Dynamic Fusion of Multi-Scale Perception and Adaptive Discrimination for Compressed GANs

In recent years, Generative Adversarial Networks (GANs) face deployment challenges due to single-network feature extraction limitations, which hinder multi-scale hierarchical feature capture and cross-semantic dependencies. This results in degraded knowledge transfer from teacher to student models, suboptimal generalization, and training instability. To address this, we propose Dynamic Fusion of Multi-Scale Perception and Adaptive Discrimination (DFAD).DFAD employs dynamic weighting to harmonize cross-architectural discrepancies and minimize perceptual gaps between student-teacher models, enhancing feature disentanglement. Additionally, an adaptive convolutional block with channel-wise attention dynamically adjusts feature map importance, improving discriminator flexibility and mitigating mode collapse. Experiments show DFAD reduces CycleGAN’s computational costs by 40× MACs and 80× parameters while achieving a 72.91 FID (outperforming SOTA’s 73.54). The framework maintains Pix2Pix-level quality on resource-constrained devices, demonstrating robust generalization across tasks like style transfer and super-resolution. By reconciling multi-scale feature hierarchies and enabling real-time adaptive discrimination. This work bridges the gap between lightweight compression and high-fidelity image translation, offering a scalable solution for edge computing applications.