Restoring Quality from Bitrate Collapse: A Two-Stage GAN for Enhancing Heavily Compressed Video

Low-bitrate video compression (e.g., H.264/AVC at ≤300 Kbps) typically introduces visible artifacts such as blocking, blurring, and texture loss. This paper proposes a two-stage Generative Adversarial Network (GAN) architecture tailored to restore visual quality in degraded video sequences. The system incorporates motion alignment, residual blocks with attention mechanisms, and multi-frame temporal modeling to enhance spatial fidelity and consistency. A novel training dataset is constructed by synthetically compressing high-quality video content to simulate real-world degradation. We analyze the architecture in detail, discuss training stability (including mode collapse mitigation), and propose a combination of distortion and perceptual losses, including L1, SSIM, LPIPS, and adversarial objectives. Quantitative evaluation on standard benchmarks shows that the proposed model achieves competitive or better performance compared to earlier methods like ESRGAN, EDVR, CVEGAN, and traditional deblocking techniques. We further present visual comparisons, ablation studies, and training dynamics to validate each architectural component. The enhanced frames exhibit restored detail and consistent temporal structure across sequences. A key novelty lies in targeting extremely compressed content and demonstrating restoration capability under these constraints. This makes the approach suitable for scenarios such as cloud video storage or ultra-low-bandwidth transmission, where post-decompression enhancement is crucial.