We present CrystalDiT, a diffusion transformer for crystal structure generation that achieves state-of-the-art performance by challenging the trend of architectural complexity. Instead of intricate, multi-stream designs, CrystalDiT employs a unified transformer that imposes a powerful inductive bias: treating lattice and atomic properties as a single, interdependent system. Combined with a periodic table-based atomic representation and a balanced training strategy, our approach achieves 8.78% SUN (Stable, Unique, Novel) rate on MP-20, substantially outperforming recent methods including FlowMM (4.21%) and MatterGen (3.66%). Notably, CrystalDiT generates 63.28% unique and novel structures while maintaining comparable stability rates, demonstrating that architectural simplicity can be more effective than complexity for materials discovery. Our results suggest that in data-limited scientific domains, carefully designed simple architectures outperform sophisticated alternatives that are prone to overfitting.
Cite
@article{arxiv.2508.16614,
title = {CrystalDiT: A Diffusion Transformer for Crystal Generation},
author = {Xiaohan Yi and Guikun Xu and Xi Xiao and Zhong Zhang and Liu Liu and Yatao Bian and Peilin Zhao},
journal= {arXiv preprint arXiv:2508.16614},
year = {2026}
}
Comments
18 pages, 18 figures. Code available at https://github.com/hanyi2021/CrystalDiT.git. Updated to remove copyright notice