New Protocol Evaluates Single-Image 3D Mesh Quality Reliably

The rapid advancements in single-image-to-3D generation highlight a critical need for standardized, automated methods to evaluate the quality of generated 3D meshes. Currently, practitioners often rely on simple, automatic proxies such as CLIP similarity in render space or basic mesh geometry validity statistics. However, the effectiveness of these proxies in accurately reflecting perceived human quality has been largely unestablished. This research introduces a novel and validated VLM-judge evaluation protocol designed to reliably assess 3D mesh quality. The protocol involves a fixed 24-view headless rendering setup, utilizes two independent vision-language judge families, and incorporates a mandatory position-bias correction to ensure consistent verdicts. The two judge families showed substantial agreement, indicating the protocol's robustness. Crucially, the study demonstrates that the commonly used cheap proxies are inadequate substitutes for this new VLM-judge protocol. Geometry validity provided only a weak signal, while render-CLIP performed at chance levels. The findings suggest that these proxies are misleading as optimization targets, especially for subtle defects, and recommend the VLM-judge protocol as a more reliable and reproducible evaluator for 3D mesh quality.