This work uses deep generative adversarial networks (GANs) to simulate conditional structural geomodels, which are representations of geometric elements of geology. Geomodeling is an ill-posed problem as different geomodels can respect the same conditional data. One approach to address this problem is through geomodel simulations, which remain a challenge. Current methods, such as the potential field method based on geostatistics, struggle to both characterize all type of uncertainties and produce realistic geomodels in complex geological settings. This study proposes to take advantage of advanced GANs such as Least-Squares GAN (LSGAN) to reproduce the unconditional spatial structure from a synthetic training dataset of realistic geomodels. The synthetic training dataset used is generated by Noddy using a set of geological parameters (including the number of domains, their thickness, the dip and the folding parameters). Once trained, the GAN can simulate unconditional geomodels. This is the same idea as unconditional simulations in geostatistics. In a previous work, we presented techniques to handle hard known- domain data. The main novelty of this work is to be able to take into account geological orientations, such as dip, to condition geomodels. To achieve this, the generator is trained to reproduce a discretized continuous function where isovalues define interfaces between geological domains. The gradient of this function, which can be computed through finite differences, provides access to geological orientations. The conditional step is performed in a Bayesian framework, where the GAN defines a prior distribution and conditional data defines the likelihood, inducing a posterior distribution. Samples from this posterior distribution are obtained using Metropolis-adjusted Langevin algorithm, a Monte Carlo Markov Chain algorithm. Finally, the proposed enables access to a wider range of geological scenarios allowing a better handling of uncertainties.