Abstract: Volumetric clouds play an important role in meteorological visualization, digital content generation, and virtual scene construction. However, under sparse-view input conditions, traditional 3DGS methods often suffer from floating artifacts, background collapse, and unstable illumination representation due to insufficient geometric constraints and the lack of modeling for volumetric cloud lighting characteristics. To address these issues, SparseCloudGS is proposed, a volumetric cloud rendering method that integrates geometric priors, regularization constraints, and illumination modeling. First, a learning-based point cloud generation module is designed to predict the point cloud distribution in missing regions, providing a reasonable spatial prior for Gaussian initialization. Then, a seed-based geometric regularization module is introduced, which explicitly guides the growth, diffusion, and pruning of Gaussian ellipsoids through key seed partitioning and local minimal cluster construction. Furthermore, a hash-grid encoding module is incorporated to enhance feature representation in view-ambiguous regions and improve local detail expression. Finally, an illumination modeling module is introduced to model the cloud illumination contribution by considering the forward single-scattering and internal multiple-scattering characteristics of volumetric clouds. Experimental results on the DigitalCloud-PBR dataset show that SparseCloudGS effectively suppresses artifacts under sparse-view conditions. Compared with the baseline 3DGS, SparseCloudGS improves the SSIM by 0.206 and the PSNR by 4.16 dB, improving the structural consistency and visual quality of volumetric cloud rendering.