Grad-CAM Guided Framework for Quality-Preserving Channel Pruning in GANs

The remarkable performance of Generative Adversarial Networks (GANs) in tasks such as picture synthesis and editing is well known. However, especially in contexts with limited resources, their massive structures and high parameter counts pose serious computing hurdles. Here, we provide a Grad-CAM Guided technique to prune StyleGan, a unique method that computes channel-wise significance scores by utilizing Grad-CAM, a class-discriminative localization technique. This allows us to efficiently trim channels without sacrificing visual quality. Our methodology prioritizes perceptual quality over standard content-based or diversity-driven pruning techniques, striking a compromise between computational efficiency and output fidelity. Our approach uses Grad-CAM maps to emphasizes important regions in the generated sample and helps in computing the channel sensitivity better and incorporates the results into a channel-wise pruning plan. StyleGAN models using popular datasets such as FFHQ-256, LSUN-HORSE, and LSUN-CHURCH have been used to validate the framework. Our method's robustness is highlighted by experimental results that show large parameter count reductions while maintaining competitive Fréchet Inception Distance (FID) scores. This work shows how Grad-CAM may be used to guide quality-aware pruning techniques and provides a fresh viewpoint on generative model optimisation by fusing interpretability with computational efficiency. This methodology has more applicability in real-world scenarios because it can be tailored to different GAN designs.