Image Upscaler

Normal resizing, as in an image editor, calculates new pixels by averaging their neighbours, so a picture enlarged 4× looks soft and blurry. An AI upscaler has learned from many pairs of low- and high-resolution images what detail usually belongs between the pixels, and draws crisp edges, text and textures instead. It cannot recover information that was never in the picture, such as a face that is only a few pixels wide, but it makes the most of what is there.

The model works on tiles of 160 × 160 pixels at a time, and the work grows with the number of pixels in the picture. To keep the result within what browsers can hold, pictures larger than about 1.9 megapixels (for example 1600 × 1200) are first reduced to that size, so a 4× result is at most 6400 × 4800 pixels. On a typical computer a 1-megapixel picture takes one to a few minutes; phones are slower. The page shows an estimate before you start and the time left while it runs, and you can cancel at any moment.

Choose 4× for small pictures, such as thumbnails, icons and old web images, that need to become much larger. Choose 2× when the picture is already medium-sized and you need a little more resolution, for example to print it larger or to use it on a high-resolution screen; 2× results are made by upscaling 4× and then reducing with high-quality resampling, which gives a cleaner result than a plain 2× stretch. If the result is still too large for your needs, use the Image Resizer afterwards.

Most online upscalers send your picture to a server with powerful graphics cards. Here, the Real-ESRGAN model (4.6 MB) and the ONNX Runtime engine are downloaded once from this site and run on your own device. Your picture is never uploaded, which makes the tool suitable for personal photos and confidential images. The model is released by its authors under the BSD 3-Clause licence.