AI Image Upscaler
Four times the pixels, detail rebuilt by AI.
How to upscale an image
- Add a small image (JPG, PNG or WebP, up to 1024 px on the long edge). Typical inputs: old photos, game assets, tiny logos, thumbnails.
- The first use downloads the 63 MB AI model once; your browser keeps it for next time.
- The image is processed in tiles and you see the tile counter advance. On a graphics-capable browser this takes seconds; on plain CPUs expect a few minutes for large inputs.
- Compare the 4x result in the preview and download the PNG.
Frequently asked questions
- How is this different from simply resizing up?
- Classic resizing interpolates pixels, which blurs. Real-ESRGAN is a neural network trained to reconstruct plausible detail: edges stay sharp, textures are rebuilt rather than smeared.
- Why is the input limited to 1024 px?
- The output has sixteen times the pixel count of the input, and the network's memory use grows with area. 1024 px in means a 4096 px result, which already covers print and wallpaper sizes.
- Why is it slow on my device?
- Everything runs locally. Browsers with WebGPU use your graphics card and finish quickly; without it the network runs on a single CPU thread. The tile counter shows real progress either way.
- Does it help with text or logos?
- Yes, sharp-edged graphics upscale especially well. Heavily compressed JPGs also come out cleaner, because the model was trained on exactly that kind of degradation.
- Which model is this?
- Real-ESRGAN x4plus (BSD-3-Clause license), unmodified, running via onnxruntime in your browser. Your image never leaves your device.