TextureSplat: Per-Primitive Texture Mapping for Reflective Gaussian Splatting 2022 Journal: TOG CCS: Computing methodologies Rasterization CCS: Computing methodologies Rasterization CCS: Computing methodologies Point-based models CCS: Computing methodologies Texturing
Mae Younes, Adnane Boukhayma · 2025
Read TextureSplat: Per-Primitive Texture Mapping for Reflective Gaussian Splatting 2022 Journal: TOG CCS: Computing methodologies Rasterization CCS: Computing methodologies Rasterization CCS: Computing methodologies Point-based models CCS: Computing methodologies Texturing by Mae Younes, Adnane Boukhayma (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-13348.txt.
Introduction
Mae Younes and Adnane Boukhayma render shiny surfaces better in twenty twenty-five. Gaussian splatting produces new views of a scene fast and well. Reflective scenes remain hard, because mirror-like surfaces create sharp, high-frequency highlights. Inverse rendering in such settings strains the standard representation.
Their method, TextureSplat, gives each Gaussian primitive its own texture map. Normals and material properties vary across the primitive's local space. That extra representation power targets exactly the specular detail that breaks simpler models. A unified material texture atlas lets the graphics hardware accelerate rendering at test time. The approach is geometrically and physically grounded. Readers in computer graphics get splatting that can handle the chrome and glass of real rooms.
Catalog record
| CREATOR | Mae Younes, Adnane Boukhayma |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-13348
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
TextureSplat: Per-Primitive Texture Mapping for Reflective Gaussian Splatting 2022 Journal: TOG CCS: Computing methodologies Rasterization CCS: Computing methodologies Rasterization CCS: Computing methodologies Point-based models CCS: Computing methodologies Texturing Mae Younes, Adnane Boukhayma Abstract . Gaussian Splatting have demonstrated remarkable novel view synthesis performance at high rendering frame rates. Optimization-based inverse rendering within complex capture scenarios remains however a challenging problem. A particular case is modelling complex surface light interactions for highly reflective scenes, which results in intricate high frequency specular radiance components. We hypothesize that such challenging settings can benefit from increased representation power. We hence propose a method that tackles this issue through a geometrically and physically grounded Gaussian Splatting borne radiance field, where normals and material properties are spatially variable in the primitive’s local space. Using per-primitive texture maps for this purpose, we also propose to harness the GPU hardware to accelerate rendering at test time via unified material texture atlas. Figure 1. Method Overview: We introduce planar primitive material textures — as opposed to single attributes — within physically based Gaussian Splatting rendering optimization. The increased representation power from spatially varying normal and material in object space enables fidelity reconstruction of high frequency specular in highly reflective scenes. Our hardware-accelerated implementation using texture atlases improves rendering efficiency at test time. 1. Introduction 3D reconstruction and inverse rendering from multi-view images are pivotal problems receiving constant interest and investigation from computer vision, graphics and machine learning research communities, with a myriad of direct applications in key industrial domains requiring high-quality 3D modeling and visualization. Neural Radiance Fields (NeRF) ( Mildenhall et al. 2020 ) revolutionized the field by representing scenes as continuous neural implicit functions optimized through differentiable volume rendering. Building on this foundation, 3D Gaussian Splatting (3DGS) ( Kerbl et al. 2023 ) reintroduced point-based graphics by replacing neural networks with explicit 3D Gaussian primitives, achieving both real-time rendering and state-of-the-art quality. These primitives are rendered through volume resampling ( Zwicker et al. 2002 ) , with their parameters optimized via gradient descent-based differentiable rendering. More recently, 2D Gaussian Splatting (2DGS) ( Huang et al. 2024 ) improved multi-view consistency by using planar Gaussian primitives that better align with surfaces. Despite these advances, accurately representing highly reflective surfaces remains challenging. Reflective objects exhibit complex view-dependent effects that depend on surface normals, material properties, and environmental lighting. Recent methods such as RefGaussian ( Yao et al. 2025 ) and 3DGS-DR ( Ye et al. 2024 ) attempt to model these effects using per-Gaussian material properties and physically-based rendering approaches. However, they often struggle with high-frequency specular highlights and sharp reflections. We hypothesise that this is due in part to the inherent resolution limitations of using a single attribute value per primitive. In this paper, we ask the question: How can we enhance the representation power of Gaussian splatting for reflective scenes while maintaining computational efficiency and leveraging hardware acceleration? Our key insight is that the planar nature of 2D Gaussian primitives naturally defines a parameterization that can be exploited for texture mapping, enabling us to store spatially varying material properties per primitive. Inspired by the distinction between Gouraud shading ( Gouraud 1971 ) (constant per-vertex attributes) and Phong shading ( Phong 1975 ) (interpolated attributes) in traditional computer graphics, we introduce per-primitive texture maps for material properties in 2D Gaussian splatting. This approach effectively decouples the geometric representation (Gaussian primitives) from the appearance representation (material textures), allowing us to model high-frequency material variations without increasing the number of primitives. Our method leverages the closed-form ray-splat intersection of 2DGS to accurately map screen-space pixels to local texture coordinates, enabling proper texture filtering. Crucially, we transform tangential normal maps to world space using the primitive’s rotation matrix, analogous to normal mapping in traditional rendering. This enables detailed normal variations across each primitive’s surface, significantly enhancing the rendering of specular highlights and reflections. For efficient rendering after optimization, we pack the primitive textures into atlases that leverage GPU hardware-accelerated texture filtering operations. Our approach is fully compatible with deferred shading pipelines, allowing us to incorporate physically-based rendering models for accurate light interactions. Through experiments on standard benchmarks for reflective scene reconstruction, we demonstrate that our method outperforms state-of-the-art approaches in terms of both quantitative metrics and visual quality. Our method achieves more accurate reflections and sharper specular highlights while maintaining real-time rendering performance. The benefits extend beyond reflective scenes, as our approach improves rendering quality for standard scenes as well. Our contributions include: \bullet A per-primitive texture mapping approach for 2D Gaussian splatting that enhances representation power while maintaining computational efficiency. \bullet Leveraging a normal mapping technique in the context of Gaussian Splatting that significantly improves the quality of specular reflections. \bullet A hardware-accelerated implementation using texture atlases that enhances real-time rendering at test time. \bullet State-of-the-art results on benchmarks for reflective scene reconstruction, demonstrating significant improvements in rendering quality and accuracy. 2. Related Work Neural Radiance Fields ( Mildenhall et al. 2020 ) (NeRFs) have been dominating the 3D shape and appearance modelling recently, based on the astounding success of implicit representations combined with differentiable volume rendering ( Levoy 1988 ; Max 1995 ) . They represent scenes using view-dependent radiance and density fields parameterized by MLPs. When density is modeled as a function of a signed distance field, NeRF variants enable more accurate geometry reconstruction ( Wang et al. 2021 ; Yariv et al. 2021 ; Li et al. 2023 ; Wu et al. 2022 ) . However, multi-scale volume rendering demands frequent MLP evaluations, limiting real-time performance. Grid-based methods ( Müller et al. 2022 ; Sun et al. 2022 ; Fridovich-Keil et al. 2023 ; Fridovich-Keil et al. 2022 ; Chen et al. 2022 ; Sun et al. 2024 ) alleviate this but often struggle with large, unbounded scenes even with level-of-detail grids ( Liu et al. 2020 ) . Gaussian splatting (3DGS) ( Kerbl et al. 2023 ) emerged lately as a strong alternative to NeRFs, offering state-of-the-art novel view synthesis and real time rendering frame rates. It extends the eliptical weighted average (EWA) volume resampling framework ( Zwicker et al. 2002 ; Zwicker et al. 2001 ) to learnable ( Kingma and Ba 2015 ) inverse rendering, modelling scenes with explicit Gaussian kernel primitives, that can be sorted and rasterized efficiently. The 2DGS representation ( Huang et al. 2024 ) leverages planar 2D primitives instead of volumetric ones (3DGS), and performs precise 2D kernel evaluation in object space as opposed to approximative ones in screen space (3DGS), thus leading to superior geometric modelling and multi-view consistency. Both vanilla NeRFs and 3DGS assume low-frequency view dependency. Hence, they can struggle with highly reflective scenes. One strategy to improve in this department is using shading functions that are reflection direction aware ( Verbin et al. 2022 ; Ge et al. 2023 ; Jiang et al. 2024b ) . For instance, Ref-NeRF ( Verbin et al. 2022 ) extends NeRFs with a new parameterization for view-dependent radiance and incorporates normal vector regularization. The shading function can be more physically grounded, and this enables additional application such as relighting and material editing. In this regard, other NeRF and GS based methods proposed to model light interaction using the explicit rendering