How to use it
- Pick the scale and output format, then drop images into the area above, click to choose them or paste them, up to 20 at a time. The first run downloads about 29 MB of runtime and model files from this site.
- Images are processed in order with progress and an estimated time left, and you can cancel at any time. Browsers with WebGPU use the graphics card and are much faster.
- Use Compare and drag the slider to check the result, then download one by one or all as a ZIP. After changing settings, choose Process again with current settings.
Choosing a scale
- 2x: the best balance of speed and quality for web, product and social images that start small.
- 4x: for very small icons, screenshots or old photos that need a big enlargement; the computation is the same as 2x for the same input, but the file is larger.
- 1x: when the size is already enough and you only want to remove noise, compression blocks and blur.
Frequently asked questions
Is it really free? What are the limits?
Yes. Use it as often as you like, with no sign-up and no watermark. The limits: up to 20 images per batch, each up to 40 MB; each result can be up to 16.78 megapixels (for example 4096×4096) and 8192 pixels on a side; without WebGPU the CPU is used and each input is limited to 1 megapixel.
Are my images uploaded to a server?
No. The Real-ESRGAN upscaling model (BSD-3-Clause licence) and the ONNX Runtime Web runtime (MIT licence) are downloaded from this site, checked against their file fingerprints and run in your browser; decoding, computation and export all happen on your device. The page shows site ads, which take no part in image processing.
Can AI upscaling recover the real detail in a photo?
It cannot bring back information that is no longer there. The model adds plausible texture and edges learned in training, so small text, logos and faces may differ from the real thing. Rely on the original for ID documents, evidence and product specification images.
How do “1x enhance” and 2x work?
The model is a 4x model. For 2x or 1x the image is first processed at 4x and the result is averaged down to the target size, as the model authors’ own tool does. 1x is useful for removing noise and compression blocks and sharpening edges without changing the size.
Why is it sometimes slow? What is WebGPU?
WebGPU lets web pages compute on the graphics card and is supported by current Chrome, Edge and other browsers. Without it the CPU is used, which can be dozens of times slower, so CPU mode limits each input to 1 megapixel. Large images are processed in tiles, and the list shows the estimated time left.
Can I upscale a PNG with a transparent background?
Yes. The colour is upscaled by the AI and the transparency channel separately with ordinary interpolation, then recombined; PNG and WebP keep the transparency. JPEG has none, so transparent areas are filled with white.
Is the result big enough to print?
At 300 dpi, 4000 pixels print about 33.9 cm (13.3 in) wide; a 1000×1000 image upscaled 4x becomes 4000×4000. AI-added detail can look unnatural up close, so proof important print jobs first.