MAP3DNet: Multimodal Aesthetic Prediction for 3D Models?
Abstract
Predicting the aesthetic appeal of 3D models is an emerging problem at the intersection of computer vision, graphics, and human preference modeling. Unlike geometric and visual quality assessment, aesthetic prediction aims to estimate how visually appealing a 3D asset might be perceived by users. We formulate this task as a regression task, learning to predict a global aesthetic score derived from user engagement statistics collected from a 3D content sharing platform, where likes considered relative to views provide a scalable proxy for collective appreciation. We build on the ViDA3D dataset, a collection of Sketchfab models represented by multi-view renderings and artist-provided tags. To complement this textual information, we generate captions from rendered views and propose MAP3DNet, a lightweight multimodal architecture that combines visual features, tags, and generated captions. MAP3DNet learns to weight different viewpoints, encodes visual and textual cues through vision-language representations, and uses cross-modal attention to predict the final aesthetic score. Experiments demonstrate that our multimodal approach improves over visual-only alternatives. While visual features remain the main cue, tags and generated captions provide important complementary semantic context. MAP3DNet outperforms adapted baselines and produces rankings closely aligned with those derived from human engagement, supporting the assumption that aesthetic evaluation should be modeled as a context-aware, multimodal phenomenon.
