A Survey of Single-Image 3D Reconstruction for Low-Latency Interactive Applications and Engineering Deployment

Abstract

Single-image 3D reconstruction aims to recover or generate the three-dimensional geometry and appearance of a target object from a single two-dimensional image. It has considerable application value in fields such as e-commerce, game development, virtual reality, augmented reality, and digital content creation. In recent years, relevant techniques have evolved from explicit representations, including voxels, point clouds, and meshes, toward neural implicit representations, multi-view diffusion models, large feed-forward reconstruction models, and general-purpose 3D generative models. Existing studies generally focus on geometric accuracy and visual quality; however, different methods exhibit substantial differences in inference latency, GPU memory requirements, input preprocessing, output asset formats, and post-processing costs. This paper reviews the major technical pipelines of single-image 3D reconstruction, with a focus on representative methods including 3D-R2N2, PSGN, Pixel2Mesh, Occupancy Networks, NeRF, Zero-1-to-3, Wonder3D, LRM, TripoSR, and CRM. These methods are compared from the perspectives of latency levels, end-to-end latency, computational resources, input robustness, asset usability, and platform integration difficulty. On this basis, this paper summarizes the major challenges associated with unobserved-region inference, generalization to real-world images, adaptation to different camera settings and materials, mesh quality, and evaluation criteria. It further discusses future directions, including hierarchical generation, 3D foundation models, sparse computation, unified latent spaces, interactive editing, and the generation of production-ready 3D assets.