<p>Earthquake ground motions demonstrate inherent non-stationarity and significant intensity variability due to the complex interaction of source mechanisms, propagation paths, and site conditions. This presents dual challenges for existing modeling approaches, including data scarcity in extreme scenarios and coupling of time-frequency-amplitude features. This paper proposes an innovative decoupled generative framework by hierarchically separating three stages: time-frequency generation, amplitude prediction, and time history reconstruction. First, leveraging a conditional generative model, peak-normalized acceleration time-frequency spectra are generated randomly based on key physical parameters as constraints, revealing temporal variations in frequency and amplitude. Then, a cross-attention mechanism is employed to fuse parameters and spectral representations, enabling precise prediction of peak ground acceleration (PGA). Lastly, a phase recovery algorithm is applied to reconstruct the time-frequency spectra into time-domain signals, with amplitude scaled by PGA to produce realistic waveforms. Additionally, data augmentation strategies are proposed to improve physical consistency, alongside a gradient field analysis method to quantify the impact of physical constraints and improve interpretability. To validate its extrapolation capabilities, comparative experiments with the end-to-end simulation method were conducted in cross-regional earthquake scenarios. The results demonstrate that the proposed framework achieves better performance in both the accuracy of response spectra prediction and the fidelity of time-frequency detail restoration. This study provides a novel approach for stochastic simulation of strong earthquakes in data-scarce scenarios.</p>