Text2MFN and MFPredictor: Text-Conditioned Mesh Face-Count Recommendation for Efficient 3D Generation

Current text-to-3D methods primarily improve geometric fidelity, topology, and generation efficiency, but generally require mesh complexity to be determined indirectly by architecture, sequence length, user settings, or post-processing. We formulate text-conditioned mesh face-count recommendation as a regression task that predicts a practical triangular-face budget from the object category, silhouette, semantic parts, structural relationships, and geometry-related details expressed in a prompt. We construct Text2MFN, a synthetic dataset containing 1,000,000 text–face-count pairs, and propose MFPredictor, a lightweight dual-branch model that separates semantic context from geometry-related complexity cues and combines them through bidirectional calibration, gated fusion, a small mixture-of-experts module, and residual regression. All models are trained exclusively on Text2MFN and evaluated on its held-out test set and on an independent benchmark containing 200,000 QEM-derived multi-LOD records from 40,000 Objaverse meshes. MFPredictor obtains a MAPE of 4.38% and a ±20% accuracy of 95.45% on Text2MFN, and a MAPE of 7.96% and a ±20% accuracy of 86.41% on the Objaverse-based benchmark. With 481,275 parameters and a measured average regression latency of 2.09 ms, the model provides an efficient front-end prediction of mesh complexity. The real-mesh results demonstrate transfer under the evaluated Text2MFN-to-Objaverse/QEM protocol and define the validated scope of the reported generalization.