← AI Terminology

Super-Resolution

Super-resolution (SR) reconstructs a higher-resolution image from a lower-resolution input, using classical priors or learned generative models.

Widely used in photography, video, medical, and satellite imaging.
Why It Matters in AI
Sensors and bandwidth limit resolution; SR hallucinates plausible detail. Modern diffusion/GAN SR improves consumer photos and restores media — with ethics around forensic trust.
Key Points
Aspect Description
Use Enhance crops, upscale generations, restore media
Eval PSNR/SSIM + perceptual metrics/human study
Risk Invented detail; not true information gain
Task Low-res → high-res image
Methods CNNs (EDSR), GANs (Real-ESRGAN), diffusion SR
Related Inpainting, generative models
Simple Analogy
Zooming into a photo and intelligently inventing crisp texture where only blur existed — sharper, not magically true new evidence.
Common Usage Examples
  • Real-ESRGAN upscaling pipelines
  • Diffusion upscalers in art tools
  • Satellite imagery enhancement
  • Disclose SR in forensic contexts
Summary
In short: Super-resolution upscales images with learned detail — enhancing resolution for media and vision pipelines, with hallucinated texture caveats.