A U-Net-based CNN upscales images using SE blocks, pixel shuffle, and residual bottlenecks. Trained on Flickr2K with realistic degradations, MAE and perceptual-loss models improve detail recovery. Achieves strong PSNR/SSIM, producing sharper high-resolution outputs.