Comparison of Loss Function Weighting Methods for Multi-Loss Function GAN

In the domain of image generation, Generative Adversarial Networks (GANs) have gained substantial attention since they were originally introduced. GANs are now required to produce images under particular circumstances that are customized to the needs of the user. One way to overcome this problem is to modify GANs to allow user-defined features to be taken into account during the training process. This is accomplished by including an additional network to assess the images produced by the GAN; the assessment outcomes are then incorporated into the GAN’s loss function. There must be a plan in place to ascertain the weight or contribution of each component in the total loss function, which grows as the number of factors that must be taken into account rises (multi-loss function). In this work, the original GAN is used as the initial comparison model to evaluate two weighting techniques used in multi-loss GANs: the static weighting approach and a dynamic weighting method. The experimental results reveal that the static weighting approach’s Inception Score and FID values outperform those of the dynamic technique and baseline GAN. More investigation is still needed into the dynamic approach used in this work, which promises automated weighting in multi-loss GANs.