How it works
- Drop one or more photos.
- Pick 2x or 4x and a model for photos, natural detail, anime or speed.
- Download the enlarged images, or all of them as a ZIP.
Why use it
- Four Real-ESRGAN based models, from 5 MB to 32 MB, download once and are cached for offline use.
- Uses WebGPU when your browser has it, and falls back to WebAssembly.
- Batch several photos at once.
- No watermark, no daily limit, no account.
AI upscaling: choosing a model and knowing its limits
What an AI upscaler does
Ordinary resizing stretches pixels and fills the gaps by blending neighbours, which turns a small photo into a large blurry one. A super resolution network has instead learned from millions of pairs of small and large images what fine detail usually looks like, and draws edges, textures and hair at the new size. The result looks far sharper than stretching. It is a prediction, though, not a recovery: detail that was never in the photo is invented to look plausible.
That is why upscaling is right for making a small picture usable, for a print, a slide or a product page, and wrong for evidence. A blurry number plate or an unreadable face will come out sharp and possibly wrong.
Choosing the model
| Model | Best for | Speed and size |
|---|---|---|
| Photos, best quality | Most photos; restores texture, keeps edges clean | Slow on a CPU; about 33 MB |
| Photos, natural detail | Skin, fabric, foliage and fur, smoothing less | Slow on a CPU; about 33 MB |
| Anime and illustrations | Line art and flat colour, without invented texture | About four times quicker; 9 to 18 MB |
| Fast, any device | Screenshots, graphics and everyday photos | About ten times quicker; 2.4 MB |
On phones only the fast model is offered: the photo models would keep a phone busy for minutes. On a computer with a graphics chip that supports WebGPU, the photo models take seconds; on a processor alone they can take minutes for a large picture, so start with the fast model if you are unsure.
2x or 4x
The networks always enlarge four times; 2x is made from the 4x result, which looks cleaner than a network built for 2x. Pick the smallest that gives the pixels you need: 4x makes 16 times as many pixels, so a 1000 pixel wide photo becomes 4000 pixels wide and its file grows to match. For printing, the width in inches times 300 is the target; for screens, 2x is usually plenty.
Getting a good result
- Start from the best copy you have. Heavy JPEG blocks, noise and blur are enlarged too; a cleaner source gives a much cleaner result.
- Crop first. Upscaling only the part you need is faster and keeps files smaller.
- Save as PNG to keep every pixel of the result, or as JPG or WebP at high quality when size matters.
- Large pictures are cut into overlapping tiles and put back together, so memory stays manageable; very large results depend on your device's memory.
The model is downloaded once from this site and kept in the browser, so later runs work offline. To shrink the result for the web, use compress image; to bring it to an exact size, resize image.
If something goes wrong
- It is very slow. The model runs on the processor. Choose the fast model, or a browser with WebGPU on a computer with a graphics chip.
- The result looks painted or waxy. Try Photos, natural detail, which smooths less.
- Text or faces came out wrong. Small text and tiny faces are where invented detail shows most. Use the original, or a larger source.
Questions and answers
Does upscaling really add detail?
It adds believable detail. A super resolution model predicts sharper edges and texture from what it has learned, which looks much better than plain resizing. It cannot recover information that was never in the photo, such as an unreadable face.
Is my photo uploaded anywhere?
No. The model runs inside your browser. The only network request is the one-time download of the model file, and nothing about your photo is sent.
How large a picture can I upscale?
Large images are processed in overlapping tiles, so memory stays manageable. Very large results depend on how much memory your device has.