Abstract: Recent advances in deep generative models have opened up new opportunities for the discovery of novel metal-organic frameworks （MOFs）. To better explore MOF structural diversity and expand their accessible chemical space, we propose BHDiT, a hybrid Diffusion Transformer with flow matching for building-block generation, and integrate it into an end-to-end workflow for MOF structure generation. BHDiT encodes 3D molecular structures as sequences of atomic coordinates and features and introduces 3D Rotary Positional Encoding （3D-RoPE） to enhance geometric awareness. On QM9, BHDiT improves uniqueness and novelty while maintaining competitive structural quality. On a MOF building-block dataset, it achieves validity rates of 78.95%, 82.76%, and 83.29% at 100, 500, and 1 000 sampling steps, respectively, achieving uniqueness of 99.67% and novelty of 99.48%. Notably, the model retains a 50.2% validity rate even for structures exceeding 50 atoms. The proposed workflow generates MOFs by predicting building-block compositions and assembling them into complete structures. Zeo++ analysis further indicates that increasing building-block diversity can effectively enhance the diversity of generated MOFs, underscoring the promise of deep generative models for MOF design.