Gaussian Splatting: From Sub-Atomic Particles to Robotic Navigation and Musical Jams
Latest 27 papers on gaussian splatting: Nov. 22, 2026
Gaussian Splatting (3DGS) has rapidly become a cornerstone in neural rendering, offering a potent combination of high-fidelity 3D reconstruction and real-time rendering. This surge in popularity stems from its explicit, point-cloud-like representation, which sidesteps the computational bottlenecks of implicit neural radiance fields. Recent research, as evidenced by a flurry of innovative papers, pushes the boundaries of 3DGS, tackling challenges from capturing extreme lighting conditions and dynamic scenes to enabling on-device applications and even extending to scientific and artistic domains.
The Big Idea(s) & Core Innovations
The central theme across these advancements is leveraging and adapting the core strengths of Gaussian Splatting – its explicit representation and differentiable rasterization – to overcome specific real-world and technical hurdles. For instance, in HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting by Yuanhao Cai et al. from Johns Hopkins University and collaborators, the challenge of high dynamic range (HDR) novel view synthesis is addressed. They introduce Dual Dynamic Range Gaussians and MLP-based tone mappers, achieving a remarkable 1000x inference speedup over previous NeRF-based methods by separating HDR radiance from display transformation. This highlights the power of a physically inspired decomposition for efficient, high-quality rendering.
Similarly, X-Gaussian: Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis by Yuanhao Cai et al. from Johns Hopkins University and HKUST extends 3DGS to X-ray novel view synthesis and direct CT reconstruction. Recognizing the isotropic nature of X-ray attenuation, they replace view-dependent spherical harmonics with a Radiation Intensity Response Function (RIRF) and propose Radiative Gaussian Volume Querying (RGVQ) for direct CT volume reconstruction. This tailored approach dramatically improves accuracy and speed for medical imaging.
Beyond perception, papers like Facet-SLAM: Real-Time Triangle Splatting SLAM by Edward Beng Wai Tan and Siew-Kei Lam from Nanyang Technological University introduce real-time RGB-D SLAM for triangle splatting. By decoupling tracking from mapping and anchoring triangle centroids to fused depth for GICP registration, they achieve over 100 FPS on high-end GPUs and 36 FPS on edge devices, paving the way for robust embodied AI. In a similar vein, GEMOR-GS: Global Extrinsic-calibration-refined Mapping and Online Relocalization for 3D Gaussian-Splatting SLAM by Yuchen Yang et al. from Nanchang University presents a LiDAR–inertial–visual SLAM framework that refines poses and Gaussian maps globally, showing how robust offline reconstruction dramatically improves map quality and enables relocalization after tracking loss.
Addressing the complexity of dynamic scenes, OCLDA-GS: Object-Centric Local Deformation-Aware Gaussian Streaming by Ashim Dahal et al. from the University of Southern Mississippi and Utah introduces an object-centric streaming framework. They use flow-guided anchor allocation and object-constrained motion propagation to improve rendering quality without per-frame Gaussian refinement, significantly enhancing dynamic scene reconstruction. Complementing this, CueSplat: Learning to Schedule Sparse Deformation for Dynamic Gaussian Splatting on Edge Devices by Qingyuan Tang et al. from Shanghai Jiaotong University achieves up to 8.15x speedup on edge devices by predicting Gaussian states from history and selectively applying exact deformation updates, preserving quality while boosting efficiency.
Other notable innovations include Gaussian Splatoms: Direct Atomic Structure Identification in Electron Tomography by Nalini Singh et al. from UC Berkeley, which represents individual atoms as 3D Gaussians to directly reconstruct atomic positions from electron tomography projections. This ground-breaking application in materials science demonstrates the versatility of the Gaussian primitive. For immersive experiences, Where do we Jam? – The Impact of Real and Virtual XR Spaces on Musical Creativity, Co-Presence, Affect, and Flow by Ruben Schlagowski et al. from the University of Augsburg explores the use of Gaussian-splatted point-cloud representations of remote musicians in AR/VR musical jamming, highlighting user preference for varied reality-virtuality continuums.
Under the Hood: Models, Datasets, & Benchmarks
The innovations discussed are often enabled or validated by specialized models, datasets, and rigorous benchmarks:
- HDR-GS: Leverages the multi-view HDR dataset from HDR-NeRF and provides its code at https://github.com/caiyuanhao1998/HDR-GS.
- X-Gaussian: Utilizes datasets like LIDC-IDRI, Open Scientific Visualization Datasets, and cone-beam CT datasets. Official implementation available at https://github.com/caiyuanhao1998/X-Gaussian.
- Facet-SLAM: Validated on TUM RGB-D, Replica, and ScanNet datasets. Code is public at https://github.com/edward62740/Facet-SLAM.
- VLN-AR: Employs a Unity-based simulator to generate a 4-million-sample 3DGS multimodal aerial-navigation dataset and evaluates on AerialVLN and OpenFly. Further details at https://wyz000.github.io/VLN-AR/.
- OCLDA-GS: Evaluated on N3DV, Meeting Room, and Google Immersive datasets. Code is available with the supplementary materials of the paper at https://arxiv.org/pdf/2611.23416.
- CueSplat: Tested on Technicolor, N3DV, and NeRF-DS datasets.
- Gaussian Splatoms: Uses public simulation data and an experimental FePt ADF-STEM tilt-series dataset. Code can be found at https://www.github.com/nalinimsingh/gaussian-splatoms.
- NIRJALA: Introduces SAJALA, a 30-scene Blender Cycles evaluation suite. Paper available at https://arxiv.org/pdf/2611.23450.
- Cross3DGS: Evaluated on Mip-NeRF 360 and Tanks and Temples datasets, with code at https://github.com/unclearness/cross3dgs.
- AbacusGS: Uses Mip-NeRF 360, Tanks and Temples, and Deep Blending. Paper is at https://arxiv.org/pdf/2611.01534.
- ElsaGS: Evaluated on Mip-NeRF360, Tanks and Temples, and DeepBlending. Paper is at https://arxiv.org/pdf/2611.22059.
- Stage-wise Lineage-aware Dropout for Sparse-View 3D Gaussian Splatting: Validated on LLFF, MipNeRF360, and Blender datasets. Paper at https://arxiv.org/pdf/2611.17752.
- MoRE-GS: Evaluated on D-NeRF and HyperNeRF datasets. Project page at https://advanced-vision-and-learning-lab.github.io/MoRE-GS-website/.
- URL-Sim: Validated on Waymo Open Dataset and KITTI-360. Paper at https://arxiv.org/pdf/2611.12932.
- RetinexGS: Uses public RawNeRF scenes and newly captured low-light RAW scenes. Paper at https://arxiv.org/pdf/2611.13280.
- RF3DGS: Introduces a 45-case Unity-rendered 3D reflection removal benchmark. Paper at https://arxiv.org/pdf/2611.13772.
- DepthGS: Evaluated on DTU, Mip-NeRF360, and Tanks and Temples. Paper at https://arxiv.org/pdf/2611.13628.
- FastWild-GS: Uses PhotoTourism and NeRF-OSR datasets. Paper at https://arxiv.org/pdf/2611.13808.
- CP-GS: Evaluated on DIV2K and Kodak24 datasets. Paper at https://arxiv.org/pdf/2611.13784.
- OctreeSplat: Evaluated on Replica, TUM, Tanks and Temples, and DTU. Paper at https://arxiv.org/pdf/2611.02658.
- MacroGS: Evaluated across EPFL, HCInew, HCIold, and INRIA light field datasets. Paper at https://arxiv.org/pdf/2611.17957.
- CoSphere-SLAM: Benchmarked on Holo360D and 360Loc. Project page at https://cosphere-slam.github.io.
- GalaxyGS: Evaluated on Auriga and TNG50 datasets. Paper at https://arxiv.org/pdf/2611.24723.
- RAPID: Uses PANDORA and PISR datasets. Paper at https://arxiv.org/pdf/2611.24404.
Impact & The Road Ahead
These papers collectively illustrate Gaussian Splatting’s growing versatility and impact across diverse fields. From improving medical diagnostics and autonomous driving safety to enabling real-time AR/VR experiences and even pushing the boundaries of scientific imaging, 3DGS is proving to be a highly adaptable and efficient primitive. The focus on efficiency, portability (as seen in Cross3DGS: Unified Vulkan-Based 3D Gaussian Splatting Training from Workstations to Mobile Devices by Hiroyuki Sato and Yusuke Takimoto), and robustness under challenging conditions (e.g., low-light in RetinexGS: Retinex-Inspired Gaussian Splatting for Low-Light RAW Reconstruction by Yulang Fei et al. and underwater in NIRJALA: Draining the Sea out of Gaussian Splatting with Three-Path Optics and Reversible Physics by Latchan Chhetri and Aman Kumar) is particularly exciting.
The road ahead for Gaussian Splatting involves further exploration of its physics-informed applications, enhanced capabilities for dynamic and non-rigid scenes, and more robust deployment on constrained hardware. The drive towards more efficient training and inference, alongside clever regularization techniques (such as Stage-wise Lineage-aware Dropout for Sparse-View 3D Gaussian Splatting by Yan Xing et al.), will undoubtedly unlock even broader applications. As we continue to distill complex scenes and phenomena into these elegant Gaussian primitives, the future of real-time, high-fidelity 3D reconstruction looks brighter than ever.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment