Super-Resolution
A deep learning technique that infers and generates high-resolution images from low-resolution inputs by reconstructing detail not present in the source.
Super-resolution (SR) generates high-resolution output from low-resolution input by leveraging learned statistical patterns to synthesize high-frequency detail (edges, textures) absent from the source. Unlike simple interpolation, SR models produce sharp, plausible detail.
Landmark architectures include SRCNN (2014), ESRGAN (2018), and Real-ESRGAN (2021). Real-ESRGAN is trained on data with real-world degradations (JPEG compression, noise, blur) and achieves practical quality at 4x magnification for photographs and illustrations.
Diffusion model-based SR approaches produce even more natural textures, though generated details are plausible estimates rather than ground truth. This distinction matters for forensic or medical imaging where accuracy is paramount. Try super-resolution with the image upscaling tool.