In the study of spiral galaxy morphology, spiral arm structures are valuable for intuitively reflecting active physical and chemical processes within galaxies. However, long-term scarcity of high-quality one-, three-, and four-armed galaxy samples has limited deep learning model performance. To address this, this study developed a spiral galaxy data simulation program with a three-stage workflow: first, screening highly reliable training samples; second, selecting the best-performing Imagen architecture as the generative model after comparing nine mainstream ones; finally, training Imagen to generate an open data set of 9402 one-/three-armed galaxies, expanding the original sample size by 6 times. Multi-dimensional evaluations verified reliability and usability: Fréchet Inception Distance scores for N = 1 and N = 3 tasks were 6.05 and 9.13; the t-distributed Stochastic Neighbor Embedding showed generated data covered and expanded real data distribution; the Structural Similarity Index Measure confirmed no sample duplication. In downstream validation, data augmentation improved seven classification models’ average accuracy by 8.7% (DenseNet peaked at 97%), and SHapley Additive exPlanations analysis showed model decisions focused on spiral arm topology. In conclusion, the program and data set support spiral galaxy morphology deep learning research and are publicly available at https://github.com/TuAstroAILab/AstroGS.