An AI upscaler makes an image bigger by predicting detail that a normal resize can't create. Stretching a small picture in an ordinary editor just spreads the existing pixels out, which looks soft and blocky. This tool uses Real-ESRGAN (realesr-general-x4v3, BSD-3-Clause licence), a model trained on real photos to rebuild sharp edges and textures and to clean up JPEG noise while it enlarges.
The model runs in your browser with onnxruntime-web. Your image is cut into 128 × 128 pixel tiles with a small overlap so no seams show, each tile is enlarged 4×, and the tiles are stitched back together; for 2× the result is drawn at half that size. The progress bar counts the tiles and estimates the time left. The model is small (about 5 MB, plus about 3 MB for the engine) and is kept in your browser after the first download. If the model can't run on your device, the run fails with a clear message and isn't counted: it never falls back to a plain resize.
Honest limits: AI upscaling invents plausible detail, so it can't recover text, faces or numbers that aren't really there, and very blurred images stay soft. To keep the result within what phones can hold in memory, inputs above 1 megapixel (for 4×) or 1.5 megapixels (for 2×) are reduced first, so the tool is best for small images: old photos, thumbnails, screenshots, logos and web images.