures, both Textured Gaussians* and A2TG reconstruct fine scene details, whereas A2TG uses less memory.
• Figure 4: The percentage and distribution. This figure shows the percentage and distribution of texture resolution produced by the adaptive texture upscaling on the scene Garden from Mip-Nerf360 dataset. Gaussians highlighted in blue have square texture of $2 \times 2$ and $4 \times 4$, and those highlighted in red have non-square texture resolution.
• Figure 5: Qualitative visualization of what the adaptive textures learn.Left: full rendering from A$^2$TG. Middle: rendering without textures (RGB textures set to zero, alpha textures set to one). Right: rendering without SH base color. The comparison, visualized on two scenes shows that textures capture high-frequency residual appearance such as foliage structure and fabric detail, while SH color provides smooth, low-frequency shading. Together, they produce the final photorealistic result.