MFDFA-Enhanced GAN: A Hardware-Optimized Architecture for Efficient Synthetic Image Generation

Generative Adversarial Networks (GANs) have achieved significant success in synthetic image generation; however, training instability, mode collapse, and high computational demand remain persistent challenges, particularly for hardware-constrained deployments. This work proposes a multifractal-augmented GAN framework (MGAN) that integrates Multifractal Detrended Fluctuation Analysis (MFDFA) within the discriminator to enhance stability and feature discrimination during training. A structured transformation reformulates image inputs as sequential signals, enabling multifractal spectrum extraction and multiscale analysis within the adversarial learning process. The embedded MFDFA module acts as a stability-aware discriminator enhancement, encouraging smoother gradient behavior and improved convergence characteristics without modifying the fundamental generator-discriminator architecture. Experimental evaluation on MNIST, Fashion-MNIST, and CelebA datasets demonstrates improved generative quality as measured by Frechet Inception Distance(FID) and Inception Score (IS), along with faster convergence trends compared to baseline GAN implementations. To validate hardware feasibility, the inference module is optimized using pipelining and tiling strategies and deployed on an AMD Kintex-7 FPGA (KC705). Hardware evaluation shows 92.05% reductions in execution latency and 86.03% reduction in power consumption relative to a conventional GAN implementation. The proposed MGAN framework therefore provides a scalable and hardware-efficient approach for stable synthetic image generation in resource-constrained environments.