SPAR3D: Stable Point-Aware 3D Object Reconstruction Based on a Single Image

Single image-based 3D object reconstruction technology has developed largely in two directions. Regression-based models efficiently infer visible surfaces but are vulnerable when handling occluded regions, while generative models handle uncertain regions well but suffer from high computational cost and misalignment with visible surfaces.

SPAR3D proposes a new two-stage approach that combines the strengths of both methods.

• Stage 1: A lightweight point diffusion model is used to quickly generate a sparse 3D point cloud.
• Stage 2: The generated point cloud and the input image are used together to construct a highly precise mesh.

Thanks to this design, SPAR3D resolves uncertainty through probabilistic modeling while simultaneously maintaining high computational efficiency and output quality. It also uses the point cloud as an intermediate representation, providing a feature that allows users to edit interactively.

According to evaluation results across various datasets, SPAR3D outperformed existing state-of-the-art (SOTA) methods, with an inference speed of only 0.7 seconds.