AI super-resolution uses deep learning to upscale images beyond their original resolution while adding realistic detail. How it works: Neural networks learn mapping from low-res to high-res images, predict plausible high-frequency details (textures, edges) not present in input. Key architectures: SRCNN (pioneering), ESRGAN (GAN-based, realistic textures), Real-ESRGAN (handles real-world degradation), SwinIR (transformer-based). Training: Pairs of low-res and high-res images, combine L1/L2 reconstruction loss with perceptual loss and GAN loss for realistic textures. Real-world vs synthetic degradation: Models trained on bicubic downsampling fail on real photos (noise, compression, blur). Real-ESRGAN handles diverse degradation. Scale factors: 2x, 4x common, larger scales increasingly hallucinate. Multiple smaller upscales sometimes better than single large. Applications: Photo enhancement, video upscaling, game texture mods, satellite imagery, medical imaging. Limitations: Cannot recover information not captured - output is plausible prediction, not ground truth. Tools: Real-ESRGAN, Topaz Gigapixel, Waifu2x, Upscayl.