A deep learning generative model, Cond-CDVAE, is introduced for crystal structure prediction (CSP). This model is trained on a large dataset of 670,979 locally stable crystal structures, including high-pressure and ambient-pressure structures. The model is conditioned on user-defined parameters such as composition and pressure, enabling the generation of physically plausible crystal structures without requiring local optimization. The model achieves high accuracy in predicting experimental structures, with 59.3% accuracy for ambient-pressure structures and 83.2% for structures with fewer than 20 atoms per unit cell. These results are comparable or better than conventional CSP methods based on global optimization. The model's performance is validated through extensive benchmarking, demonstrating its ability to efficiently sample the material configuration space and generate high-fidelity crystal structures with root mean squared displacements (RMSDs) well below 1 Å. The model is particularly effective in generating structures under various pressure conditions, showing improved performance in predicting complex high-pressure structures. The model's success rate in predicting experimental structures increases with the number of samplings, reaching 59.3% for ambient-pressure structures and 83.2% for structures with fewer than 20 atoms per unit cell. The model's performance is further validated through comparisons with other CSP methods, such as CALYPSO, showing that it requires fewer samplings to locate the global minimum for certain structures. The model's ability to generate structures with user-specified physical conditions makes it a valuable tool for CSP tasks, particularly in materials research where determining the ground-state structure under specific pressure conditions is crucial. The model's performance is also demonstrated through the generation of structures for lithium, boron, and silica, showing its effectiveness in predicting complex high-pressure phases. The model's success rate in predicting these structures is comparable to or better than conventional methods, highlighting its potential in CSP applications. The model's performance is further supported by its ability to generate structures with high thermodynamic stability and accurate predictions of lattice parameters and atomic positions. The model's success in generating structures with user-specified conditions and its ability to efficiently sample the material configuration space make it a promising tool for CSP tasks. The model's performance is also validated through its ability to generate structures with high accuracy and efficiency, demonstrating its potential in materials research and discovery.A deep learning generative model, Cond-CDVAE, is introduced for crystal structure prediction (CSP). This model is trained on a large dataset of 670,979 locally stable crystal structures, including high-pressure and ambient-pressure structures. The model is conditioned on user-defined parameters such as composition and pressure, enabling the generation of physically plausible crystal structures without requiring local optimization. The model achieves high accuracy in predicting experimental structures, with 59.3% accuracy for ambient-pressure structures and 83.2% for structures with fewer than 20 atoms per unit cell. These results are comparable or better than conventional CSP methods based on global optimization. The model's performance is validated through extensive benchmarking, demonstrating its ability to efficiently sample the material configuration space and generate high-fidelity crystal structures with root mean squared displacements (RMSDs) well below 1 Å. The model is particularly effective in generating structures under various pressure conditions, showing improved performance in predicting complex high-pressure structures. The model's success rate in predicting experimental structures increases with the number of samplings, reaching 59.3% for ambient-pressure structures and 83.2% for structures with fewer than 20 atoms per unit cell. The model's performance is further validated through comparisons with other CSP methods, such as CALYPSO, showing that it requires fewer samplings to locate the global minimum for certain structures. The model's ability to generate structures with user-specified physical conditions makes it a valuable tool for CSP tasks, particularly in materials research where determining the ground-state structure under specific pressure conditions is crucial. The model's performance is also demonstrated through the generation of structures for lithium, boron, and silica, showing its effectiveness in predicting complex high-pressure phases. The model's success rate in predicting these structures is comparable to or better than conventional methods, highlighting its potential in CSP applications. The model's performance is further supported by its ability to generate structures with high thermodynamic stability and accurate predictions of lattice parameters and atomic positions. The model's success in generating structures with user-specified conditions and its ability to efficiently sample the material configuration space make it a promising tool for CSP tasks. The model's performance is also validated through its ability to generate structures with high accuracy and efficiency, demonstrating its potential in materials research and discovery.