Meshy 7 Sets New Standards for Image-to-3D Geometry Alignment

The core challenge for 3D generative models has shifted from simply avoiding broken meshes to ensuring the output faithfully mirrors the input. While leading tools now rarely produce mangled limbs or scrambled structures, they often fail to capture subtle proportions or specific engraved features. Meshy 7 addresses this by integrating alignment as a tracked signal directly into its training cycles, ensuring the model optimizes for the same fidelity users demand.

In testing against four competing foundation models, Meshy 7 outperformed the field across all three key metrics: overall proportion, spatial distribution, and surface detail. The gap is most pronounced in single-view input scenarios, where the model achieved a 5.3-point lead in surface detail over its nearest competitor. This performance stems from three architectural updates: a multi-scale image encoder capable of processing higher resolutions, a refined training dataset stripped of lighting and background noise, and a methodology that treats geometry alignment as a primary objective throughout the training process.

To validate these claims, Meshy developed a benchmark that compares generated 3D assets directly against ground-truth geometry rather than using vision-language judges. By rendering held-out reference models and measuring the geometric delta, the team ensures that the evaluation is objective and reproducible. While the initial release focuses on geometry, Meshy plans to expand this framework to include dedicated benchmarks for texture and material alignment in future updates.