Long acquisition times, a staple characteristic of Magnetic Resonance Imaging (MRI) can be shortened by aggressively undersampling k-space and solving the resulting ill-posed reconstruction problem using learning based sparsity priors. Unrolled optimization networks such as FISTA-Net provide an interpretable framework with strong data consistency but struggle to restore fine textures at very low sampling ratios ( \(\le \) 10%). Conversely, Generative Adversarial Network (GAN) based approaches sharpen details but are unstable and often hallucinate when the adversarial loss dominates. In this work, we introduce FISTA-GAN, an interpretable cosine-decayed GAN that combines unrolled FISTA iterations with a scheduled adversarial process. The generator is an unrolled FISTA-Net whose learnable convolutions act as proximal maps; the discriminator used is a spectrally normalized patch GAN. Training strategy starts with a generator warm-up, followed by a linear adversarial ramp and a cosine decay that smoothly anneals the GAN weight to \(\le \) 5% of the total loss. An exponential moving average (EMA) of the generator further stabilises convergence. In 10 \(\%\) brain data from the CC-359 benchmarked dataset, the proposed model achieves 39.5 dB PSNR and 0.9142 SSIM, surpassing state-of-the-art unrolled baselines (FISTA-Net) and GAN (HARA-GAN, RSCA-GAN) by up to 2.6 dB and 0.05 SSIM, respectively. The results demonstrate that the integration of interpretability and adversarial detail enhancement yields fast, high-fidelity reconstructions suitable for time-critical clinical workflows.