VIP-GS: Enhancing Sparse Novel View Synthesis via In-Place Virtual Viewpoint Inpainting

While radiance field representations have achieved remarkable success in photo-realistic novel view synthesis with densely captured images, their performance sharply degrades under sparse input conditions, resulting in floater artifacts and missing regions caused by its inherent shape-radiance ambiguities and lack of informative observations, respectively. To address these challenges, we propose a few-shot 3DGS framework that leverages scene-adapted generative priors gated by multi-view geometric consistency to enhance the reconstruction quality for both interpolated and extrapolated views given limited observing viewpoints. Built upon coarse 3D Gaussians initiated via multi-view stereo, the core of our approach lies in lightweight personalized adaptation of a pretrained inpainting model from only the sparse input images, requiring no task-specific paired training data. During scene reconstruction, the current 3DGS reconstruction itself localizes unreliable virtual-view regions through multi-view warping consistency, where our personalized inpainting model is selectively invoked, and the repaired views are iteratively integrated to enrich the training corpus of Gaussian Splatting, progressively expanding the observing coverage and intensifying multi-view constraints. Extensive experimental results on three popular benchmark datasets including Tanks and Temples, MVImgNet and MipNeRF-360 demonstrate that our method, without using any specialized external training assets, dramatically improves the rendering quality of sparse-view novel view synthesis with a single consumer-grade GPU.