This paper proposes a novel method for reconstructing complete 3D scenes from a single image using Gaussian splats. Existing Gaussian splatting techniques require dense observation data and struggle to reconstruct occluded or unobserved regions. This study utilizes a latent diffusion model to reconstruct a complete 3D scene, including occluded regions, from a single image. Completing the surface of occluded regions is a challenging problem due to ambiguity. Existing methods rely on regression-based approaches that predict a single "mode," leading to blurriness, unrealistic interpretation, and inability to interpret multiple regions. In contrast, this study proposes a generative method that learns the 3D representation distribution of Gaussian splats based on a single input image. To address the lack of accurate data, we propose a variational autoreconstructor that learns the latent space using only 2D images using a self-supervised learning approach, and then train a diffusion model based on this model. Consequently, our method achieves faithful reconstructions and a diverse sample set, enabling the completion of occluded surfaces for high-quality 360-degree rendering.
