Evaluating Diffusion Models for Single-Image 3D Gaussian Splatting Scene Reconstruction

Abstract

Sparse or limited observations often cause 3D Gaussian Splatting (3DGS) scenes to exhibit holes, missing content, and unstable appearance in novel views. Repairing these regions requires generating synthetic frames that remain consistent across viewpoints, a challenge closely related to cross-view drift in generative models. In this work, we study synthetic frame generation as a repair strategy for incomplete 3DGS renderings under controlled sparse-view conditions. We focus primarily on image-based repair, including sequential frame-by-frame generation, anchor-guided repair, and masked local inpainting. In addition, we include a limited endpoint-conditioned video-generation comparison using just repaired first and last frames to condition intermediate-view synthesis. Our analysis focuses on cross-view drift, since repaired frames must remain geometrically coherent and visually stable to support plausible scene completion. We assess geometric consistency, and perceptual drift across repaired sequences. Sequential repair shows drift accumulation over longer trajectories particularly when occlusions are revealed, despite the strong correlation between consecutive frames, while anchor-guided repair reduces this effect. The additional endpoint-conditioned comparison also yields stable sequences. These findings suggest that stronger cross-frame constraints are beneficial for repairing incomplete 3DGS renderings under sparse-view conditions.