Upscale an image 2× or 4× with AI.
Make a small or low-resolution picture larger with sharper edges than stretching. The image never leaves your device.
Drop an image, or click to choose · Ctrl+V to paste
JPEG, PNG, WebP, AVIF, GIF, BMP, HEIC/HEIF (one image at a time, up to 50MB)
Your image is processed on this device and is never sent anywhere.
An AI upscaler that does not need your image
Most AI upscalers send the picture to their server, run the model there and send the result back, often after a sign-up and with a limit on how many you can do. Here the model is downloaded into your browser and run on your device, so the image never leaves it. Unreleased artwork, product photos or anything you would rather not hand to someone else's server can go in as they are. There is no account, nothing to pay and no watermark on the result.
What it can and cannot do
To be straight about it: AI upscaling does not recover lost detail. The model guesses what a shape would have looked like at a larger size and invents plausible detail to match. That is why small text on a sign does not become readable, and why a blurred face does not turn back into the person's face; the invented detail can differ from what was really there. Do not use it on an image that has to serve as evidence or a record. Where it helps is appearance: using a logo or an illustration at a larger size, or putting an old, small picture on a website.
What works well
| Image | Result | Why |
|---|---|---|
| Logos, icons, illustrations | Very good | Clear outlines are easy to guess right |
| A clean photo | Good | Texture comes through; the fine detail is invented |
| A heavily compressed image | Mixed | The blocky artefacts are enlarged along with it |
| Small text, faces to identify | No | Not made readable; a face can become someone else's |
Why there is a size limit
The result has the scale squared as many pixels: 4× is sixteen times the area. A 1500×1500 image at 4× is 6000×6000, thirty-six million pixels, and holding that on a phone is where the tab crashes. So 4× takes images up to about 1 megapixel and 2× up to about 4 megapixels, and files up to 50MB. Above a limit that scale's button cannot be pressed: a button you cannot press is better than a crash that loses the picture. The larger the image, the longer the run: small images take seconds, images near the limit much longer.
What comes out
It opens JPEG, PNG, WebP, AVIF, GIF, BMP and HEIC/HEIF, one image at a time. The result is always a PNG, so the new pixels are not compressed again, and the file name ends in -x2 or -x4. An animated GIF or WebP comes back as its first frame only. 2× and 4× use separate models, and the first upscale at each scale downloads its model, about 1MB.
What it does not do
- · Scales other than 2× and 4×. For an exact width and height, upscale first and then resize
- · Several images at once. One at a time; each run is the slowest thing on this site
- · Fix blur, noise or faces. It enlarges; it does not sharpen a shaky photo or restore a face
- · Keep animation. A moving image is enlarged as one still frame
How to use it (three steps)
- 1. Drag and drop an image (or click to choose, or paste with Ctrl+V)
- 2. Choose 2× or 4×; the size of the result is shown before you start
- 3. Press Upscale, check the result and press Save upscaled image
Frequently asked questions
Is my image uploaded anywhere?
No. The model is downloaded into your browser and the upscaling is calculated on your device. The image is never sent to a server; there is no upload endpoint to send it to, so there is also nothing to delete afterwards. Most AI upscaler sites work the other way round, which is the main difference.
How is this different from just resizing?
A plain resize fills the new pixels from the colours around them, so edges come out jagged or soft. This tool uses a model trained on many images to guess what a shape would have looked like at a larger size, and draws the new pixels from that guess. Edges and outlines stay crisp, and the result looks less obviously enlarged.
Does it enlarge an image without losing quality?
It loses less than stretching does, but it cannot add back what was never in the picture. The model invents plausible detail; it does not recover real detail. Small text that was unreadable stays unreadable, and a blurred face does not become that person's face again; the invented detail can differ from what was really there. Do not use it on an image that has to serve as evidence or a record.
How large can the image be?
2× works on images up to about 4 megapixels (around 2000×2000) and 4× on images up to about 1 megapixel (around 1000×1000), and a file can be up to 50MB. The result has the scale squared as many pixels, and past these sizes the browser tab runs out of memory. Above a limit, that scale's button cannot be pressed, and the page says why.
What kind of images work best?
Illustrations, logos, icons and anything with clear lines. Noisy photos, images already broken up by heavy compression and pictures with fine text are harder and can come out looking unnatural. Try it once and look at the result before you use it.
How long does it take?
2× and 4× each have their own model. The first upscale at a scale downloads that model (about 1MB), so it takes a moment; after that, that scale starts without the wait. The calculation itself depends on the size of the image and the scale, and the button shows the progress as a percentage.
What format is the result saved in?
PNG, whatever went in, so the upscaled pixels are not compressed again; the file is larger for that reason. The name ends in -x2 or -x4. An animated GIF or WebP is upscaled as its first frame only; the movement is not kept.
Is there a watermark or a sign-up?
No to both. Nothing is added to the image, there is no account, no credit system and no daily limit.
To change the width and height without AI, or to shrink an image that is over the limit first, use theimage resizer; to make the saved PNG lighter, theimage compressor; and to cut out the background, thebackground remover. All three run in the browser too.