Neural Shell Texture Splatting: More Details and Fewer Primitives

Xin Zhang, Anpei Chen, Jincheng Xiong, Pinxuan Dai, Yujun Shen, Weiwei Xu · 2025

Read Neural Shell Texture Splatting: More Details and Fewer Primitives by Xin Zhang, Anpei Chen, Jincheng Xiong, Pinxuan Dai, Yujun Shen, Weiwei Xu (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-2507-20200.txt.

Introduction

Xin Zhang and five co-authors present a way to render 3D scenes with fine surface detail while using far fewer building blocks. Gaussian splatting, a popular rendering method, builds scenes from many small primitives. Its quality is high, but the count of primitives grows large because shape and surface look are tangled together in each one.

The team's answer is a neural shell texture: a single global record of surface appearance that wraps around the shape. The Gaussian primitives then serve two roles, marking the geometry and sampling the texture record as the scene is drawn. With shape and look kept apart in this way, scenes need far fewer primitives for the same detail. Tests showed sharp texture, good use of model size, and easy extraction of textured meshes, the authors report.

Catalog record

CREATORXin Zhang, Anpei Chen, Jincheng Xiong, Pinxuan Dai, Yujun Shen, Weiwei Xu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-20200

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neural Shell Texture Splatting: More Details and Fewer Primitives Xin Zhang, Anpei Chen, Jincheng Xiong, Pinxuan Dai, Yujun Shen, Weiwei Xu Abstract Gaussian splatting techniques have shown promising results in novel view synthesis, achieving high fidelity and efficiency. However, their high reconstruction quality comes at the cost of requiring a large number of primitives. We identify this issue as stemming from the entanglement of geometry and appearance in Gaussian Splatting. To address this, we introduce a neural shell texture, a global representation that encodes texture information around the surface. We use Gaussian primitives as both a geometric representation and texture field samplers, efficiently splatting texture features into image space. Our evaluation demonstrates that this disentanglement enables high parameter efficiency, fine texture detail reconstruction, and easy textured mesh extraction, all while using significantly fewer primitives. 1 Introduction Novel View Synthesis (NVS) generates images from new camera angles that are plausibly consistent with a set of conditioning images, allowing broad applications in virtual reality, robotics, and digital content creation. Among the most influential advancements in this domain is Neural Radiance Fields (NeRF) [ 34 ] , which represents 3D scenes using a neural network for photorealistic novel view generation. Despite its success, NeRF suffers from computational inefficiency, motivating the search for faster alternatives. Recently, 3D Gaussian Splatting (3DGS) [ 26 ] has emerged as a compelling solution, offering real-time, high-quality rendering by leveraging efficient Gaussian primitives. This breakthrough has sparked significant interest in extending 3DGS to dynamic scenes, geometry, anti-aliasing techniques, and generative 3D modeling, pushing the boundaries of NVS toward practical deployment. Specifically, 3DGS represents complex scenes using a set of 3D Gaussian primitives, which are efficiently rendered onto the screen via splatting-based rasterization. Each Gaussian is defined by attributes such as position, size, orientation, opacity, and color, all of which are stored independently and optimized using a multi-view photometric loss to accurately reconstruct scene appearance. A promising follow-up, 2DGS [ 21 ] , replaces the 3D Gaussian representation with 2D oriented planar Gaussian surfels and introduces a backward ray tracing technique for ray-splat intersection and rasterization, leading to more precise geometry modeling. However, the high rendering quality and accurate surface reconstruction of these methods come at the cost of requiring a large number of Gaussian primitives. This is due to the inherent coupling of geometry and appearance in their primitive representation, which requires significant densification in regions with complex textures or detailed geometry. In practice, the frequency of surface textures is often much higher than that of geometry. Therefore, prior GS-like methods suffer from underparameterization for texture and overparameterization for geometry. Figure 1 : Our method achieves enhanced details in rendering by replacing the splat-constant, view-dependent color in Gaussian Splatting with a global shell texture. To resolve these issues, we propose to disentangle geometry and appearance in a novel way: the high-frequency texture is represented and shared across different positions using a global neural shell texture, while geometry can be modeled with 2D surfels for efficiency. The texture and geometry are then separately represented, and they are connected and rendered to images using the ray-splat intersections. This is different from previous works [ 52 , 39 , 9 , 41 ] that employ spatially varying textures for each Gaussian primitive. These methods still associate textures to 3D points and fail to significantly reduce the number of primitives while maintaining high-quality reconstructions, since the higher texture capability provides a highly discontinuous gradient field, leading to unreliable optimization. We adopt Instant-NGP [ 36 ] to model neural shell texture in a canonical representation space, while using Gaussian primitives as explicit geometry samplers. In addition, we model and query geometry using efficient splatting instead of sampling points along the rays like in NeRF. Consequently, our method is able to reconstruct objects with high-frequency texture details using 3 x fewer primitives and improves the rendering results at high-frequency texture areas, as shown in Fig. 1 . In summary, our contributions include: • We propose Ne ural S hell T exture Splatting (NeST-Splatting), enabling Gaussian primitives with spatially varying texture colors, which significantly enhances the ability to represent fine details. • We disentangle texture and geometry in Gaussian Splatting, representing scenes with fewer primitives and a more compact implicit appearance model. • Our method naturally mitigates the needle-like artifacts commonly observed in Gaussian-based methods by enabling low-frequency geometry modeling through texture-geometry decoupling. Experiments demonstrate that our method achieves state-of-the-art performance on the NVS task, exhibiting superior rendering quality in texture-rich regions. 2 Related Work Novel View Synthesis. Given multi-view images, novel view synthesis generates images for unseen viewpoints. Previous methods achieve this by reconstructing light fields [ 14 , 29 ] or blending nearby views based on geometry proxies [ 17 , 8 , 18 , 19 , 10 ] . Neural Radiance Fields (NeRF) [ 34 ] revolutionized the field of novel view synthesis by optimizing a neural scene representation using only photometric supervision. 3D coordinates with frequency encoding are fed into multi-layer perceptrons (MLPs) to predict geometric opacity and appearance color, aggregating in differentiable volumetric rendering. Following works improve NeRF on anti-aliasing rendering [ 4 , 5 ] , reflection modeling [ 46 ] , large-scale reconstruction [ 45 , 43 , 50 ] , acceleration with grid features [ 16 , 11 , 35 , 40 ] etc. Instant-NGP [ 35 ] employs multi-resolution hash grid and small MLPs to significantly speedup both the training and rendering of NeRF. The hash grid flats 3D grid features into a 1D array using the corresponding 3D coordinates with a hash projection mapping. It eliminates the redundancy of unused features in space and enables flexible compression rate control by the hash table length. Compact-NGP [ 42 ] reduces storage requirements by using smaller hash tables indexed by learnable probes, at the cost of longer training times. However, real-time rendering remains challenging for NeRF-based methods due to the hundreds of MLP queries needed per pixel. 3D Gaussian Splatting (3DGS) [ 26 ] represents scenes explicitly using anisotropic 3D Gaussian kernels with spherical harmonic (SH) color coefficients. By leveraging tile-based rasterization, it efficiently renders and optimizes Gaussian kernels through pixel-wise primitive blending. 3DGS achieves state-of-the-art novel view synthesis quality while maintaining over 100 FPS, even on low-end GPUs. Mip-Splatting [ 55 ] addresses aliasing issues in 3DGS by applying 3D smoothing and 2D mip filters. 2DGS [ 21 , 13 ] flattens ellipsoidal 3D Gaussian kernels into elliptical disks for more accurate geometry modeling. Other advancements improve rendering quality by using more expressive kernel functions [ 54 , 23 ] or enhanced densification strategies [ 32 , 33 , 57 ] . However, the explicit nature of 3DGS requires storing a large number of per-Gaussian parameters, leading to high storage overhead as the number of points increases. Efforts to reduce model size include quantization techniques [ 37 , 28 ] and replacing SH with hash grid assisted appearance models [ 12 , 28 ] . Another line of work [ 27 , 49 , 56 ] leverages deferred rendering to handle high reflective surfaces, achieving promising results. Our method leverages the best from both worlds, using Gaussian primitives for geometry modeling and a continuous neural field for

Read the complete text