FACT-GS: Frequency-Aligned Complexity-Aware Texture Reparameterization for 2D Gaussian Splatting

TL;DR

The paper tackles inefficiencies in texture-based Gaussian splatting by reallocating texture sampling density to match local visual frequency. It introduces FACT-GS, a frequency-aligned, complexity-aware texture reparameterization implemented as a differentiable warp whose Jacobian controls per-texel sampling density. The approach is end-to-end trainable and preserves real-time rendering while yielding sharper high-frequency details, outperforming prior uniform-texture methods across multiple benchmarks and budget settings. The findings suggest broad applicability to other spatially parameterized appearance representations and highlight potential extensions to dynamic scenes and explicit frequency predictors.

Abstract

Realistic scene appearance modeling has advanced rapidly with Gaussian Splatting, which enables real-time, high-quality rendering. Recent advances introduced per-primitive textures that incorporate spatial color variations within each Gaussian, improving their expressiveness. However, texture-based Gaussians parameterize appearance with a uniform per-Gaussian sampling grid, allocating equal sampling density regardless of local visual complexity. This leads to inefficient texture space utilization, where high-frequency regions are under-sampled and smooth regions waste capacity, causing blurred appearance and loss of fine structural detail. We introduce FACT-GS, a Frequency-Aligned Complexity-aware Texture Gaussian Splatting framework that allocates texture sampling density according to local visual frequency. Grounded in adaptive sampling theory, FACT-GS reformulates texture parameterization as a differentiable sampling-density allocation problem, replacing the uniform textures with a learnable frequency-aware allocation strategy implemented via a deformation field whose Jacobian modulates local sampling density. Built on 2D Gaussian Splatting, FACT-GS performs non-uniform sampling on fixed-resolution texture grids, preserving real-time performance while recovering sharper high-frequency details under the same parameter budget.

Figures (6)

• Figure 1: Existing methods for novel view synthesis, such as 2DGS huang20242d, use a spatially constant per-Gaussian appearance, while Textured GS chao2025textured adds per-Gaussian textures but still relies on a uniform sampling grid. This uniform allocation ignores local signal complexity, causing high-frequency details to blur and wasting capacity in flat regions. In contrast, our frequency-aligned texture reparameterization allocates capacity based on visual complexity, preserving sharp details under the same primitive budget.
• Figure 2: Method comparison of Textured GS and FACT-GS. (A) For the $i$-th Gaussian, the ray–Gaussian intersection was firstly computed to obtain texture coordinates $(u,v)$. Textured GS bilinearly sampled from RGB and opacity textures $\mathbf{T}_i^{\text{RGB}}$ and $\mathbf{T}_i^{\alpha}$ (B) using $(u,v)$ to get the texture color $c_i^{\text{tex}}$ and opacity $\alpha_i^{\text{tex}}$ (C), which lead to blurred high-frequency details (edges). Instead, FACT-GS uses a learnable deformation field $\mathbf{D}_i$ to predict a continuous warp $\Phi_i(u,v)$ that allocate sampling density guided by local frequency (D), after which the warped coordinates $\Phi_i(u,v)$ are used to sample from RGB and opacity textures (E), producing the FACT texture color and opacity (F), which has more clear high-frequency details.
• Figure 3: Novel view synthesis comparison under reduced primitive budgets (10% / 1% of default 2DGS). 2DGS huang20242d loses high-frequency appearance, and Textured GS chao2025textured blurs fine detail due to uniform texture sampling. Our method preserves sharp edges, textures, and material details (zoom-ins), achieving significantly higher perceptual fidelity without increasing parameters.
• Figure 4: Qualitative comparison under high primitive budgets ($100\%$ for Lego, $50\%$ for Church). 2D GS huang20242d lacks spatial texture variation, and Textured GS chao2025textured shows blurred high-frequency patterns due to uniform texture sampling. Our method preserves thin structures and fine texture details.
• Figure 5: Per-Gaussian average RGB gradient magnitude. Textured GS huang2024textured concentrates most textures in the low-frequency range ($<10$), indicating insufficient capacity for high-frequency regions. Our method reallocates capacity according to local signal complexity, yielding more high-frequency textures under the same parameter budget.