Curriculum learning for 3D vehicle geometry editing

Despite the strong potential of 3D editing to impact various industries, recent research has largely focused on converting text or images into 3D models, with limited attention to fine-grained control over existing 3D objects. Our previous work introduced a framework for continuous and attribute-specific editing of 3D vehicle models, but quantitative analysis revealed that the editor’s adherence to the pre-trained regressor remained suboptimal. In this study, we design a set of curriculum learning strategies to address this gap, gradually increasing editing difficulty and regularizing non-target dimensions. Experiments on 3D car models show that these strategies improve the reliability of latent-space editing, with consistent gains in attribute hit rate and success under evaluation, while preserving overall object identity. These results demonstrate that curriculum-guided training is a simple yet effective recipe for strengthening controllable 3D editing.