gaussian splatting

While point-based representations (e.g., 3D Gaussian splatting) enable real-time novel-view synthesis, they lack compatibility with AR/VR systems and game engines that rely on mesh-based rendering pipelines. Method: We propose the first end-to-end differentiable mesh reconstruction framework, jointly optimizing geometry and appearance. To ensure topological connectivity, we employ constrained Delaunay triangulation; to bridge point-based efficiency and mesh explicitness, we introduce surface consistency regularization and Gaussian splatting projection constraints. Contribution/Results: Our method achieves a +0.69 dB PSNR gain over MiLo (SOTA) on Mip-NeRF360, while accelerating training by 2× and halving GPU memory consumption. It enables high-fidelity real-time rendering and seamless integration into standard graphics pipelines—unifying the computational efficiency of point primitives with the interoperability of explicit meshes.

To address the inefficiency of voxel/implicit representations in neural rendering and the lack of graphics compatibility in explicit primitives (e.g., 3D Gaussians), this paper introduces Triangle Splatting—a novel differentiable rendering framework that employs triangles as the fundamental differentiable primitive for radiance field reconstruction and end-to-end mesh optimization. The method integrates differentiable splatting modeling, GPU-accelerated rasterization, and non-voxel explicit representation, enabling joint geometric and appearance optimization while preserving the computational efficiency of traditional rasterization. On the Mip-NeRF360 benchmark, Triangle Splatting achieves superior visual quality compared to contemporary non-voxel methods: it outperforms Zip-NeRF in PSNR on indoor scenes and attains over 2400 FPS (1280×720) on the Garden scene—significantly improving the speed–quality trade-off in novel-view synthesis.

3D Gaussian Splatting (3DGS) suffers from inherent limitations in modeling sharp edges, planar surfaces, and geometric compactness—manifesting as edge blurring, “float-in-air” artifacts, irregular spatial distribution around surfaces, and reliance on hand-crafted regularization. To address these, we propose constructing geometrically and semantically explicit radiance fields using differentiable 3D smooth convex bodies as primitives. This work pioneers the replacement of Gaussians with smooth convex bodies as fundamental radiance field units, inherently achieving tight surface-aligned distribution without explicit regularization while preserving edge sharpness and volumetric density representation. Technically, our approach integrates differentiable convex body parameterization, multi-view self-supervised optimization, and a custom CUDA rasterizer enabling efficient forward rendering and gradient backpropagation. On benchmarks including Mip-NeRF360, our method achieves up to +0.81 PSNR gain and −0.026 LPIPS reduction, matches 3DGS rendering speed, and significantly reduces primitive count.

Differentiable ray casting on irregularly distributed Gaussian kernels suffers from severe rendering artifacts (e.g., splatter) and low efficiency in novel-view synthesis. Method: We propose a voxelized Gaussian modeling framework that establishes, for the first time, a physically consistent, density-radiance decoupled differentiable ray casting model. Our approach introduces hierarchical voxel slab integration with BVH acceleration for efficient and accurate volumetric rendering; jointly represents full-spectrum color using spherical Gaussians and spherical harmonics; and co-optimizes Gaussian geometry and radiance properties. Contribution/Results: On the Blender dataset, our method achieves real-time inference at 25 FPS with reasonable training efficiency. It significantly outperforms state-of-the-art methods in PSNR and SSIM while effectively suppressing splatter artifacts. This work bridges a critical gap between differentiable ray casting and Gaussian-based scene representation.

To address the challenge of inaccurate surface reconstruction in 3D Gaussian Splatting (3DGS) caused by multi-view geometric inconsistency, this paper proposes 2D Gaussian Splatting: modeling voxels as oriented planar Gaussian disks to achieve geometrically consistent radiance field reconstruction. Methodologically, we introduce ray-Gaussian intersection–driven perspective-correct splatting, jointly optimized with depth distortion regularization and normal consistency loss to enforce geometric and appearance coherence. Leveraging differentiable 2D splatting rendering and multi-view geometric regularization, our approach preserves real-time rendering performance while significantly improving stability in reconstructing thin structures and enhancing surface accuracy. The result is noise-free, high-fidelity geometric reconstructions with fine geometric detail and competitive appearance quality.

This work proposes Gaussian Mesh Rendering (GMR), a novel differentiable rendering framework that integrates the efficient rasterization mechanism of 3D Gaussian splatting with triangle mesh representations. Traditional mesh-based differentiable renderers suffer from high computational costs and non-smooth gradients, hindering efficient optimization under limited memory. GMR addresses these limitations by analytically generating Gaussian primitives for each triangular face, yielding a lightweight, structure-aware renderer. This approach preserves geometric fidelity while producing smoother gradients, significantly improving optimization efficiency and reconstruction quality—particularly under small-batch, low-memory conditions.

This work addresses the inherent conflict in 3D Gaussian Splatting (3DGS) between jointly modeling appearance and geometry, where direct geometric extraction often degrades rendering quality. To resolve this, the authors propose a lightweight decoupling mechanism that assigns each Gaussian an independent geometric opacity parameter and employs an opacity-guided optimization strategy. Leveraging geometric priors from vision foundation models, this approach enables disentangled modeling of appearance and geometry. The method significantly improves both novel-view synthesis quality and geometric reconstruction accuracy across multiple datasets, demonstrating particularly strong performance in complex scenes containing transparent objects.

This work addresses the gradient instability and limited expressiveness of planar rational splines—such as NURBS—in differentiable rendering by introducing a continuous Gaussian field–based differentiable vector rendering framework. The method reformulates rendering as a smooth, differentiable accumulation process by sampling Gaussian kernels over both the curve parameter domain and the interior of closed regions. It is the first approach to jointly support long splines, rational weights, non-uniform knot vectors, and filled closed regions within a unified differentiable pipeline. Compared to conventional analytical rasterization techniques, the proposed framework substantially enhances gradient stability and geometric expressiveness, achieving superior reconstruction quality and robustness in tasks including calligraphy reconstruction, vectorization, and image abstraction with long splines.

This work addresses the limitations of conventional 3D Gaussian Splatting, which relies on heuristic primitive manipulations—such as manually tuned densification strategies—that hinder stable optimization and impede natural exploration of volumetric structure. The authors propose a probabilistic density field–based Gaussian Splatting framework, where Gaussian positions are treated as samples drawn from a learnable volumetric density field. By leveraging a multi-scale hash grid, the method enables end-to-end gradient-based optimization without heuristic interventions. Unbiased gradient estimation and control variates are introduced to reduce variance, allowing probability mass to flow adaptively according to the loss landscape without fragile priors. Experiments demonstrate that the approach achieves state-of-the-art reconstruction quality on the mip-NeRF 360 benchmark while maintaining real-time rendering speeds comparable to the original 3DGS.

To address limitations of 3D Gaussian Splatting (3DGS) in training efficiency, dynamic scene modeling, theoretical foundations, and deployment adaptability, this work proposes a systematic enhancement framework. Methodologically, it introduces a feedforward network for single-pass radiance field reconstruction, integrates distributed optimization with lightweight differentiable rendering to reduce training overhead, constructs a 4D Gaussian lattice representation for dynamic scene modeling and real-time evolution, establishes a geometry-optics coupled mathematical analysis model to characterize its convergence properties and expressive capacity as an implicit function approximator, and achieves end-to-end deployment on mobile and VR platforms. Experiments demonstrate a 62% reduction in training time for large-scale scenes, real-time rendering at 23 FPS on iPhone 15, and state-of-the-art reconstruction quality—thereby advancing both the practical applicability and theoretical understanding of 3DGS.