How AI upscaling works
Ordinary resizing stretches the pixels you already have, so a small image only gets blurrier. Real-ESRGAN is a super-resolution network trained on pairs of sharp and degraded photos: it predicts what the missing detail most likely looked like and draws it in, reconstructing edges, text and texture while it enlarges the picture by 2× or 4×.
Upload a JPG, PNG or WebP, pick the factor and download the result as a PNG. Large images are processed in overlapping tiles so the whole picture receives the same treatment regardless of size. For a smaller file afterwards, run the PNG through PNG to JPG.
What the model does well and where it struggles
The clearest gains come from photos and illustrations that are simply too small: thumbnails, old phone pictures, product shots pulled from a catalogue, logos saved at web resolution. JPEG compression blocks are smoothed away at the same time, which often improves an image even before the enlargement is considered.
The network invents detail, so it cannot recover text that is already unreadable or restore a face that is only a few pixels wide; in those cases it produces a plausible but not necessarily correct rendering. Heavy noise, motion blur and watermarks are enlarged along with everything else.
Typical uses
Print shops need at least 150 to 300 pixels per inch; a 4× upscale turns a 1000-pixel web image into one that prints cleanly at postcard size. Online sellers enlarge supplier images to meet marketplace minimum dimensions, and designers rescue logos and icons for which only a small copy survives.
Upscaled screenshots stay readable on high-resolution displays, and archive scans from old cameras gain enough resolution for slides and presentations.
Limits and file sizes
The output is capped at 16 megapixels: an image up to 4 megapixels can be enlarged 2×, and one up to 1 megapixel 4×. Larger inputs are rejected with a message that names the limit, because the result would exceed what most devices can open comfortably.
The result is written as a lossless PNG, which is several times larger than the source JPG. Compress Image shrinks it for the web without changing the pixel size. Processing runs on a GPU; a 4× upscale of a 1-megapixel photo takes a few seconds.