ASAP-Textured Gaussians: Enhancing Textured Gaussians with Adaptive Sampling and Anisotropic Parameterization

Recent advances have equipped 3D Gaussian Splatting with texture parameterizations to capture spatially varying attributes, improving the performance of both appearance modeling and downstream tasks. However, the added texture parameters introduce significant memory efficiency challenges. Rather than proposing new texture formulations, we take a step back to examine the characteristics of existing textured Gaussian methods and identify two key limitations in common: (1) Textures are typically defined in canonical space, leading to inefficient sampling that wastes textures’ capacity on low-contribution regions; and (2) texture parameterization is uniformly assigned across all Gaussians, regardless of their visual complexity, resulting in over-parameterization. In this work, we address these issues through two simple yet effective strategies: adaptive sampling based on the Gaussian density distribution and error-driven anisotropic parameterization that allocates texture resources according to rendering error. Our proposed ASAP Textured Gaussians, short for Adaptive Sampling and Anisotropic Parameterization, significantly improve the quality efficiency tradeoff, achieving high-fidelity rendering with far fewer texture parameters.

💡 Research Summary

The paper “ASAP-Textured Gaussians: Enhancing Textured Gaussians with Adaptive Sampling and Anisotropic Parameterization” addresses a critical efficiency challenge in 3D Gaussian Splatting (3DGS) enhanced with texture parameterizations. While attaching local textures to Gaussians significantly enriches their representational power for modeling appearance, geometry, and materials, it introduces substantial memory overhead, hindering scalability.

Instead of proposing yet another novel texture formulation, the authors identify two fundamental limitations common to existing textured Gaussian methods: 1) Inefficient Sampling: Textures are typically defined in the Gaussian’s canonical (UV) space. This leads to a mismatch, as the Gaussian’s opacity (and thus its contribution to rendering) decays from the center outward, but texture samples are uniformly distributed. Consequently, texture capacity is wasted on peripheral regions with negligible visual contribution. 2) Uniform Over-Parameterization: A fixed, often square, texture resolution is assigned to all Gaussians regardless of their visual complexity. This results in over-parameterizing simple areas and under-parameterizing complex ones.

To solve these issues, the paper introduces two complementary techniques under the name ASAP (Adaptive Sampling and Anisotropic Parameterization).

1. Adaptive Sampling: This technique warps the texture coordinate space to align the sampling density with the Gaussian’s probability density distribution. It introduces warping functions, such as axis-wise or radial Cumulative Distribution Function (CDF) warping, that transform canonical coordinates u to warped coordinates ũ. These functions concentrate texture samples (texels) towards the Gaussian’s center where the rendering contribution is highest, dramatically improving the utilization of each texture map without increasing its size.

2. Error-driven Anisotropic Parameterization: This strategy dynamically allocates texture resources across different Gaussians based on their need. Instead of a fixed resolution, each Gaussian’s texture can grow anisotropically (different resolutions in width and height). The decision for growth is driven by gradient statistics. During training, the average magnitude of texture gradients along each axis (u and v) is accumulated. Periodically, if this average exceeds a threshold, the texture resolution is increased along that specific axis. This ensures that Gaussians in visually complex regions or with large surface support receive higher texture resolution, while simpler areas remain compact.

The proposed ASAP-Textured Gaussians framework is built upon the 2D Gaussian Splatting (2DGS) baseline, which provides a natural tangent plane for texture parameterization. The two techniques are orthogonal and synergistic: adaptive sampling optimizes resource use within each Gaussian, while anisotropic growth optimizes resource allocation between Gaussians across the scene.

Extensive experiments on standard datasets (Mip-NeRF 360, DTU, BlendedMVS) demonstrate that ASAP achieves rendering quality comparable to or better than state-of-the-art textured Gaussian methods (e.g., Textured Gaussians, BBSplat, SuperGaussians) while using a significantly smaller number of texture parameters—often reducing them by 70-90%. This represents a major improvement in the quality-efficiency trade-off. The paper concludes that by addressing these core inefficiencies, ASAP enhances the practicality and scalability of textured Gaussians, making them a more viable and efficient general-purpose 3D representation for high-fidelity rendering and downstream tasks.