What an image upscaler does
Upscaling makes an image bigger: more pixels wide and tall. The hard part is filling in the new pixels. A 500 × 500 photo enlarged to 2000 × 2000 needs 16 times as many pixels, and 15 of every 16 have to be invented.
There are two ways to invent them, and this tool offers both:
| Method | How it fills new pixels | Looks | Speed | Download |
|---|---|---|---|---|
| Fast (Lanczos) | A weighted average of nearby pixels, using the Lanczos-3 filter | Smooth, accurate, but soft; no new detail | Instant | Nothing |
| AI (Swin2SR) | A neural network predicts the most likely detail from patterns it learned | Sharper edges, crisper text, cleaner texture | Seconds with WebGPU, minutes on a CPU | 7–53 MB once |
Everything runs in your browser, with no account and no daily cap. AI mode downloads the transformers.js library (jsDelivr) and the model (Hugging Face) once; both are cached. Your image never leaves your device, unlike upscalers that upload photos or charge credits for 4×.
How the AI upscaler works
The AI is Swin2SR, a published super-resolution model built on the Swin Transformer and released under the Apache 2.0 licence. It looks at the picture in small windows of 8 × 8 pixels, compares each window with its neighbours, and predicts a larger, sharper version. Two versions are available:
- Best quality: the real-world ×4 model. It was trained on photos that had been deliberately blurred, noised and JPEG-compressed, so it learned to clean up the kind of damage real photos have, not just enlarge clean ones. About 21 MB on a processor (8-bit weights), 32–53 MB on a graphics card.
- Fast AI: the lightweight ×2 model, about a tenth of the size and several times faster on a processor. For 4× it simply runs twice. About 7–8 MB.
Large images do not fit through the network in one go, so the tool cuts them into overlapping tiles of 128 × 128 pixels. Each tile is enlarged, the overlapping edges are thrown away, and the centres are stitched together. Because every kept pixel had neighbours on all sides when it was processed, you do not see seams. Transparent PNGs keep their transparency: the network enlarges the colours and the original alpha channel is smoothly enlarged and put back.
Where it runs: WebGPU or WebAssembly
If your browser supports WebGPU (current Chrome and Edge, and other browsers as they enable it), the model runs on your graphics card and a typical photo takes a few seconds. Otherwise it runs on your processor with WebAssembly, which always works but is much slower: in our tests on a laptop, one 128 × 128 tile took about 4 seconds with Fast AI and 29 seconds with Best quality. The tool tells you before starting if a job will take more than a few seconds, and you can cancel at any time.
Worked examples
A small product photo for a shop. A 600 × 600 JPG downloaded from a supplier looks blurry on a 1200-pixel product page. Upscale 2× with AI: the result is 1200 × 1200, with sharper edges on the product and less visible JPEG blocking. On a processor with Fast AI, that is 36 tiles, roughly two and a half minutes; with WebGPU it takes seconds.
An old scanned photo for printing. A family photo scanned at 800 × 600 pixels would print at only 160 pixels per inch at 5 × 3.75 inches. Printers like about 300 pixels per inch for sharp prints:
Upscale 4× to 3200 × 2400 and it prints at about 10.7 × 8 inches at 300 PPI. The AI cannot restore detail the scan never captured, but it gives a sharper base than enlarging in print software.
A screenshot or digital drawing. Line art and interface screenshots have hard edges that ordinary resampling blurs. AI keeps lines crisp. For pixel art, where you want blocky pixels, use Fast at 2× or 4× instead; it keeps the image faithful without inventing smooth curves.
Choosing the right settings
- 2× or 4×? Pick the smallest scale that gets you to the size you need. 4× takes longer and makes huge files.
- Best quality or Fast AI? With WebGPU, choose Best quality. On a processor, start with Fast AI; use Best quality for small images where it matters most, such as a face or a logo.
- PNG or JPG? PNG is lossless and keeps transparency, but a 4× photo can be 20–40 MB. JPG at quality 92 is usually 5–10 times smaller and looks the same. You can also convert afterwards with JPG to WebP for websites.
- Huge results? The AI output is limited to 40 megapixels (for example 1500 × 1000 at 4× or 4000 × 2500 at 2×) so it fits in browser memory; the Fast method goes to 64 megapixels. Crop away what you do not need first with Crop Image.
What upscaling cannot do
Upscaling is a good guess, not a recovery. An AI upscaler makes images look sharper by drawing the most likely detail, so:
- Tiny text and faces made of a handful of pixels become sharp-looking guesses, which may not match the real letters or features. Never use an upscaled image as evidence or to read a number plate.
- Heavy blur from camera shake or bad focus is only partly improved.
- Official photos for passports and visas should come from a real, sharp photo. Upscaling a tiny photo can make it look artificial, and many offices reject edited images. Use Passport Photo Maker with an original photo instead.
After upscaling
Drag the comparison slider across faces, text and edges to compare the original (left) with the result (right), or click Zoom 100% to inspect real pixels. Then:
- make the file lighter with Compress Image,
- set an exact pixel size with Resize Image,
- or cut out the subject with Background Remover, which works better on larger images.