CDR-GS: 3D Gaussian Splatting for Cross-Device Rendering

Abstract

3D Gaussian Splatting has achieved remarkable progress in real-time scene rendering, yet existing methods are predominantly tailored to a single camera model, typically the pinhole camera, which limits their applicability to images captured by heterogeneous devices. In this paper, we introduce CDR-GS, the unified 3DGS framework explicitly designed for cross-device rendering. Our key insight is to leverage 360° images as a universal intermediate representation that inherently encompasses the full field of view (FoV) of diverse camera models, thereby eliminating the need for error-prone block-wise rendering and mapping. A differentiable cross-device mapping then transforms the intermediate 360° images into target camera domains, forming a unified rendering pipeline supervised directly by original cross-device images. To mitigate the excessive GPU memory overhead of activating nearly all Gaussians during 360° rendering, we propose a Gaussian selection strategy that exploits prior Gaussians and target FoV to retain only high-contribution Gaussians. We further design an adaptive-FoV mapping method that substantially reduces textureless regions when projecting cross-device images into pinhole images for prior Gaussians generation. To facilitate future research, we establish CDR-Bench, a benchmark comprising three datasets and adapted baselines. Extensive experiments demonstrate that CDR-GS achieves state-of-the-art rendering quality and efficiency across diverse camera models.