Research map / 3D vision / Gaussian splatting

CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction

NeurIPS 2026 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CompactSplat, a feedforward sparse-view 3DGS model that allocates Gaussians adaptively via a texture-driven quadtree (TAP) and multi-level attribute pyramid sampling (HPS), enabling inference-time sparsity control via a partition threshold.

Problem

Feedforward generalizable 3D Gaussian Splatting assigns one or more Gaussians to every pixel, producing spatially redundant primitives in homogeneous regions and sub-optimal planar modeling from sparse views.

Approach

Replace pixel-wise Gaussian prediction with a texture-guided quadtree partition over the input image; sample Gaussian attributes from a multi-resolution attribute pyramid at the LoD of each leaf, and scale each Gaussian by its block footprint so flat regions use fewer, larger primitives. A single partition threshold tau controls sparsity at inference without retraining.

Key results

• On RE10K, Ours-L averages 26.392 PSNR and 0.198 LPIPS across 256/512/1024 resolutions, versus 26.255 PSNR for SurfSplat, while using 72.6k Gaussians versus 131.1k (GPP 0.554 vs 1.000).
• On ACID, the method averages 26.146 PSNR and 0.268 LPIPS, with 91.4k Gaussians and GPP 0.625, and is best on average LPIPS and SSIM at 1024x1024 among compared methods.
• Cross-dataset results show Ours averaging 20.208 PSNR with GPP 0.685, slightly below SurfSplat (21.078 PSNR, GPP 1.000) in fidelity but far more compact.

Surprising findings

• At 256x256 on ACID the method trails SurfSplat and TranSplat in PSNR (27.992 vs 28.336), and its cross-dataset average PSNR is lower than the pixel-aligned baselines.
• The YonoSplat baseline is evaluated at 224x224 rather than 256x256, which the authors note may inflate its metrics.

Limitations

• Quadtree partitioning uses rigid orthogonal grids and is not shape- or semantics-aware.
• Redundancy across views is not addressed; only within-view redundancy is reduced.
• At standard 256x256 resolution on ACID and in some cross-dataset metrics, the method is slightly below strongest baselines in PSNR.

Experimental setup

Datasets: RealEstate10K ACID

Benchmarks: RealEstate10K ACID DL3DV ScanNet DTU

Metrics: PSNR SSIM LPIPS HRRC GPP GS Num

Baselines: pixelSplat HiSplat MVSplat TranSplat DepthSplat SurfSplat YonoSplat TokenGS C3G VolSplat

Builds on: DepthSplat SurfSplat Depth Anything V2 MVSplat

Closest papers by meaning

• Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
• Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.92)
• Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.91)
• GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
• Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.91)
• pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.91)