What does the Image Upscaler do?
The Image Upscaler makes a picture two or four times bigger without the blur you get from stretching it. Drop in a 600 × 600 product photo and you get back a 2400 × 2400 one, with edges, lettering and texture drawn crisply at the new size instead of smeared across it. The work is done by a super-resolution model: a small neural network trained to guess what the sharp version of a soft image looks like.
Most people arrive in the same spot. A supplier sent a 600-pixel product shot, the marketplace wants 1,600 pixels or more, and the listing tab is still open. Or it's an old profile photo, a logo that only exists as a thumbnail, a screenshot that goes fuzzy at full width. You need to increase image resolution, and you need the new pixels to look like they were always there.
The model runs on your own device, inside the browser tab. Your image is never uploaded. There's no account, no credit counter and no watermark on the download. You drop the file, it starts at 4× on its own, and a progress bar counts the tiles until it's done. That's the whole tool.
How AI upscaling works, and why a resize can't do it
A normal resize only has the pixels you gave it. To take a 600-pixel image to 2400 pixels, it spreads every source pixel across a 4 × 4 block and blends the seams between neighbours. Nothing new is added, so every edge turns into a soft ramp. Bigger, yes. Sharper, never. Our own Image Resizer does exactly this when you enlarge, and it tells you so.
A super-resolution model works from experience instead. It was trained on pairs of images: a sharp original, and a copy that was shrunk, blurred and compressed on purpose. After enough pairs, it learned what a crisp edge, a strand of hair or a line of printed letters usually looks like once it has gone soft. Then it draws that at the larger size.
This tool runs a compact, general-purpose model from the open-source Real-ESRGAN family. It's a 2.4 MB file your browser downloads once. The model is natively 4×. For 2×, the tool runs it at 4× and then halves the result exactly.
Why your image is cut into tiles. The model works on 256 × 256 pieces of the source, each padded with 34 pixels of context on every side that get cropped away afterwards. 34 pixels is how far the network can see from any single pixel, so the stitched result matches a single pass over the whole image. No seams, no faint grid. An 800 × 600 photo is 12 tiles, and the progress bar counts them: Upscaling… 7 of 12.
Where your browser supports WebGPU, the model runs on your graphics card. Where it doesn't, it runs on your processor in a background worker, so the page and the Cancel button stay responsive. If the graphics card fails partway through, the tool quietly starts that image over on the processor. It never hands you a plain resize with an "upscaled" label on it. If the model can't run at all, you get an error and a Try again button, not a fake result.
How big your image gets: common upscales
However you increase image resolution here, the output is always exactly the source size times the scale. There's one ceiling. The result has to fit inside 16,777,216 pixels (a 4096 × 4096 square), and no side can pass 8192. That's the largest image the strictest mobile browsers can hold. Past it, a browser doesn't throw an error. It quietly saves a blank file. So the tool checks the size before it runs anything.
| Source | Typical image | At 4× | At 2× | What happens when you drop it |
|---|---|---|---|---|
| 600 × 600 | Supplier product shot | 2400 × 2400 | 1200 × 1200 | Runs 4×, 9 tiles |
| 640 × 480 | Old webcam or early digital camera | 2560 × 1920 | 1280 × 960 | Runs 4×, 6 tiles |
| 800 × 600 | Small blog or catalogue photo | 3200 × 2400 | 1600 × 1200 | Runs 4×, 12 tiles |
| 1024 × 1024 | The largest square 4× allows | 4096 × 4096 | 2048 × 2048 | Runs 4×, 16 tiles |
| 1080 × 1080 | Square social post | Too large | 2160 × 2160 | Switches to 2×, 25 tiles |
| 1920 × 1080 | Full HD screenshot | Too large | 3840 × 2160 | Switches to 2×, 40 tiles |
| 2048 × 2048 | The largest square 2× allows | Too large | 4096 × 4096 | Switches to 2×, 64 tiles |
| 4000 × 3000 | 12-megapixel phone photo | Too large | Too large | Refused: "This image is already 12.0 MP." |
The 1080 × 1080 row catches people out. It's barely bigger than the 1024 square, but at 4× it would be 4320 × 4320, just over the ceiling. The tool doesn't make you guess. It picks 2× on its own, runs it, and says so: 4× would be too large for a browser, so this is 2×. The 4× button is greyed out, and hovering it tells you why.
A Full HD screenshot at 2× lands at 3840 × 2160, which is exactly 4K. That's the most useful upscale on the list if you're making a slide, a wallpaper or a video frame.
One thing that surprises people: 2× isn't faster. The tile count is the same at both scales, because the model always runs at 4× and the halving happens afterwards. Pick the scale by the size you need, not by the wait.
How to upscale an image for a product listing
Here's the kitchen-table case from start to finish. The file is blue-mug.jpg, 600 × 600 pixels, and the marketplace wants at least 1,600 on the long side.
- Drop blue-mug.jpg on the page, paste it from your clipboard, or use Choose an image. Upscaling starts at 4× the moment it lands. The very first image also fetches the model and runtime, which is a one-time wait.
- The progress bar counts 9 tiles. If you dropped the wrong photo, Cancel stops at the next tile and keeps your image on screen.
- Drag the divider across the picture. Before on the left is your original, stretched to the same size the way any resize would. After on the right is the model. The arrow keys move the divider too. Look at the handle and the rim of the mug. That's where the difference shows.
- Beside the buttons you'll see 600 × 600 → 2400 × 2400 · JPG, followed by the file size. Click Download and you get blue-mug-upscaled-4x.jpg, saved as a JPG at quality 0.92.
2400 pixels clears the 1,600 minimum with room to spare. If a platform wants one exact size, run the result through the Image Resizer afterwards. Shrinking a sharp 2400-pixel image down to 1600 stays sharp. Stretching a 600-pixel one up to 1600 never does. If the upload form complains about file weight, the Image Compressor trims it without changing the dimensions.
Why most upscalers make you pay per picture
Search for an image upscaler and you'll meet the same pattern over and over. You upload your photo. It's processed on the company's graphics cards. Then the full-resolution result sits behind something: an account, a watermark on the free version, a daily cap, a pack of credits, or a monthly plan with "AI credits" bundled into it. Often there's a free trial, sized to run out one image before you're done.
That isn't villainy. It's arithmetic. Every image costs them server time, so every image has to be metered, and your photo has to sit on their machine while it happens. The whole business is a GPU bill wearing a web page.
Compact super-resolution models now run fast enough inside a browser tab that the server can simply go away, and the bill goes with it. The Image Upscaler doesn't pay anyone per image, so there's nothing to meter. The first download on your first visit is the full-resolution 4× file. No sign-up screen stands between you and it.
The same goes for the other route people take: installing a desktop app to fix one photo, once. The Image Upscaler is the model, the tiles, a before-and-after divider and a download button. Open it, use it, close it.
What upscaling can't fix
An image upscaler predicts detail. It doesn't recover detail the camera never captured. For a product shot, a thumbnail or an old profile picture, that prediction is usually far better than any resize. But it has limits worth knowing before you rely on it.
- Small text turns into confident squiggles. If you couldn't read it before, the model will make it look sharp, not correct.
- Motion blur, missed focus and heavy noise leave too little for any model to work from. You'll get a bigger version of the same problem.
- It's an upscaler, not a photo enhancer. It doesn't restore faces, colourize old prints or remove scratches. Those are different jobs.
- Tiny sources are refused. Every side needs at least 16 pixels. A 12 × 12 icon gets That's 12 × 12 — too small to work from.
- Files over 50 MB are turned away before anything is decoded.
- HEIC won't open in a browser. Convert it with HEIC to JPG first. GIF and SVG aren't accepted either. Use a PNG, JPG or WebP.
Don't use it for anything that has to be true. ID and passport photos, license plates, evidence, a face you need to identify. The model draws what's plausible, and plausible isn't the same as what was there. The tool says this on the page as well.
Related image tools
Upscaling is usually one step in a longer chain. The Image Cropper is worth running first: cut away what you don't need, and every remaining pixel gets the model's full attention. A smaller crop is also more likely to fit under the 4× ceiling. Afterwards, the Image Resizer sets an exact size and the Image Compressor brings the file weight down. If a site only takes one format, the Image Format Converter switches between PNG, JPG and WebP. And if the product needs a clean backdrop, the Background Remover handles that part.
Frequently asked questions
How big can I print an upscaled photo?
Divide the pixels by 300 for a sharp print you'll hold in your hand. A 2400 × 2400 result prints at 8 × 8 inches. A 3200 × 2400 result prints at about 10.7 × 8 inches. Posters seen from across a room hold up at 150 pixels per inch, which doubles those sizes. The upscale gives the printer real pixels to work with. The model's usual limits still apply up close.
Is upscaling the same as increasing the DPI?
No. DPI is a number stored in the file that tells a printer how big to make it. Changing it doesn't add a single pixel, so a 600 × 600 image "at 300 DPI" is still a 2-inch print. Upscaling changes the pixels themselves: 600 × 600 becomes 2400 × 2400. If a form asks you to increase image resolution, it almost always means more pixels.
Does the upscaled file keep my photo's location and camera data?
No. The download is drawn fresh from a canvas, and camera tags such as GPS location, date and phone model aren't copied into it. If you want to strip those tags from the original without changing its size, the EXIF Remover does just that.
Why did my WebP come back as a PNG?
Some browsers can read WebP but can't write it. When that happens, the tool saves a PNG instead and says so: Your browser can't save WebP, so this is a PNG. The file is named for what it actually is, so nothing gets mislabelled. Everywhere else, PNG stays PNG, JPG stays JPG and WebP stays WebP.
Why is the upscaled file so much bigger?
Because it has far more pixels. At 4×, every source pixel becomes sixteen, so a 0.4-megapixel image turns into a 5.8-megapixel one. PNGs are saved without losing anything, which makes them the heaviest. JPGs and WebPs are saved at quality 0.92. If the size matters more than the last bit of detail, run the result through the image compressor.
Can I upscale more than one image at a time?
One at a time. Drop the next image whenever you like, and the run in progress stops so the new one can start. Nothing from the first image ends up in the second. When you're done with a picture, Start over clears the page.
Does it work on a phone?
Yes. It runs in a mobile browser the same way it runs on a laptop, and the image stays on the phone. The ceiling of 16,777,216 output pixels comes from mobile browsers in the first place, so the size limits are already set with your phone in mind. The Copy image button only appears where the browser supports copying pictures. Everywhere else, use Download.