IGAN: An Inception-Based Generative Adversarial Network for Multi-Scale Image Synthesis

Abstract

Generative Adversarial Networks (GANs) provide a powerful framework for image synthesis, but their practical effectiveness is often limited by the interplay between high visual fidelity and stable adversarial training. Heavily parameterized GANs can improve feature representation; yet, they may also increase the risk of gradient instability or and mode collapse and hence representation design and adversarial optimization must be considered together. This study presents IGAN, a specific Inception-based configuration for unconditional synthesis of 64×64 RGB images. Four-branch modules combine pointwise, local, wider-field, and pooled features in both the generator and discriminator. Bottleneck projections allocate channels before the spatial convolutions; generator modules preserve resolution, whereas discriminator modules reduce it. The documented generator module uses 199,072 convolutional parameters, compared with 590,080 for a corresponding unreduced four-branch block, a 66.26% reduction in parameters count. The model is experimentally evaluated on CUB-200 and selected ImageNet categories. The reported results show Fréchet Inception Distance (FID) values of 13.12 and 15.08 on CUB-200 and ImageNet, respectively, and Inception Scores of 9.27 and 68.25, respectively, compare to certain GAN variations baselines.