A Unified Framework for Automated 3D Modelling Pipelines: Analysis of Geometric, Topological and Perceptual Quality Assessment

Multimodal AI-driven 3D modelling has become an important production tool, yet it remains unclear which automatic retopology and remeshing strategies best balance geometric, topological and perceptual consistency. Prior studies of AI-integrated workflows emphasise asset-generation efficiency while overlooking the quality of resulting models. We developed a unified retopology, remeshing, texturing and multi-metric evaluation framework comparing quadric error metrics (QEM), instant meshes (IM) and quadriflow (QF) across eight geometric, topological and perceptual indicators. A tri-factorial design (source, algorithm, category) yielded 270 observations from 30 asset concepts generated on three platforms (90 reference meshes). Linear mixed-effects models included a random intercept for concept and fixed effects for algorithm, category, their interaction and platform. Significant algorithm main effects emerged for seven of eight metrics. QEM achieved the lowest geometric error and highest perceptual scores, IM the highest scaled Jacobian mean, and QF the strongest valence regularity. The algorithm–category interaction was significant only for the perceptual metrics, where all pairwise contrasts were significant for inorganic assets but only the QEM–IM contrast for organic assets. The effects persisted after adjustment for the mesh budget. Our findings offer practitioners a data-driven basis for category-conditional algorithm selection, potentially reducing production overhead. Future work will integrate multi-agentic workflows for immersive deployment.