Abstract Objective. Following previous works on virtual sources model with Generative Adversarial Network (GAN), we extend the proof of concept for generating back-to-back pairs of gammas with timing information, typically for Monte Carlo simulation of Positron Emission Tomography(PET) imaging. Approach. A conditional GAN is trained once from a low statistic simulation in a given attenuation phantom and enables the generation of various activity source distributions. GAN training input is a set of gammas exiting a phantom, tracked from a source of positron emitters, described by position, direction and energy. A new parameterization that improves the training is also proposed. An ideal PET reconstruction algorithm is used to evaluate the quality of the GAN. Main results. The proposed method is evaluated on National Electrical Manufacturers Association (NEMA) International Electrotechnical Commission (IEC) phantoms and with CT patient image showing good agreement with reference simulations. The proportions of 2-gammas, 1-gammas and absorbed-gammas are respected to within one percent, image profiles matched and recovery coefficients were close with less than 5% difference. GAN tends to blur gamma energy peak, e.g. 511 keV. Significance. Once trained, the GAN generator can be used as input source for Monte Carlo simulations of PET imaging systems, decreasing the computational time with speedups up to ×400 according to the configurations.