Image upscaling (also called super-resolution) is the process of producing a larger, sharper version of an image by rebuilding detail that was never captured — or was lost to compression — rather than spreading the existing pixels over a bigger canvas. Classic resampling makes a small photo big and soft; an upscaling model makes it big and readable: edges stay hard, texture comes back, and the picture holds up at print or 4K size.
How it works
The model is trained on pairs of images: a high-resolution original, and a copy of it degraded on purpose — downscaled, blurred, noised, crushed by JPEG. Learning to reverse that degradation millions of times teaches it what a hard edge, a strand of hair, a woven fabric or pore-level skin texture looks like, so at inference it can infer plausible high-frequency detail from a low-resolution input instead of averaging between neighbouring pixels the way bicubic resampling does. That is also the honest limit of the technique: detail is reconstructed from what the model has learned, not recovered from the original scene. Fine print, licence plates and faces reduced to a handful of pixels come back as a plausible reading, not necessarily the true one.
What it's best at
- Small and compressed photos — phone shots saved and re-saved through chat apps, images pulled off an old website, screenshots scaled down years ago.
- Print and large displays — taking a web-sized image up to the resolution a poster, a product page or a 4K screen expects.
- Product and catalogue images — legacy shots that need to match the sharpness of newly photographed ones.
- AI-generated images — lifting a generated picture to delivery resolution with its composition untouched.
What a good input includes
Start from the least-damaged copy you have: the original file rather than a screenshot of it, and the version from before it went through a chat app. One clear problem upscales better than four stacked — small is an easy case; small, blurred, noisy and watermarked is a hard one. Do not sharpen or denoise before upscaling, because those filters strip out the very cues the model reads. And keep the boundary with photo restoration in mind: torn corners, creases and stains are damage rather than missing resolution, and want a restorer instead.