Yuhan Zhang Long Zhuo Ziyang Chu Tong Wu Zhibing Li Liang Pan Dahua Lin Ziwei Liu

Abstract

Despite rapid advances in 3D content generation, quality assessment for thegenerated 3D assets remains challenging. Existing methods mainly rely onimage-based metrics and operate solely at the object level, limiting theirability to capture spatial coherence, material authenticity, and high-fidelitylocal details. 1) To address these challenges, we introduce Hi3DEval, ahierarchical evaluation framework tailored for 3D generative content. Itcombines both object-level and part-level evaluation, enabling holisticassessments across multiple dimensions as well as fine-grained qualityanalysis. Additionally, we extend texture evaluation beyond aestheticappearance by explicitly assessing material realism, focusing on attributessuch as albedo, saturation, and metallicness. 2) To support this framework, weconstruct Hi3DBench, a large-scale dataset comprising diverse 3D assets andhigh-quality annotations, accompanied by a reliable multi-agent annotationpipeline. We further propose a 3D-aware automated scoring system based onhybrid 3D representations. Specifically, we leverage video-basedrepresentations for object-level and material-subject evaluations to enhancemodeling of spatio-temporal consistency and employ pretrained 3D features forpart-level perception. Extensive experiments demonstrate that our approachoutperforms existing image-based metrics in modeling 3D characteristics andachieves superior alignment with human preference, providing a scalablealternative to manual evaluations. The project page is available athttps://zyh482.github.io/Hi3DEval/.