gaussian splatting

3D Gaussian Splatting (3DGS) suffers from limited local modeling capacity due to fixed Gaussian kernels, struggling to simultaneously achieve geometric accuracy under sparse inputs, dynamic scene reconstruction, and high-fidelity novel-view synthesis. To address this, we propose Neural Texture Lattice (NTL), a lightweight neural field built upon triplane encoding and a shared neural decoder, enabling global–local co-adaptive representation. NTL explicitly models viewpoint- and time-dependent appearance and geometry by generating dynamic textures and displacements for each Gaussian primitive. This design enhances representational capacity while suppressing parameter redundancy, significantly improving generalization. Evaluated across diverse reconstruction scenarios—including static/dynamic scenes and sparse/dense inputs—NTL consistently outperforms existing 3DGS variants on multiple benchmarks. It achieves state-of-the-art results in rendering quality, geometric fidelity, and temporal consistency, while maintaining efficient optimization and real-time rendering.

To address the trade-off between high visual fidelity and low latency in VR real-time rendering, this paper proposes a dynamic foveated dual-stream rendering framework grounded in human visual perception. The peripheral region employs streamlined 3D Gaussian splatting for efficient, smooth rendering, while the foveal region—aligned with gaze—is rendered using a lightweight CNN-driven neural point representation to recover high-fidelity details. We pioneer the deep integration of perceptual modeling into the radiance field rendering pipeline, enabling optimal balance between computational load and perceptual quality. Furthermore, we introduce the first foveated dual-stream representation unifying 3D Gaussian splatting and neural points, augmented with point-cloud-based occlusion culling for enhanced efficiency. Evaluated on Quest 3-class hardware, our method achieves over 90 FPS, surpassing standard 3D Gaussian Splatting (3DGS) in edge sharpness and texture detail, with a user preference rate of 92%, thereby fulfilling the stringent latency and immersion requirements of interactive VR applications.

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.

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.

To address overfitting in 3D Gaussian Splatting (3DGS) under single-view supervision—leading to artifacts in novel-view synthesis and inaccurate geometric reconstruction—this paper proposes a multi-view collaborative optimization framework. Our method introduces three key innovations: (1) a novel multi-view regularization paradigm that jointly enforces consistency across multiple views; (2) an intrinsic-cross-guided coarse-to-fine training strategy integrating multi-scale geometric and appearance priors; and (3) ray-intersection-driven cross-view densification coupled with view-difference-aware adaptive densification. While preserving real-time rendering performance, our approach significantly improves both novel-view image fidelity and 3D geometric accuracy. Extensive experiments demonstrate strong generalization across diverse scenes and mainstream 3DGS variants, outperforming existing single-view methods in both qualitative and quantitative evaluations.

Existing general-purpose novel view synthesis methods employ fixed allocation strategies for Gaussian primitives, which struggle to adapt to spatial complexity variations across scenes, resulting in redundant resources in smooth regions and insufficient representation in detailed areas. This work proposes SplatWeaver, a framework that introduces dynamic primitive allocation within a feed-forward 3D Gaussian splatting architecture for the first time. By integrating a mixture-of-Gaussians expert model with a pixel-level routing mechanism—guided by high-frequency structural priors and enhanced through routing regularization—the method adaptively allocates Gaussian primitives according to local geometric complexity. Experiments demonstrate that SplatWeaver significantly outperforms state-of-the-art approaches using fewer Gaussians, consistently achieving higher rendering fidelity and improved detail reproduction across diverse scenes.

Existing 3D representations struggle to simultaneously achieve high-quality novel view synthesis and object-level editability. This work proposes MLP-Splatting, which introduces compact, standalone MLPs as local neural primitives to model radiance and opacity, enabling efficient rendering through sparse voxel composition. The method automatically decomposes scenes into semantically coherent objects or parts using only RGB supervision—without requiring segmentation masks—and supports interactive editing, open-vocabulary scene interaction, and instant segmentation. Compared to semantic 3D Gaussian Splatting, MLP-Splatting reduces memory consumption to 1/15 and accelerates rendering by 3×, while preserving high-fidelity synthesis and fine-grained object-level manipulation capabilities.

Existing single-step feedforward networks struggle to regress static Gaussian primitives suitable for all viewing angles in novel view synthesis from pose-free images, limiting reconstruction fidelity. This work proposes a viewpoint-adaptive dynamic Gaussian splatting method that transforms static representations into a view-aware dynamic splatting mechanism. Specifically, a lightweight dynamic MLP predicts residual updates to Gaussian attributes—including position, scale, rotation, opacity, and color—conditioned on the target viewpoint coordinates. By adapting Gaussian parameters dynamically to each novel view, the method significantly enhances synthesis quality while maintaining high computational efficiency, achieving state-of-the-art fidelity with inference at 17 FPS and real-time rendering at 154 FPS.

While Gaussian splatting achieves strong performance in novel view synthesis, its requirement of millions of primitives for highly textured scenes incurs prohibitive storage and computational overhead. This work addresses real-time novel view synthesis for sparse-geometry scenes by proposing *Nexels*: a hybrid representation that decouples geometry (surfels) from appearance—modeled jointly via a global NeRF-style neural field and per-surfel color parameters. To our knowledge, this is the first approach to co-model neural-textured surfels with a fixed-sampling neural field. Rendering employs pixel-level sparse texture sampling, enabling efficient representation without compromising visual fidelity. Experiments demonstrate significant gains: for outdoor scenes, voxel count reduces by 9.7× and memory usage by 5.5×; for indoor scenes, voxel count drops by 31× and memory by 3.7×; rendering speed improves by 2×, while visual quality surpasses existing textured primitive methods.

In 3D Gaussian splatting, opacity-agnostic occlusion culling remains challenging due to the inherent semi-transparency of Gaussians, severely limiting rendering efficiency for complex scenes. To address this, we propose a neural visibility model: a lightweight, shared MLP that learns view-dependent visibility functions per Gaussian primitive, enabling dynamic occlusion culling of semi-transparent Gaussians prior to rasterization. Integrated with a custom-instanced software rasterizer, frustum culling, and Tensor Core acceleration, the pipeline achieves end-to-end optimization. This is the first work to incorporate neural networks into Gaussian splatting occlusion culling, and the first to support view-dependent visibility prediction for semi-transparent primitives—complementary to existing LOD strategies. Experiments demonstrate a 27% reduction in VRAM usage and a 2.1× speedup in rendering, while maintaining state-of-the-art image quality, significantly advancing real-time, high-performance Gaussian splatting rendering.