ROAD: Reciprocal-Objective Alignment of Discriminative Semantics for 3D Shape Generation

Researchers introduced ROAD, a 3D generation framework that transfers semantic and structural priors from a discriminative 3D foundation model into diffusion transformers.

It uses a reciprocal alignment scheme with global semantic condensation and bipartite-matching-based structural alignment, while the foundation model is used only during training.

The paper reports competitive results versus Step1X-3D using just 1.5% of the training data, suggesting much lower data and compute costs for high-fidelity 3D generation.