Generative Adversarial Networks (GANs) are remarkably versatile tools for data generation in the materials science domain, in particular for reconstruction of microstructures based on desired properties or material discovery. In such applications, the aim is to obtain a microstructure corresponding to a continuous value, i.e., a numerical value in a specific range. Although various methods exist for conditioning GANs on continuous values, little is known regarding the key mathematical properties of the conditioning architecture involved. In this work, we train GANs with different conditioning architectures on a range of material-science-related and computer vision datasets. We then compute the properties of the embedding architectures, such as monotonicity and linearity, and relate them to overall GAN performance metrics. Our findings show that the underlying relationships between images and labels of datasets dictate the optimal conditioning strategies. Building on these observations, we propose a novel framework for continuous conditioning in GANs -- a strategy that strongly simplifies and accelerates the training pipeline while maintaining high fidelity. Consequently, our framework allows for solving the ill-posed, inverse problems in a computationally efficient manner while adhering to the underlying physical principles.