Image upscaler

Bigger, without the blur

Ordinary resizing spreads the existing pixels over a larger area, so enlarged photos look soft and blocky. An AI upscaler instead predicts what the missing detail should look like: edges stay crisp, textures stay textured, and compression blocks are cleaned up.

Choose 2× to double the width and height (four times the pixels), for example to print a phone photo larger or use an old picture on a modern screen. Choose 4× for small images such as thumbnails, icons or screenshots.

The model works best on photos and natural images. It can't recover details that were never captured, so a tiny face won't gain an identity, but the result looks clean and natural. Photos larger than 2 megapixels are first reduced to that size, then upscaled.

On your device

The AI model runs in your browser, on your graphics card when available; it's small (5 MB) and downloaded once. Your photos are never uploaded, there's no watermark, and you can upscale five at a time.

The photo is processed in small tiles by Real-ESRGAN (BSD-3-Clause), an open-source super-resolution network that outputs four times the pixels of each tile. Tiles overlap and only their centres are kept, so no seams appear. ONNX Runtime runs the model on your graphics card (WebGPU) or processor. The 4× result is then resized smoothly to the factor you chose and saved as JPG at quality 95, or PNG if the picture has transparency.