Gaussian Splatting: Surfing the Wave of Realism, Efficiency, and Control in 3D AI
Latest 41 papers on gaussian splatting: Apr. 25, 2026
Gaussian Splatting (3DGS) has rapidly emerged as a game-changer in 3D AI, offering unprecedented photorealism and real-time rendering capabilities. However, its journey from novelty to ubiquitous tool involves tackling numerous challenges, from handling dynamic scenes and sparse inputs to optimizing for resource constraints and enabling intuitive editing. Recent breakthroughs, synthesized from cutting-edge research, are pushing the boundaries of what’s possible, ushering in an era of more robust, efficient, and controllable 3D worlds. This post dives into the latest innovations that are shaping the future of 3DGS.
The Big Idea(s) & Core Innovations
The core of recent advancements revolves around making 3DGS more adaptable and intelligent. A significant theme is the pursuit of robustness in challenging environments. DualSplat: Robust 3D Gaussian Splatting via Pseudo-Mask Bootstrapping from Reconstruction Failures from Beihang University and Peking University addresses transient objects by converting initial reconstruction failures into explicit pseudo-mask priors, enabling a cleaner second-stage optimization. Similarly, PDF-GS: Progressive Distractor Filtering for Robust 3D Gaussian Splatting by Sungkyunkwan University amplifies the self-filtering property of 3DGS through a multi-phase progressive strategy to suppress distractors like dynamic content or view-dependent elements. For physically challenging scenarios, Dehaze-then-Splat: Generative Dehazing with Physics-Informed 3D Gaussian Splatting for Smoke-Free Novel View Synthesis from Huazhong University of Science and Technology and The Hong Kong University of Science and Technology, proposes a two-stage generative dehazing and physics-informed 3DGS pipeline to achieve multi-view consistency in smoky environments. Further improving robustness, ELoG-GS: Dual-Branch Gaussian Splatting with Luminance-Guided Enhancement for Extreme Low-light 3D Reconstruction from Shanghai Jiao Tong University combines zero-shot restoration and VGGT-based depth estimation with a dual-branch GS architecture for robust reconstruction in extreme low-light, sparse-view settings.
Another major area of innovation is efficiency and scalability. You Only Gaussian Once: Controllable 3D Gaussian Splatting for Ultra-Densely Sampled Scenes by Ke Holdings Inc. introduces YOGO, a system-level framework for deterministic, budget-aware Gaussian growth, making 3DGS production-ready. Gaussians on a Diet: High-Quality Memory-Bounded 3D Gaussian Splatting Training from University of Texas at Arlington and University of Georgia presents a memory-bounded training framework with dynamic growing and pruning, achieving 80% lower peak memory for edge deployment. For accelerated rendering, AdaGScale: Viewpoint-Adaptive Gaussian Scaling in 3D Gaussian Splatting to Reduce Gaussian-Tile Pairs by Korea University speeds up rendering by 13.8x on city-scale scenes by adaptively scaling Gaussians to reduce redundant computations. Pushing efficiency to new frontiers, GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens from The Hebrew University of Jerusalem and Westlake University, along with TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens by NVIDIA, propose feed-forward architectures that achieve ultra-compact representations and real-time inference by decoupling Gaussian prediction from pixel-aligned features, using global scene tokens or learnable Gaussian tokens respectively. GSCompleter: A Distillation-Free Plugin for Metric-Aware 3D Gaussian Splatting Completion in Seconds from East China Normal University, offers rapid scene completion via a novel ‘Generate-then-Register’ paradigm, achieving significant speedups. Extending 2DGS for video, GS-STVSR: Ultra-Efficient Continuous Spatio-Temporal Video Super-Resolution via 2D Gaussian Splatting from University of Science and Technology of China and Huawei Noah’s Ark Lab, achieves state-of-the-art C-STVSR with near-constant inference latency by leveraging the temporal stability of Gaussian covariance parameters.
Control, editability, and specific applications are also seeing rapid innovation. TransSplat: Unbalanced Semantic Transport for Language-Driven 3DGS Editing by Guangdong University of Technology and Peking University formulates language-driven 3DGS editing as an unbalanced semantic transport problem for precise content manipulation. FluSplat: Sparse-View 3D Editing without Test-Time Optimization from Goertek Alpha Labs introduces a fully feedforward sparse-view 3D editing framework that removes per-scene optimization by enforcing cross-view consistency in the image domain. For high-fidelity human avatars, High-Fidelity 3D Gaussian Human Reconstruction via Region-Aware Initialization and Geometric Priors by Sun Yat-sen University integrates SMPL-X priors and region-aware initialization to recover fine details in faces and hands. CLOTH-HUGS: Cloth Aware Human Gaussian Splatting from University of Texas at San Antonio explicitly disentangles body and clothing with physics-guided supervision for realistic garment deformation and real-time rendering. One-shot Compositional 3D Head Avatars with Deformable Hair from Xi’an Jiaotong University goes further, creating compositional 3D head avatars with physically plausible hair dynamics from a single image. SSD-GS: Scattering and Shadow Decomposition for Relightable 3D Gaussian Splatting by Victoria University of Wellington enables photorealistic relighting by decomposing reflectance into diffuse, specular, shadow, and subsurface scattering components. MSGS: Multispectral 3D Gaussian Splatting, also from Victoria University of Wellington, extends 3DGS to multispectral data for wavelength-aware novel view synthesis. To address fine-grained surface details, Neural Gabor Splatting: Enhanced Gaussian Splatting with Neural Gabor for High-frequency Surface Reconstruction by The University of Tokyo augments Gaussians with MLPs and frequency-aware densification. Finally, Instant Colorization of Gaussian Splats from University of Osnabrück offers a 10x speedup for mapping 2D information onto 3DGS scenes, enabling fast relighting, feature enrichment, and segmentation.
New data acquisition and benchmarking are also key enablers. An Object-Centered Data Acquisition Method for 3D Gaussian Splatting using Mobile Phones by Northwestern Polytechnical University facilitates high-quality 3DGS capture with real-time IMU-based guidance. For challenging underwater environments, BALTIC: A Benchmark and Cross-Domain Strategy for 3D Reconstruction Across Air and Underwater Domains Under Varying Illumination from Heriot-Watt University introduces a controlled benchmark, showing 3DGS performs well with simple preprocessing. DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis by University of Technology Sydney and partners, provides a massive dataset of clean/cluttered image pairs to advance distractor-free 3D reconstruction. E3VS-Bench: A Benchmark for Viewpoint-Dependent Active Perception in 3D Gaussian Splatting Scenes from The University of Tokyo and partners, focuses on embodied AI, evaluating agents’ active viewpoint control for question answering in 3DGS scenes.
Under the Hood: Models, Datasets, & Benchmarks
The innovations above are driven by or contribute to a rich ecosystem of tools and resources:
- R3D3 Dataset & R3D2 Diffusion Model: Developed by Zenseact, Linköping University, Chalmers University, and UC Berkeley for realistic 3D asset insertion in autonomous driving simulations, allowing text-to-3D and cross-dataset asset transfer. Code: https://github.com/wljungbergh/r3d2
- Immersion v1.0 Dataset: From Ke Holdings Inc., an ultra-dense indoor dataset with 30K+ multi-sensor frames per scene, pushing evaluation towards absolute physical fidelity. Used to train YOGO. Code: https://jjrcn.github.io/yogo-project-home/
- DF3DV-1K Dataset: A large-scale real-world dataset (1,048 scenes, ~90K images) with paired clean/cluttered views for distractor-free novel view synthesis, introduced by University of Technology Sydney and collaborators. URL: https://arxiv.org/pdf/2604.13416
- ArtifactWorld Training Dataset & Benchmark: Ke Holdings Inc. built a 107.5K paired video dataset and 1,284 manually audited test clips to restore 3DGS artifacts under sparse-view constraints. Code and dataset will be public.
- BALTIC Benchmark: Heriot-Watt University et al. provides 13 datasets for 3D reconstruction across air/underwater domains under varying lighting. URL: https://arxiv.org/pdf/2604.19133
- E3VS-Bench: The University of Tokyo and partners created a benchmark with 99 3DGS scenes and 2,014 question-driven episodes for embodied 3D visual search. Source code and scripts are supplementary materials.
- HY-World 2.0: Tencent Hunyuan Team’s open-source multi-modal world model for 3D generation and reconstruction. Features HY-Pano 2.0, WorldNav, WorldStereo 2.0, WorldMirror 2.0. Code: https://github.com/Tencent-Hunyuan/HY-World-2.0
- RadarSplat++: Integrated into RadarSplat-RIO by Meta Reality Labs and University of Michigan, the first radar rendering pipeline supporting full range-azimuth-Doppler measurements for radar bundle adjustment. URL: https://arxiv.org/pdf/2604.13492
- GaussianFlow SLAM: KAIST’s monocular 3DGS SLAM system leveraging optical flow with closed-form analytic gradients. Code: https://github.com/url-kaist/gaussianflow-slam
- NG-GS: Beijing Jiaotong University and Westlake University’s NeRF-guided 3DGS segmentation framework. Code: https://github.com/BJTU-KD3D/NG-GS
- Habitat-GS: Zhejiang University and partners’ high-fidelity navigation simulator upgrading Habitat-Sim with 3DGS for photorealistic rendering and dynamic gaussian avatars. Code: https://zju3dv.github.io/habitat-gs/
- Instant Colorization of Gaussian Splats: University of Osnabrück’s efficient algorithm for mapping 2D info onto 3DGS scenes. Code: https://github.com/dlieber01/Instant-Colorization-of-Gaussian-Splats
- Neural Gabor Splatting: The University of Tokyo’s enhanced 3DGS with MLPs per Gaussian for high-frequency details. Code: https://github.com/haato-w/neural-gabor-splatting
Impact & The Road Ahead
The collective impact of these advancements is profound. 3DGS is rapidly maturing from a novel rendering technique to a versatile foundation for complex 3D AI applications. We’re seeing more robust reconstruction in challenging real-world conditions (dynamic scenes, low-light, underwater), dramatically improved efficiency for real-time and edge deployment, and powerful editing capabilities driven by natural language or sparse inputs. Applications range from autonomous driving simulations (R3D2), advanced robotics and SLAM (RadarSplat-RIO, GaussianFlow SLAM, ReefMapGS, GGD-SLAM), photorealistic avatar creation (High-Fidelity 3D Gaussian Human Reconstruction, CLOTH-HUGS, One-shot Compositional 3D Head Avatars), to immersive VR experiences (LIVE-GS) and digital content creation. The ability to control and manipulate 3DGS scenes with fine-grained precision (TransSplat, FluSplat) is unlocking new creative workflows. Furthermore, efforts in data acquisition (Object-Centered Data Acquisition) and benchmarking (DF3DV-1K, BALTIC, E3VS-Bench) ensure that progress is measurable and impactful.
The road ahead promises even more exciting developments. We can expect further integration of large language models for intuitive 3D scene understanding and generation (LIVE-GS), more robust handling of complex real-world dynamics, and even more efficient representations that scale to massive environments while maintaining real-time performance. The ongoing efforts to address the inherent challenges of 3DGS, combined with a growing ecosystem of tools and benchmarks, position it as a foundational technology for the next generation of spatial computing and immersive experiences. The future of 3D is bright, and Gaussian Splatting is leading the charge.
Share this content:
Post Comment