Generative models for crystal structure prediction (CSP) have expanded the search space for novel materials and improved screening efficiency. However, simultaneously generating stable and novel structures remains challenging because of the complex structural and physical constraints governing crystals. To address this challenge, we propose site-orbit masked twin-branch Transformer model (ST2M), a symmetry-guided discrete–continuous diffusion framework incorporating degree-of-freedom masks and conditioned on Wyckoff patterns. ST2M represents crystal structures as sequences of independent occupied Wyckoff orbits and employs a twin-branch Transformer to jointly denoise discrete and continuous variables. In our experiments, ST2M achieves a structural validity of 96.37%. For known patterns in dataset, 23.86% of generated structures simultaneously satisfy stability, uniqueness, and novelty criteria, representing a relative improvement of 68.15% over the best-performing baseline. First-principles calculations identify a total of 21 previously unreported stable structures among selected candidates with low predicted energies. These results demonstrate that ST2M can generate chemically viable crystal candidates under prescribed crystallographic constraints and generalize to unknown Wyckoff patterns, providing a practical route to accelerated materials discovery.