CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction
Bing He, Jingnan Gao, Yunuo Chen, Zhengxue Cheng, Rong Xie, Li Song, Wenjun Zhang
Shanghai Jiao Tong University
Published October 9, 2026 · on the feed October 22, 2026 · ▲ 34 · arXiv 2610.13177
The abstract
Reconstructing 3D scenes from sparse views without per-scene optimization remains highly challenging, especially for recovering accurate geometry and fine textures. While recent feedforward approaches leverage generalizable 3D Gaussian Splatting (3DGS) for scene generation, they typically assign one or multiple Gaussians to each pixel. Such uniform allocation not only produces highly redundant representations by wasting primitives in homogeneous areas, but also fails to exploit the inherent geometric priors of planar surfaces, often resulting in sub-optimal structural modeling. To address these limitations, we present CompactSplat, a novel feedforward framework that enables a content-aware allocation of 3D Gaussians. By integrating texture-guided spatial partitioning with hierarchical sampling, CompactSplat adaptively distributes primitives - concentrating dense Gaussians in complex regions while tiling flat areas with expanded, scale-modulated primitives. Moreover, our framework inherently supports dynamic sparsity control at inference time, enabling seamless trade-offs between rendering fidelity and memory footprint without requiring network retraining. Extensive experiments on RealEstate10K, DL3DV, and ScanNet demonstrate that CompactSplat consistently outperforms prior methods on both standard metrics and high-resolution rendering consistency, achieving high-fidelity reconstructions with significantly fewer primitives.
The Brief
2:23A short podcast summary of the paper. Written and voiced by AI.
Transcript
CompactSplat. Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction. By Bing He and six others, from Shanghai Jiao Tong University. Published on the ninth of October. Its code repository is worked out: about one hundred and ten source files, with a readme and a licence.
Rebuilding a 3D scene from a few sparse views, without optimising separately for each scene, remains hard, particularly when the goal is accurate geometry and fine texture. Recent feedforward approaches use generalizable 3D Gaussian Splatting for this, and they typically give each pixel one or more Gaussians. That uniform allocation has two costs. It is highly redundant, spending primitives on homogeneous areas that do not need them. And it ignores the geometric priors that planar surfaces naturally carry, which often leaves the structure modelled less well than it could be.
CompactSplat is a feedforward framework that allocates 3D Gaussians according to content rather than by pixel. Dense clusters of Gaussians go where a region is complex. Flat areas are tiled instead with fewer primitives that are expanded and scale-modulated, so one primitive can cover what many would have covered before. The representation becomes compact where the scene is simple and detailed where it is not, and the allocation follows the content of the image rather than its grid.
Two ingredients produce that distribution: texture-guided spatial partitioning, combined with hierarchical sampling. Together they let the framework adapt where primitives are placed. A further property comes with the design. The framework supports dynamic control of sparsity at inference time, so rendering fidelity can be traded against memory footprint on demand, without retraining the network. One trained model can therefore serve a setting that wants the most faithful render and a setting that wants the smallest memory cost, simply by changing how sparse it is asked to be.
Experiments cover RealEstate10K, DL3DV and ScanNet. Across them, CompactSplat consistently outperforms prior methods, both on standard metrics and on rendering consistency at high resolution. It achieves high-fidelity reconstructions while using significantly fewer primitives. The two halves of the claim go together: better quality, and a smaller representation, rather than one bought at the expense of the other.
The shift is from asking how many Gaussians each pixel deserves to asking how many each region deserves. A blank wall and a cluttered shelf are no longer given the same budget, and the reconstructions reported get both sharper and leaner, with the balance between the two left adjustable after training.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free