AI Image Upscaler Beta
Enlarge images with reconstructed detail, not blur.
Worth knowing before you startA 4.6 MB model downloads once and is cached. Processing is tiled, so a large image can take 10 to 60 seconds. Inputs above 768 px on the long edge are downscaled first to keep memory in check.
Choose image
or drop it anywhere on this page
JPG · JPEG · PNG · WEBP · AVIF · GIF · BMP · TIFF · TIF · SVG · +3
Processed on your device — never uploaded
How to ai image upscaler
- Add an image — smaller inputs give the most dramatic improvement.
- Pick 2× or 4×. The model downloads on first use and is then cached.
- Wait for the tiles to process, then download.
About this tool
Super-resolution is one of the clearest wins from small neural networks. The Real-ESRGAN family was trained specifically on the kinds of degradation real images suffer — compression artefacts, sensor noise, prior downscaling — so it does more than sharpen: it reconstructs edges and texture that interpolation simply cannot invent.
Running it locally keeps a genuinely useful capability free. The compact model used here is only a few megabytes, and tiled inference keeps memory bounded so even large images process without exhausting the tab. Old family photographs, low-resolution logos and compressed screenshots are where it shines hardest.
Questions
How is this different from just resizing?
Ordinary resizing interpolates between existing pixels, so enlarging produces a bigger, softer copy. A super-resolution network has been trained on millions of image pairs and reconstructs plausible edges and texture instead. On photographs and artwork the difference is substantial.
Can it recover detail that was never there?
No — and be sceptical of anything claiming otherwise. The model produces convincing detail, but it is inferring what was probably there, not recovering ground truth. That makes it excellent for photos and illustrations, and a poor choice for anything forensic, like reading a licence plate.
Why is it slow?
Neural upscaling is genuinely heavy computation, and it is running on your CPU or GPU rather than a datacentre rack. The image is processed in overlapping tiles to stay within browser memory limits. Browsers with WebGPU enabled are dramatically faster.