Resize an image with AI
Normal enlarging just stretches the pixels a picture already has, so it goes blurry; here an AI model (Swin2SR) runs inside your browser and fills in likely detail instead. Shrinking does not need the model, and your photo is never uploaded.
- Runs entirely in your browser
- Works offline
- No upload
- No sign-up
- No watermark
Loading the tool…
How it works
- Load the model once. It is about 53 MB and stays saved in your browser afterwards.
- Choose a picture and type the width you want. Up to 2048 px the AI model does the work; if you ask for less than the original size, it is a plain shrink.
- Wait while it works through the image piece by piece, then download the result as a PNG.
Why nothing is uploaded
Every operation on this page is done by code running inside your browser tab, using the same engine that renders web pages. The file is read from disk into your tab’s memory, transformed there, and written back out as a download. It is never sent anywhere — not to us, not to a third party.
Verify it yourself
- Open your browser’s developer tools (F12) and select the Network tab.
- Load your file and run the tool.
- The only requests you will see fetch the tool’s own code — and, for a few heavy tools, their open-source engine from a public CDN — plus one small page-view ping to loreatec.jp (page address and title, nothing more). None of them carry your file.
Frequently asked questions
How is this different from the plain resize tool?
For shrinking there is no difference, and the plain tool is quicker; this page does not use the model for smaller sizes, and says so. The difference is enlarging: a plain resize stretches the pixels it already has, while this runs a trained model that adds detail the small version never contained. In our test (a photo shrunk to a quarter, then enlarged back) it measured 23.87 dB against 22.95 for the browser’s own enlarging — a small but real gain (PSNR, a sharpness measure).
Why can it not go past 2048 px?
The model enlarges by four, so 2048 px of output means 512 px of input; beyond that the browser runs out of room and stops. To stay safe it works on the picture in small tiles (192 px), and the time grows with the size: 2048 px wide took a bit over a minute on our machine, and smaller targets are quicker because they need fewer tiles.
Is my photo uploaded?
No. The model is downloaded to you; your photo is not sent to us. If you want proof, load the model and switch your device to flight mode: it keeps working. (For technical users: the browser’s Network tab shows only the one-off model download.)
Can it recover an unreadable number plate or face?
No, and be wary of anything that claims otherwise. The model sharpens edges and removes compression blockiness based on what it learned from other images; it cannot recover detail that was never captured. What it produces is a plausible enlargement, not evidence.
Which model, and under what licence?
The “real-world” 4× version of an open super-resolution model, Apache-2.0 licensed — the exact name is on the licences page. We picked it by measuring: the “classical” and “compressed” versions of the same family, and another well-known candidate, all scored below the browser’s plain enlarging on real photographs.