Glossary
What is AI upscaling?
In short
AI upscaling raises the resolution of an image or a video with a neural network that does not just stretch pixels but draws in believable small details.
Images, video, audio · upscaling
MMOOZZGG.com editorsChecked: 1 min read
For a child
If you enlarge a small picture from a book on a copy machine, it becomes big but blurry, as if you look through wet glasses. Now imagine an artist who looks at the small picture and redraws it large, adding every eyelash and blade of grass with care. Such an artist lives inside a computer too. It makes small pictures big and sharp.
In plain words
A usual editor stretches each pixel when you enlarge a photo, so the picture looks soft. A neural network works differently. It was trained on pairs of images: a large sharp one and the same one reduced and damaged. So it learned how hair, fabric, leaves and letters usually look up close. Given a small image, it adds the missing details and returns a version several times larger. People use it to restore old family photos, prepare phone pictures for print and improve old video recordings.
For business
Neural enlargement is cheaper than a new photo shoot. From an old archive, supplier photos or camera frames you can make images for print, online marketplaces and large screens. Video can be raised to high resolution for ads and training videos. The cost depends on volume: one image is processed fast even in the cloud, but video needs a lot of computing, and for large archives your own graphics card pays off. The main risk is invented detail. The model can change a label, facial features or the texture of fabric, so the result is not fit for products, documents or evidence without a check. On medical and technical images, drawn-in details must never count as real.
For a specialist
AI upscaling (image super-resolution, SR) rebuilds a high-resolution image from a low-resolution one. One of the first deep models was described by Dong et al. (2014): a convolutional network is trained end to end, and the authors showed classic sparse coding methods as its special case. Then adversarial networks raised the quality. Real-ESRGAN (Wang et al., 2021) is trained only on synthetic pairs made by a high-order degradation model with blur, noise, compression and ringing artifacts, and a U-Net discriminator with spectral normalization keeps training stable. The diffusion model SR3 (Saharia et al., 2021) turns noise into an image step by step, conditioned on the low-resolution input. In an eightfold face enlargement test, people were wrong almost half the time, against no more than 34% for GAN models. PSNR and SSIM metrics reward smoothness, so sharpness is also judged with perceptual metrics and by people.
Example
An online store got product photos from a supplier the size of a postage stamp. A neural network enlarges them for product pages, and a content manager checks the text on the packaging against the original and fixes what the model distorted.
Related terms
Questions and answers
Can you enlarge video this way?
Sources
Not sure whether your company needs this?
AI solutions map