Image upscaling

Classic enlargement relies on interpolation: to fill the new pixels, the software averages the values of the nearest ones. Bilinear and bicubic methods are quick and predictable, but they cannot invent detail, so a large enlargement looks soft. An AI upscaler is a neural network trained on pairs of the same picture at low and high resolution. It has seen how edges, skin, fabric and foliage usually look when sharp, and it draws that kind of texture into the enlarged result. Some upscalers are built on diffusion models, which rebuild detail step by step.

The detail is plausible, not recovered. The network adds what is likely, and it cannot know what the original scene really contained. That is fine for a generated picture or a landscape, but it can change a face slightly, smooth a texture too much or garble small lettering, so a quick check of the result is wise. Fine repeating patterns such as brick walls, striped shirts or fences can turn into waves or moiré after enlargement, so zoom in on them first. For the same reason, upscaling is not a way to read a blurry license plate or to prove anything about a photograph.

The factor tells you how much the dimensions grow. A 2x upscale doubles the width and the height, which gives four times as many pixels, and 4x gives sixteen times as many. Larger factors ask the network to invent more, so quality varies. A sensible routine is to settle on the composition first, upscale last and keep the original file. Many people also reduce noise before enlarging, because an upscaler tends to treat grain as texture worth preserving. Photographers use it to prepare prints, game artists to rescue old textures and designers to enlarge a logo cropped from a thumbnail. It is also why a fast generator pairs well with upscaling: make a quick draft at modest size, then enlarge only the one you keep.

In the Studio, the upscale tool is part of what the image generator page lists for paid plans, next to background removal and editing by instruction, and it enlarges an image two times. The API has a matching endpoint, POST /image/upscale, which enlarges an image two or four times and returns the image itself, as PNG unless you ask for another output format. The image is first checked against NeverNo's hard limits, and the call is billed in credits like any other.

Z-Image Turbo, the default Studio model, makes a typical 1-megapixel picture in about 3 seconds, measured on Runware, so you can try several drafts and upscale only the best. The images you make are not kept in a gallery on our servers: they are stored encrypted in your browser, in the database that holds your history, and the API returns images as base64 for you to save yourself.