GaussianLens: Localized High-Resolution Reconstruction via On-Demand Gaussian Densification

ICLR 2026 Conference Submission13044 Authors

Keywords: 3D Gaussian Splatting, 3D Reconstruction

TL;DR: We reconstruct local high-resolution details given an initial 3DGS reconstruction by learned on-demand Gaussian densification.

Abstract: We perceive our surrounding environments with an active focus, paying more attention to regions of interest, such as the shelf labels in a grocery store or a family photo on the wall. When it comes to scene reconstruction, this human perception trait calls for spatially varying degrees of detail ready for closer inspection in critical regions, preferably reconstructed on demand as users shift their focus. While recent approaches in 3D Gaussian Splatting (3DGS) can achieve fast, generalizable scene reconstruction from sparse views, their uniform resolution output leads to high computational costs, making them unscalable to high-resolution training. As a result, they cannot leverage available image captures at their original high resolution for detail reconstruction. Per-scene optimization methods reconstruct finer details with heuristic-based adaptive density control, yet require dense observations and lengthy offline optimization. To bridge the gap between the prohibitive cost of high-resolution holistic reconstructions and the user needs for localized fine details, we propose the problem of localized high-resolution reconstruction through on-demand generalizable Gaussian densification. Given an initial low-resolution 3DGS reconstruction, the goal is to learn a generalizable network that densifies the reconstruction to capture fine details in a user-specified local region of interest (RoI), based on sparse high-resolution observations of the RoI. This formulation avoids the high cost and redundancy of uniformly high-resolution reconstructions and enables the full leverage of high-resolution observations in critical regions. To address the problem, we propose GaussianLens, a feed-forward densification framework that fuses multi-modal information from the initial 3DGS and multi-view images. We further propose a pixel-guided densification mechanism that effectively captures details under significant resolution increases. Experiments demonstrate our method's superior performance in local high-fidelity detail reconstruction and strong scalability to images of up to $1024\times1024$ resolution.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 13044