Reference-Free Evaluation Metric for Fine-Grained 3D Shape Editing

Evaluating the quality of fine-grained 3D shape editing, such as adjusting a vehicle’s roof length or wheelbase, is essential for assessing generative models but remains challenging. Most existing metrics depend on auxiliary regressors or large-scale human evaluations, which may introduce bias, reduce reproducibility, and increase evaluation cost. To address these issues, a reference-free metric for evaluating fine-grained 3D shape editing is proposed. The method is based on the Rich-Attribute Sufficiency Assumption (RASA), which posits that when a geometric attribute set is sufficiently comprehensive, models with the same attribute vector should exhibit nearly identical shapes. Following this assumption, the dataset itself serves as a validation source: each source model is edited to match a small set of target attribute vectors, and the post-editing similarity to the targets reflects the editor’s accuracy and stability. Reproducible indicators are defined, including mean similarity, variation across targets, and calibration with respect to attribute distance. Empirical validation demonstrates the effectiveness of the proposed metric, showing approximately 9% degradation under semantic perturbations and less than 2% variation across different target-sampling settings, confirming both its discriminative sensitivity and robustness. This framework provides a low-cost, regressor-free benchmark for fine-grained editing and establishes its applicability through an explicit assumption and evaluation protocol.