Upscaling

In short

Upscaling is enlarging an image or video to a higher resolution while trying to keep it clear. AI upscalers estimate fine detail that was not recorded, rather than merely stretching each pixel.

Also called: Image upscaling, Video upscaling, Super-resolution
Topic: Images, video, voice and music
Level: Beginner
First described: Chao Dong and colleagues introduced SRCNN, an influential deep-learning approach to single-image super-resolution, in 2014.

A conventional resize calculates new pixels from nearby pixels. AI upscaling instead learns a relationship between lower-resolution and higher-resolution examples, then predicts a detailed version of the input. The result can have sharper edges and textures, but predicted detail is not the same as detail captured by the original camera¹.

In 2014, Dong and colleagues introduced SRCNN, a deep convolutional network that learned an end-to-end mapping from a low-resolution image to a high-resolution one. Their paper framed single-image super-resolution as reconstructing a larger image from a smaller input¹.

Upscaling does not automatically fix every quality problem. Blur, compression blocks, noise and motion may be amplified or interpreted incorrectly, and a model can create textures that look realistic but are not faithful. Video adds a consistency problem because the estimated details should not flicker from frame to frame.

Upscaling appears in photo and video software, archive restoration and online publishing. Topaz Video describes models that denoise, sharpen and upscale old footage to 4K, while its web tool offers 1080p and 4K output options². The same broad idea can enlarge a small family photograph, game asset or social-media clip.

Example

A low-resolution scan of a printed family photograph is enlarged before being put into a digital album. An AI upscaler may make faces and clothing edges clearer, but it cannot know the exact detail erased by the scan.

Why it matters

Upscaling can make old or small media more useful on modern screens. It can also produce convincing invented detail, so enlarged images should not be mistaken for higher-fidelity originals.

Questions people ask

Does upscaling increase the original detail?

It increases the number of output pixels and can estimate plausible detail, but it cannot retrieve information that the source never contained. AI methods may invent texture or edges while making the result look sharper¹.

Is upscaling the same as resizing?

Both change dimensions. Resizing uses a mathematical interpolation rule, while AI upscaling uses a learned model to estimate detail and may also denoise or sharpen the result¹².

Can video upscaling turn any video into 4K?

Software can render a larger 4K file from many lower-resolution videos, but the visual improvement varies with blur, noise, compression and motion. A 4K output file is not proof that the source contained 4K detail².

Related terms

Tools that use it

Sources

Back to the glossary