GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
| Authors | Roni Itkin et al. |
| Year | 2026 |
| HF Upvotes | 21 |
| arXiv | 2604.15284 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
The efficient spatial allocation of primitives serves as the foundation of 3D Gaussian Splatting, as it directly dictates the synergy between representation compactness, reconstruction speed, and rendering fidelity. Previous solutions, whether based on iterative optimization or feed-forward inference, suffer from significant trade-offs between these goals, mainly due to the reliance on local, heuristic-driven allocation strategies that lack global scene awareness. Specifically, current feed-forward methods are largely pixel-aligned or voxel-aligned. By unprojecting pixels into dense, view-aligned primitives, they bake redundancy into the 3D asset. As more input views are added, the representation size increases and global consistency becomes fragile. To this end, we introduce GlobalSplat, a framework built on the principle of align first, decode later. Our approach learns a compact, global, latent scene representation that encodes multi-view input and resolves cross-view correspondences before decoding any explicit 3D geometry. Crucially, this formulation enables compact, globally consistent reconstructions without relying on pretrained pixel-prediction backbones or reusing latent features from dense baselines. Utilizing a coarse-to-fine training curriculum that gradually increases decoded capacity, GlobalSplat natively prevents representation bloat. On RealEstate10K and ACID, our model achieves competitive novel-view synthesis performance while utilizing as few as 16K Gaussians, significantly less than required by dense pipelines, obtaining a light 4MB footprint. Further, GlobalSplat enables significantly faster inference than the baselines, operating under 78 milliseconds in a single forward pass. Project page is available at https://r-itk.github.io/globalsplat/
Engineering Breakdown
Plain English
This paper introduces GlobalSplat, a method for efficiently allocating 3D Gaussian primitives in 3D Gaussian Splatting that improves the trade-off between representation compactness, rendering speed, and visual quality. The core problem is that existing feed-forward and optimization-based methods use local, view-aligned allocation strategies that lack global scene understanding, causing redundancy when multiple input views are processed and making the representation fragile as more views are added. GlobalSplat replaces pixel-aligned or voxel-aligned approaches with a globally-aware primitive allocation strategy that maintains consistency across views while reducing representation size. The result is a more compact 3D asset that renders faster without sacrificing reconstruction fidelity.
Core Technical Contribution
The key novelty is replacing heuristic-driven, locally-aware Gaussian allocation with a globally-aware strategy that understands the entire scene geometry rather than processing each view independently. Instead of unprojecting pixels into dense view-aligned primitives (which causes redundancy), GlobalSplat allocates primitives based on global scene characteristics, eliminating the bloat that accumulates as more input views are added. The architectural innovation shifts from pixel-space or voxel-space alignment to a scene-aware allocation framework that maintains consistency without baking view-specific artifacts into the final 3D representation. This is fundamentally different from prior feed-forward methods that optimize locally and prior iterative optimization methods that are computationally expensive.
How It Works
GlobalSplat takes multiple input views of a scene and instead of independently allocating Gaussians per view (which creates redundancy), it first performs a global analysis of the scene geometry to determine where primitives are actually needed. The method unprojection strategy is replaced with a globally-consistent allocation mechanism that understands which regions of 3D space are geometrically important across all views simultaneously. Gaussian primitives are then placed based on this global understanding, with each primitive positioned to maximize reconstruction quality while minimizing the total count. The rendering pipeline remains standard—splatting these optimally-placed Gaussians to pixels—but because the primitives are globally allocated rather than view-aligned, the representation stays compact and consistent regardless of how many input views are added. The key computational difference is that the allocation step considers the full scene geometry once, rather than making local decisions per view.
Production Impact
For teams building 3D reconstruction pipelines, GlobalSplat significantly reduces the size of trained 3D Gaussian Splatting models—directly lowering memory requirements, disk storage, and transmission bandwidth for 3D assets. Rendering latency improves because fewer primitives mean fewer splat operations per frame, which is critical for real-time applications like VR/AR or game engines where frame rate matters. The reduced model size makes it practical to serve 3D assets on edge devices or mobile platforms where memory is constrained. Integration is relatively straightforward since it replaces the primitive allocation stage of standard 3D Gaussian Splatting pipelines without requiring changes to the rendering backend. The trade-off is that the global allocation requires an upfront scene analysis pass that adds preprocessing time; however, this is a one-time cost that pays dividends in every subsequent render.
Limitations and When Not to Use This
The paper's approach assumes that a single global allocation strategy works well across diverse scene types, but highly varied geometry (both extremely sparse and extremely dense regions in one scene) may not be handled optimally by a single allocation principle. Preprocessing cost and complexity increase compared to simple per-view allocation; if you need extremely fast training or only process single-view scenarios, the global allocation overhead may not be justified. The method likely still requires sufficient input views to infer accurate global scene structure—performance on extremely sparse view settings (e.g., 2-3 views) is not addressed in the abstract. Long-term consistency when scenes have dynamic or temporal components is not discussed, suggesting this approach may be limited to static scene reconstruction.
Research Context
This work builds on the 3D Gaussian Splatting family of methods (which have become dominant for novel view synthesis and 3D reconstruction in recent years) by addressing a core inefficiency: prior feed-forward methods trade off consistency for speed, while iterative optimization methods trade off speed for quality. It likely benchmarks against recent feed-forward Gaussian splatting variants that use pixel-aligned or voxel-aligned allocation, demonstrating improvements in representation size and rendering speed without quality degradation. The broader research direction this opens is moving away from view-centric 3D primitive allocation toward scene-centric allocation strategies, which could inform future work in dynamic scenes, scalable 3D reconstruction, and compact 3D representations. This fits into the growing push to make 3D reconstruction and rendering practical for production systems where memory and latency matter as much as quality.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::