Enlarge a small image
Choose a PNG, JPG or WebP image, select 2× or 4×, then process. The first run downloads Real-ESRGAN general-x4v3 and its runtime. Processing uses overlapping tiles and can take minutes on slower devices; progress counts completed tiles. You can cancel at any time.
Is this ordinary resizing?
The model predicts higher-resolution RGB detail at 4×. The 2× option downsamples that prediction. A 500 × 400 image becomes 1000 × 800 at 2× or 2000 × 1600 at 4×. Predicted textures are not recovered ground truth: text, faces and fine patterns can change or show artifacts. Always inspect the result.
Why is the input limited?
Input is limited to 1,048,576 pixels, 10 MB and 4096 pixels on the longest side to bound browser memory and processing time. Fourfold enlargement produces sixteen times as many pixels. The five-minute processing limit stops unfinished work; try a smaller image if your device cannot finish. A smaller file size alone does not guarantee a smaller pixel count.
Privacy and PNG output
The model runs locally without sending your photo to an API. Internet is needed to open the site and download the model. Original alpha is resized separately to retain transparency. Output is PNG without source metadata, and may be much larger than the input. Download before leaving, then use a compressor if needed.