ACM International Conference on Multimedia

Zhenyu Li,Shanshan Gao,Guangshun Wei✉,Jianing Zhang,Wenping Wang,Yuanfeng Zhou✉

Abstract

High quality 3D asset generation remains a long-standing challenge in computer graphics, with the core difficulty lying in accurately recovering geometry and appearance that are consistent with a given input under limited single view conditions. However, existing generation paradigms struggle to impose reliable structuraland appearance constraints from a single image, often resulting in inconsistencies in geometry and texture. To address these challenges, we propose HiQu3DGen, a consistency-driven frameworkthat guides 3D asset generation through the synthesis of viewconsistent multi-view representations. The proposed frameworkadopts a bidirectional recursive multi-view generation strategy toexplicitly strengthen inter-view dependencies during generation,thereby reducing the emergence of mutually contradictory views.Leveraging these structurally consistent multi-view observations,it constructs a robust geometric intermediate and further refines geometry and texture by tightly coupling multi-view appearancecues, ultimately producing high quality textured 3D assets. Extensive quantitative and qualitative evaluations on both simple andcomplex 3D asset generation benchmarks emonstrate that theproposed framework consistently outperforms existing methods interms of geometric consistency and texture fidelity