GSFixer: Improving 3D Gaussian Splatting with Reference-Guided Video Diffusion Priors

3D reconstruction and novel view synthesis (NVS) are fundamental tasks in computer vision and graphics, with wide-ranging real-world applications in virtual reality, autonomous driving, and robotics. Recently, 3D Gaussian Splatting (3DGS) (Kerbl et al., 2023) has achieved impressive results in both reconstruction quality and rendering efficiency when dense input views are available. However, its performance degrades significantly in sparse-view settings, where limited viewpoint information leads to under-constrained 3D representations. In such cases, 3DGS often suffers from severe artifacts, including distorted geometric structures and incomplete reconstructions, particularly in less-observed regions or extreme novel viewpoints. These limitations hinder its applicability in real-world scenarios where acquiring dense multi-view data is challenging. To alleviate the limitations, some previous regularization methods have been proposed to introduce additional constraints into the 3DGS optimization process, such as monocular depth (Li et al., 2024; Zhu et al., 2024) , frequency smoothness (Zhang et al., 2024) , and random dropout (Xu et al., 2025) . While these approaches can help prevent 3DGS representations from overfitting to sparse input views, they often remain sensitive to noise and yield only marginal improvements in NVS rendering quality. Inspired by the success of ReconFusion (Wu et al., 2024) , which introduces diffusion model into NeRF (Mildenhall et al., 2020) optimization, more recent studies (Liu et al., 2024b;a; Wu et al., 2025a;b) explore incorporating 3DGS optimization with powerful generative priors from diffusion models, which are trained on internet-scale data. These strong priors enable the correction of spurious geometry or the inpainting of plausible content in novel views. However, a key challenge still remains: maintaining visual and 3D consistency between the generated and original input images, especially when the novel views are far from the observed inputs. Meanwhile, recent advances in controllable video generation have demonstrated the effectiveness of incorporating various conditional signals (