Project page · Training and evaluation code available
1 Huazhong University of Science and Technology 2 Megvii 3 Zhejiang University
* Corresponding author
High-fidelity 3D generation predominantly relies on scaling model capacity and data, which incurs prohibitive computational costs. This paradigm typically requires learning geometry from scratch and overlooks the rich semantic and structural priors already encapsulated in discriminative 3D foundation models. We contend that leveraging the understanding of the 3D world possessed by these discriminative models can significantly reduce generative cost.
To this end, we propose ROAD, a framework that transfers discriminative priors into diffusion transformers through reciprocal-objective alignment. Holistic Semantic Condensing enforces global semantic coherence, while Structural Optimal Alignment formulates fine-grained correspondence as a bipartite matching problem. The 3D foundation model is used only for training-time supervision and introduces no additional inference cost. Compared with the industrial Step1X-3D baseline, ROAD achieves highly competitive generation performance with only 1.5% of the training data.
3D Shape Diffusion Models have emerged as a dominant paradigm for generating high-fidelity 3D assets, but their success relies heavily on brute-force scaling. Such systems consume massive datasets and computational resources to establish the mapping between visual conditions and geometric structures, even though discriminative 3D foundation models have already acquired rich semantic and structural knowledge of the 3D world.
Transferring these priors is non-trivial. Point clouds are inherently unordered, and generative and discriminative models employ fundamentally different tokenization strategies. Generative tokens tend to aggregate holistic surface context and remain order-agnostic, whereas foundation-model tokens preserve localized geometric structure. As a result, rigid element-wise alignment lacks a natural one-to-one correspondence.
HSC aggregates disordered tokens into unified semantic prototypes through global pooling. This holistic anchor resolves the unordered nature of 3D representations and guides the generative model toward the semantic manifold learned by the foundation model.
SOA preserves fine-grained geometric information by treating local correspondence as a bipartite matching problem. Hungarian matching dynamically pairs tokens according to semantic similarity, avoiding the incorrect assumption of fixed positional correspondence.
Together, the two reciprocal objectives transfer both macroscopic semantic abstractions and microscopic geometric primitives without disrupting the native generative flow.
Evaluated on the representative Step1X-3D framework, ROAD reaches a Uni3D-Score of 32.3 and a ULIP-Score of 33.7. Effective prior transfer substantially reduces the amount of data and computation required for high-quality generation, demonstrating that semantic alignment can serve as a practical alternative to massive data scaling.
Textured GLB outputs rendered as 360-degree turntables. ROAD maintains coherent geometry and appearance across viewpoints.
The release includes the Step1X-3D rectified-flow baseline, ROAD / Uni3D-REPA training, and Uni3D-I and ULIP-I evaluation.
conda env create -f environment.yml
conda activate step1x
pip install -r requirements.txt
# Train ROAD on four GPUs
GPU_IDS=0,1,2,3 bash scripts/train_uni3d_repa.shView the complete setup and evaluation instructions on GitHub →
@misc{luo2026road,
title = {ROAD: Reciprocal-Objective Alignment of
Discriminative Semantics for 3D Shape Generation},
author = {Luo, Xiao and Du, Mingyang and Zhou, Xin and
Feng, Tianrui and Chen, Xiwu and Li, Xiaofan and
Zhang, Jiangning and Liang, Dingkang},
year = {2026}
}