A Comprehensive Survey on Diffusion Model-Driven 3D Reconstruction: Methods, Datasets, and Prospects

Three-dimensional (3D) reconstruction serves as a key technology bridging the real and digital worlds, with broad application in remote sensing, autonomous driving, and robotics. In recent years, its technical paradigm has shifted from geometry-based methods to data-driven approaches, with diffusion models emerging as a major driving force due to their stable training process, strong ability to learn priors from large-scale datasets, and excellent controllability over outputs. Despite the proliferation of diverse architectures, a systematic analysis and comparison of these methods remains absent. In this paper, we present a comprehensive review of diffusion-based 3D reconstruction methods. Based on the space where diffusion operates, we first categorize existing approaches into four types—image diffusion, latent diffusion, 3D diffusion, and hybrid diffusion—and we provide a detailed analysis of their methodologies. We then summarize commonly used 3D datasets and provide a comparative evaluation of these methods across three dimensions: reconstruction accuracy, computational efficiency, and generalization capabilities. Finally, we discuss future developments of diffusion-based 3D reconstruction methods.