AI Image Upscaler
Enlarge a picture two or four times with a neural network that invents the detail, in your browser. Nothing is uploaded.
Drop a small or blurry picture here
PNG, JPEG, WebP, HEIC, GIF, AVIF or BMP, up to 50 MB. Best on pictures under a megapixel; larger ones are reduced first.
About this tool
Makes a small picture bigger without making it blurry, in your browser. A plain resize spreads each pixel over four and softens every edge; the network here, Real-ESRGAN, has learned from millions of photographs what edges, textures and type look like at higher resolution, and draws them in. An old 640-pixel photo, a logo pulled off a website, a screenshot that has to be printed, a thumbnail that is all you have: each comes out at two or four times the size, with detail the original only hinted at. The model runs on your graphics card when the browser allows it and on the processor when not, and the picture never leaves the device.
How to use it
- 1Drop a picture, or paste one from the clipboard. Small ones benefit most; anything over a megapixel is reduced first.
- 2Choose 2× or 4×. The buttons show the size that comes out.
- 3Press Upscale. The first run fetches about ten megabytes of runtime and model and keeps them; later runs start at once. Then download the PNG, or send it on.
Good to know
What it does well, and what it cannot do
Photographs, artwork and text come out sharper, with clean edges and plausible grain, and JPEG blocking is smoothed away. What the network cannot do is recover what was never there: a face that is eight pixels wide becomes a convincing but invented face, and a licence plate that cannot be read at one size cannot be read at four times it either. It is a restoration of how the picture looks, not of what it recorded.
Why it works in tiles, and why there is a size limit
The network runs 34 convolutions over every pixel, so its memory grows with the picture. It is fed 192-pixel tiles with a border of context around each, and the borders are discarded, which is what keeps the seams invisible. A megapixel is about fifty tiles. Above a megapixel the picture is reduced first, both because four times each edge of a large photo is more pixels than some browsers will hold in a canvas, and because a large photo has the detail already; upscaling is for the small ones.
GPU or processor
With WebGPU, which current Chrome and Edge have and Safari and Firefox are gaining, the whole job takes a few seconds. Without it each tile takes about a second on the processor, and the page pauses while a tile runs. Before trusting the graphics card, the tool runs one small tile on both and compares them, and falls back to the processor if they disagree, so a driver quirk cannot hand you a wrong picture quietly. The file is read into memory by your own browser and processed there, so it is never uploaded and never reaches a server.