scaling up
in English: upscaling
Upscaling is increasing the resolution of an image or video so that the program – nowadays typically artificial intelligence – fills in the missing details.
In plain terms
If you simply stretch a small image larger, it becomes blurry or blocky. The program still calculates new pixels, but only copies or averages the existing ones, so no new detail enters the image. Upscaling instead calculates new pixels that make the larger image appear sharp and detailed.
It is like a restorer completing a faded fresco: they have not seen the original, but based on knowledge of many similar paintings, they can likely guess what was missing. Similarly, AI-based upscalers have learned from countless images how a hair, a brick wall, or a letter should look in sharp detail.
This is also the most important difference compared to traditional upscaling: old methods only average neighboring pixels, while AI invents plausible details. The result is more impressive, but not necessarily what was actually there.
Upscaling is used to restore old photos and films, enlarge AI-generated images to printable sizes, and in televisions and video games.
How it works
- To train the model, sharp, high-resolution images are intentionally downscaled and degraded, creating image pairs: a low-quality and a high-quality version of the same image.
- The neural network learns from millions of such pairs how to return from the weak version to the good one, i.e., what patterns and textures usually lie behind a blurred spot.
- During use, the program analyzes your image in smaller segments, identifies edges, surfaces, and shapes, then calculates the new pixels between them.
- Generative upscalers (such as those based on diffusion models) go further: they partially redraw the image and add fine details that were not visible in the original at all.
- For video and games, the system also uses adjacent frames. Real additional information can be extracted from these, so the filled-in details do not flicker but remain consistent during motion.
An everyday example
Your grandparents' small, grainy wedding photo surfaces, and you want to enlarge and frame it. You upload it to an image enhancement application, which returns a sharp image four times larger – but it is worth checking whether the faces still resemble them.
Who is affected
A common misconception
Limits and risks
- The added details are fabricated, so faces, text, and numbers in the image may change: the gaze will differ, and text may appear readable but will be incorrect.
- Good images cannot be created from very poor-quality sources; excessive upscaling results in plastic-looking skin, paint-like surfaces, and strange patterns.
- For medical, regulatory, or evidentiary purposes, an upscaled image can be misleading because genuine and hallucinated details cannot be distinguished.
- If you upload family or company photos to an unknown online service, check what happens to them, as the photos are stored on the provider's server.
Similar, but not the same
- scaling (traditional zoom) – Resizing merely stretches and averages existing pixels, while AI-based upscaling also fills in new details. Everyday usage sometimes refers to traditional zooming as upscaling.
- image generation – Image generation typically creates entirely new images from text descriptions, while upscaling enlarges and sharpens an existing image.
- frame interpolation – Frame interpolation calculates new frames between two video frames for smoother motion, while upscaling increases the resolution of individual frames.
Related concepts
generative AI, image processing, Computer vision
video upscaling, face restoration
image generation, diffusion model, noise reduction, deepfake, hallucination
A brief history
Image upscaling has long been done via mathematical interpolation. Deep learning entered the field in 2014 with the SRCNN neural network, while ESRGAN, released in 2018, could already reproduce lifelike textures.
Where things stand
By the end of 2026, upscaling is a built-in feature of phones, TVs, video games, and image editors, and the usual final step for AI image and video generators. Development focuses on real-time, flicker-free upscaling of videos and improving the fidelity of details.
In the news
- Holyvolt raises $45 million for AI-based battery materials as PNNL-led group applies AI to scale-upOct 26, 2026 · Ipar
- L'Oréal rethinks its marketing engine as AI reshapes the consumer journey; mid-size advertisers scale AI creativeOct 5, 2026 · Marketing
Also known as
scaling up · upscaling · upscale · AI upscaling · AI upscaling · super-resolution · super-resolution · resolution enhancement
Sources
- Dong et al.: Image Super-Resolution Using Deep Convolutional Networks (SRCNN, arXiv)
- Wang et al.: ESRGAN – Enhanced Super-Resolution Generative Adversarial Networks (arXiv)
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.