Manfredi, GildaCapece, NicolaErra, UgoGruosso, AntonioComino Trinidad, MarcMancinelli, ClaudioMaggioli, FilippoRomanengo, ChiaraCabiddu, DanielaGiorgi, Daniela2025-11-212025-11-212025978-3-03868-296-72617-4855https://doi.org/10.2312/stag.20251327https://diglib.eg.org/handle/10.2312/stag20251327Generating 3D objects with complex, nonlinear shapes directly from images is still an open research area. To address this problem, several state-of-the-art methods use Deep Learning (DL) to predict a set of parameters from images, which are then used to generate the 3D geometry, leveraging the characteristics of procedural modeling. Recently, Kolmogorov-Arnold Networks (KANs) have emerged as an alternative to traditional Multilayer Perceptrons (MLPs) in DL, and have been successfully integrated into architectures such as Convolutional Neural Networks (CNNs), Graph Neural Networks, and Transformers. In this work, we propose a DL architecture consisting of a hybrid CNN-KAN network for parametric 3D model generation from images. The model combines the ability of KANs to capture complex nonlinear functions with the strong visual feature extraction capabilities of CNNs. The method is evaluated using both quantitative error metrics and qualitative visualizations comparing predicted shapes with ground truth, and its performance is compared against a more standard CNN-MLP architecture.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Artificial intelligence; Modeling methodologiesComputing methodologies → Artificial intelligenceModeling methodologies10.2312/stag.202513279 pages