gaussian splatting: Unpacking the Latest Breakthroughs in 3D/4D Scene Understanding

Latest 27 papers on gaussian splatting: Aug. 30, 2026

Gaussian Splatting (3DGS) has rapidly emerged as a game-changer in 3D reconstruction and novel view synthesis, offering impressive rendering quality and real-time performance. Yet, as with any rapidly evolving field, challenges persist – from managing massive model sizes and handling dynamic scenes to operating in challenging environments and ensuring geometric accuracy. Recent research, however, reveals a flurry of ingenious solutions pushing the boundaries of what’s possible with Gaussian Splatting.

The Big Idea(s) & Core Innovations

One major theme across recent papers is the pursuit of efficiency and compactness. The ‘model bloat’ problem, where millions of redundant Gaussians inflate memory and computational costs, is a significant hurdle. “Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior” by Wang et al. (National Tsing Hua University, Taiwan, and National Institute of Informatics, Japan) tackles this by replacing the conventional ‘densify-then-prune’ with a synergistic L2 loss and a novel Polarized Opacity Prior (POP). This approach naturally drives important Gaussians to full opacity while suppressing redundant ones, eliminating heuristic opacity resets and leading to significantly faster training with fewer Gaussians.

Complementing this, KISS-GS: 3D Gaussian Splatting Compression Kept Simple from Fraunhofer HHI, Germany, focuses on post-training compression. Their modular pipeline achieves astounding 85x to 319x file size reductions by decoupling compaction, encoding, and adaptation. A key innovation, Parallel Representative Assignment Smoothing (PRAS), exploits the 48 equivalent parameterizations of Gaussian covariances to improve compressibility without altering the rendered scene. For dynamic scenes, QuARC-GS: Quantized Anchored Residual Coding for Compact Dynamic Scene Streaming with Gaussian Splatting by Nguyen et al. (University of Michigan) introduces a quantization-aware 4D scene optimization framework that uses a canonical frame and highly compressed per-frame residuals, achieving up to 11x per-frame storage reduction for streaming Free-Viewpoint Video (FVV).

Robustness in challenging environments and dynamic scenes is another critical area. Underwater reconstruction, plagued by severe light scattering and moving objects, receives significant attention. Gaussian Splatting Underwater: A Controlled Cross-Regime Study by Álvarez-Tuñón and Graßhof highlights a counter-intuitive finding: photometric quality often improves with turbidity while geometric accuracy degrades. Their work suggests simple image restoration pre-passes can outperform complex medium-aware systems for geometry. Building on this, AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction from a collaboration of Chinese universities and research labs, fine-tunes a 3D vision foundation model and employs a hybrid scene representation to disentangle scene geometry from underwater optical effects, achieving state-of-the-art SLAM performance. Taking it further, NemoSplat: Feed-Forward 4D Gaussian Splatting for Media-Aware Underwater Reconstruction by Guo et al. (The Hong Kong University of Science and Technology, China) is the first feed-forward 4DGS framework to jointly estimate camera poses, depth, dynamic masks, and water medium parameters in a single pass, enabling robust reconstruction from uncalibrated underwater videos without per-scene optimization.

Beyond challenging natural environments, dynamic scene understanding and manipulation are advancing rapidly. 4DGS-WAM: Bridging Past and Future with an Object-Centric World Action Model based on 4D Gaussian Splatting by Ma et al. (The Chinese University of Hong Kong, Fudan University, and Shanghai Academy of AI for Science) introduces a world action model that explicitly decomposes dynamic scenes into static background and dynamic object components, enabling efficient future prediction. Meanwhile, LagrangeGS: Non-Conservative Lagrangian System on Dynamic 3D Gaussian Splatting from NTT, Japan, formulates dynamic 3DGS as a non-conservative Lagrangian system for physics-based scene manipulation, enabling stable long-term extrapolation, time reversal, and counterfactual editing without retraining.

Improving geometric accuracy and semantic understanding is also a continuous pursuit. TopoSurfel: Closing the Loop between Gaussian Surfels and Meshes for Surface Reconstruction by Fan et al. (University of Science and Technology of China) bridges discrete Gaussian surfels and continuous meshes through a differentiable proxy mesh, establishing bidirectional supervision for accurate reconstruction in textureless or occluded regions. CoMVS-GS: Collaborative Multi-View Stereo and 3D Gaussian Splatting for Surface Reconstruction from Wuhan University and Hangzhou Dianzi University, integrates Multi-View Stereo (MVS) with 3DGS, using dense MVS point initialization and PatchMatch-3DGS Mutual Supervision for explicit depth-level guidance, leading to more accurate geometry. For sparse views, Seeing the Unseen: Semantic-in-Gaussian for Sparse-View 3D Generalization (SeeU) by Bai et al. (University of Chinese Academy of Sciences) uses semantic-conditioned refinement in Gaussian space with a Cross-view Entropy-Aware module to guide residual updates in uncertain regions, significantly improving sparse-view reconstruction.

Removing unwanted elements and refining reconstructions are crucial for practical applications. Per-View Gaussian Predictions Enable Training-Free Distractor Filtering in Feed-Forward 3DGS by Seo and Heo (Sungkyunkwan University) introduces a training-free method to filter transient objects from feed-forward 3DGS reconstructions by exploiting the native association between predicted Gaussians and input views. For more complex scene editing, CoGeo-GS: Concept-Driven and Geometry-Aware Multi-Object Removal in 3D Scenes by Ni et al. (Southern University of Science and Technology, China) offers a concept-driven framework for multi-object removal using semantic tagging, combining monocular depth priors with diffusion-based refinement for plausible geometry and multi-view consistency. And for refining any 3D representation, FixAnything: 3D-Consistent Rendering Refinement via Video Generative Priors from Carnegie Mellon University leverages a single generalist video diffusion model to produce photorealistic, 3D-consistent outputs from various 3D representations (3DGS, NeRF, meshes, point clouds), requiring minimal training data.

Under the Hood: Models, Datasets, & Benchmarks

These advancements are often enabled by novel data, evaluation methodologies, and foundational models:

• KISS-GS achieves impressive compression using the image-based SOG-XT codec and introduces techniques like PRAS and self-organizing 2D codebooks for spherical harmonics. Code is available at https://fraunhoferhhi.github.io/KISS-GS/.
• Per-View Gaussian Predictions evaluates its training-free distractor filtering across models like DepthSplat, YoNoSplat, ReSplat, and AnySplat, using the RobustNeRF and NeRF On-the-go datasets.
• CoGeo-GS utilizes models like Depth Anything, Grounding DINO, Segment Anything Model (SAM), and InFusion diffusion for its concept-driven multi-object removal. It is benchmarked on Mip-NeRF 360 and SPIn-NeRF.
• The comparative study on AR/MR laboratory objects, Comparative Evaluation of 3D Reconstruction Methods for Immersive Visualization of Laboratory Objects, compares photogrammetry, NeRF-based, Gaussian splatting, and LiDAR methods.
• Cross-Platform Benchmark of Neural 3D Reconstruction for Autonomous Laboratory Robots by Kim et al. (Argonne National Laboratory) systematically benchmarks NeRF and 3DGS on NVIDIA Jetson Orin, RTX 2080, and A100, leveraging the NerfStudio framework (https://github.com/nerfstudio-project/nerfstudio) and COLMAP. It also provides a preliminary assessment of SAM3D.
• 4DGS-WAM uses vision foundation models like SAM 3, DA3, VGGT, and WAFT for segmentation, depth, pose estimation, and optical flow, evaluating on the KITTI-MOT benchmark.
• Gaussian Splatting Underwater performs a cross-regime analysis on SeaSplat, WaterSplatting (https://github.com/water-splatting/water-splatting), and UW-GS (https://github.com/WangHaoran16/UW-GS) on datasets like SeaThru-NeRF Curaçao, SOTRUE, Eiffel Tower deep-sea, and EIVA industrial survey. Their code is at https://github.com/olayasturias/uw3dgs.
• PAGS: Autofocusing Photoacoustic Tomography by Ge et al. (Shanghai Jiao Tong University) uses Gaussian splatting for medical imaging, validating with k-Wave MATLAB toolbox simulations and physical phantom data. Code is available at https://github.com/work-submit/PAGS/.
• PIVOT: A Multi-Trajectory Dataset and Testbed by Raymond introduces a crucial benchmark for evaluating 3DGS and NeRF under realistic deployment conditions, providing sensor-derived measured poses, COLMAP-optimized poses, and flexible intrinsic calibration options. The dataset and code are on Hugging Face (https://huggingface.co/datasets/MaryRaymond/PIVOT/tree/v1.0.0) and GitHub (https://github.com/maryraymond/PIVOT/tree/v1.0.0).
• Physics-Integrated Operator Learning via Gaussian Splatting Representations uses FFGS for continuous representation of physical fields, evaluated on 2D and 3D PDE benchmarks from APEBench. It utilizes the pdequinox library for its neural-operator component.
• FixAnything employs the Wan2.1-I2V-14B video diffusion model and DL3DV-10K dataset, demonstrating its versatility with various 3D representations including the gsplat library for Gaussian Splatting. Project page: https://fix-anything.github.io.
• AquaFlow fine-tuned MASt3R on extensive underwater datasets (sweet-corals: https://huggingface.co/datasets/wildflow/sweet-corals, FLSea, TartanAir-Ocean, MIMIR-UW, UW-Stereo-VI) and established a comprehensive underwater evaluation benchmark with 62 diverse real-world sequences.
• NemoSplat introduces a large-scale dynamic underwater dataset with 256 training and 20 evaluation scenes.
• 4DAnyone builds the MVGameHuman dataset (38k videos, 318 actors, 24 cameras) and is benchmarked on DNA-Rendering and DyMVHumans. Code and project page are at https://4danyone.github.io.
• M3ISR: A Multi-Modal Multi-View Benchmark by Liu et al. (The University of Hong Kong) provides a controlled synthetic benchmark with 25 scenes, 6 synchronized views, and ground-truth annotations for 3D/4DGS synthesis, streaming, and compression, accessible at https://arxiv.org/pdf/2608.22465.
• Depth Anything V4 introduces Riemannian Flow Matching for 4DGS, training on WebVid-10M and MegaDepth datasets, and evaluating on Dynamic-Sim, KITTI Dynamic, ETH3D, and ScanNet. Code: https://github.com/fjm123456789/Depth-Anything-4.
• QuARC-GS is evaluated on N3DV (Neural 3D Video) and Meeting Room datasets, with code available at https://github.com/high-performance-computational-optics/QuARC-GS.
• LagrangeGS is tested on Dynamic Object and Dynamic Indoor Scene datasets and achieves low forward-backward cycle errors.
• TopoSurfel uses DTU, Tanks and Temples (TNT), Mip-NeRF 360, and NeRF-Synthetic datasets. Code is available at https://github.com/Fan-Treasure/TopoSurfel.
• CoMVS-GS is evaluated on DTU, GauU-Scene V2, and MatrixCity datasets, utilizing OpenMVS 2.3.0 (https://cdcseacave.github.io/openMVS).
• Point-Based 3D Reconstruction from Sparse Views under Known Illumination utilizes NeRF (LEGO), Blendkit (Horse, Plant), Stanford 3D Scanning Repository (Dragon), and Stanford range image meshes (Teapot).
• GaussVid: Sparse-View Gaussian Splatting with 3D-Aware Video Diffusion Priors builds a large-scale paired dataset (5,000 clips) and leverages Wan2.1-VACE-14B/I2V-14B models, evaluated on DL3DV (https://dl3dv.github.io/) and Zip-NeRF. Code: https://github.com/Xinhui-99/GaussVid.
• Towards Alias-Free 4D Gaussian Representations with Motion-Aware Filtering uses Plenoptic Video (https://novel-view synthesis from multi-view video), D-NeRF, and HyperNeRF datasets. Project page with code: https://maaf-4dgs.github.io/.
• In-Situ Reconstruction of the International Space Station Using 3D Gaussian Splatting and Astrobee leverages existing grayscale images from NASA’s Astrobee free-flying robot and a GLOMAP-based SfM pipeline. This represents a pragmatic application of 3DGS in space.
• DesignAgent3D uses Nerfstudio, Mip-NeRF 360, and Replica datasets, with a GitHub repository mentioned but no URL provided in the paper text.

Impact & The Road Ahead

These advancements in Gaussian Splatting have profound implications across numerous domains. In robotics, more efficient and accurate 3D/4D reconstruction is critical for autonomous navigation and manipulation, especially in complex environments like autonomous laboratories or underwater exploration. The benchmark from Argonne National Laboratory shows that while GS outperforms NeRF, on-device training remains a bottleneck, pushing towards tiered pipelines. The first 3DGS reconstruction of the ISS demonstrates its viability for space habitat mapping, offering crucial tools for mission control and crew training. Medical imaging, with PAGS for photoacoustic tomography, showcases the versatility of GS in scientific applications by improving autofocusing under heterogeneous acoustic media.

For immersive visualization and creative applications, the improvements in compactness, geometric accuracy, and dynamic scene understanding are transformative. Imagine truly interactive virtual spaces where objects can be effortlessly edited, dynamic events accurately simulated, and high-fidelity 4D human reconstructions generated from a single video. The concept-driven editing of CoGeo-GS and the generalist refinement of FixAnything empower creators with unprecedented control. The new benchmarks like PIVOT and M3ISR are essential for driving rigorous evaluation, ensuring that progress is measured against realistic deployment challenges rather than idealized scenarios.

Looking ahead, the explicit nature of Gaussians continues to offer fertile ground for integrating physics, semantics, and efficient representations. The trend toward training-free or minimally trained solutions, leveraging robust priors from large foundation models, is particularly exciting. We can anticipate even more sophisticated methods for handling extreme environments, higher-fidelity dynamic scene capture, and seamless integration of learned generative models with explicit 3D structures. Gaussian Splatting is not just a rendering technique; it’s rapidly becoming a foundational primitive for a new era of explicit, manipulable, and interpretable 3D AI.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.