What the upscaler actually does
This guide explains how the upscaler enlarges an image with its two methods, AI (Real-ESRGAN) and Classic, what result to expect from each scale factor, and when each method is the better choice. Both run entirely on your device; the sections below are honest about what each one can and cannot do.
Enlarging an image means creating new pixels between the existing ones. The default AI method does this with Real-ESRGAN, an open-source super-resolution model (BSD-3-Clause) trained on pairs of sharp and degraded photos. It enlarges the image 4× in small tiles and rebuilds edges and texture, then scales to the factor you chose; 1× keeps the size and uses the model only to clean up and sharpen. The model is about 5 MB, downloads once and runs in your browser, on the graphics card through WebGPU where the browser supports it and on the processor otherwise.
The Classic method uses resampling: it uses the browser's high-quality image smoothing to calculate each new pixel from the surrounding original pixels. It does not use a neural network and does not guess what the missing detail should look like.
With Classic, two details make the result cleaner than a single stretch in an ordinary image viewer. First, the enlargement happens in steps of at most 2× at a time, so 4× is done as 2× and then 2× again, and 3× as 2× and then 1.5×. Second, after the final step, a sharpening pass increases the contrast between each pixel and its four neighbours. That restores some of the crispness that smoothing takes away, especially along edges and lines. The Sharpen slider sets its strength, from 0 to 100 percent, with 30 percent as the default.
Classic needs no model, so it starts immediately. With either method the image never leaves your device.
Step by step
- Drop one or more images on the page. JPG, PNG, WebP, HEIC and most other formats are accepted.
- Optional: click Edit to crop the image first. Upscaling only the part you need keeps the result smaller and faster to process.
- Choose the method: AI for photos, Classic for instant results or for graphics where you want no invented detail.
- Choose a Scale factor: 1× (AI clean-up at the same size), 1.5×, 2×, 3× or 4×.
- With Classic, set Sharpen. Leave it at 30 percent for most images; see the table below for when to change it.
- Choose the output format. PNG keeps every pixel; JPG and WebP give smaller files. With JPG, pick a Background colour for any transparent areas.
- If the result must stay under a file size, turn on Also limit file size and enter the target.
- Press Upscale, then drag the compare slider on the result to judge the edges before you download. Files are named with the factor, for example photo-2x.png.
Worked example: a 640 × 480 image at 2×
Say you have an old 640 × 480 photo of a shop front, taken with an early digital camera, and you want a larger version for a slide. Here is what the Classic method does with it; the AI method is compared at the end.
At 2× the result is 1280 × 960 pixels. That is four times as many pixels, because both the width and the height double. Every new pixel is a blend of its neighbours, so diagonal edges such as the sign's lettering come out smooth instead of in visible stair steps, and the default 30 percent sharpening gives the outlines of the window frames a firm edge.
What you will not get is detail that was never captured. If the lettering on a small sign was a blur of a few pixels in the original, it is a larger, smoother blur at 1280 × 960. Fine texture such as brickwork, leaves or hair keeps the softness of the original, just bigger. On a laptop screen the 2× version looks clean; viewed at 100 percent zoom it looks slightly soft, as an enlargement always does.
If the original was a heavily compressed JPEG, the faint square blocks of compression are enlarged too, and sharpening makes them easier to see. For that kind of source, lower Sharpen to between 0 and 15 percent.
Scale factors and what to expect
| Factor | Output size | How it is done | What to expect |
|---|---|---|---|
| 1.5× | 960 × 720 | One step | Barely distinguishable from the original at normal viewing size |
| 2× | 1280 × 960 | One step | Clean edges, noticeably soft fine detail |
| 3× | 1920 × 1440 | 2× then 1.5× | Fine for screens and small prints; textures look smooth and painterly |
| 4× | 2560 × 1920 | 2× then 2× | Good for logos and graphics; photos look clearly soft up close |
With the AI method the same factors look noticeably crisper, especially at 3× and 4×. Its input is limited to about 1 megapixel; larger images are reduced to that first, so AI suits small and medium photos best. The Classic output is capped at 64 megapixels. If the factor you choose would go beyond that, the tool enlarges as far as the cap allows instead. A 4000 × 3000 photo at 4× would be 192 megapixels, so it is enlarged to about 9238 × 6928, roughly 2.3×.
Which images upscale well
| Image | Result | Sharpen |
|---|---|---|
| Logos, icons and flat graphics | Very good; smooth curves, solid colours stay solid | 30 to 50% |
| Screenshots and interface images | Good; text stays readable, edges stay firm | 30 to 50% |
| Photos with clear shapes | Good at 1.5× and 2×, softer beyond | 20 to 30% |
| Faces and portraits | Acceptable at 2×; skin and hair turn smooth | 10 to 20% |
| Heavily compressed JPEGs | Compression blocks are enlarged along with the image | 0 to 15% |
| Noisy low-light photos | Grain is enlarged; sharpening exaggerates it | 0 to 10% |
| Pixel art | Poor; smoothing blurs the hard pixel edges that define the style | Not a good fit |
Upscaling for print
Print shops often ask for 300 dpi, meaning 300 pixels for each inch of the print. A print 6 inches wide therefore needs at least 1800 pixels across. The 640 × 480 example falls short at 2×, but at 3× it reaches 1920 × 1440, which covers a 6 × 4.5 inch print at 300 dpi.
Be realistic about what the printer will show. The print will not reveal pixel blocks or jagged edges, but it will be as soft as the original information allows. If you also need the dpi value written into the file, run the result through Change DPI afterwards.
Getting the best result
- Start from the best copy you have. The original file from the camera or phone holds more detail than a version saved from a chat app or a social network, which has usually been shrunk and recompressed.
- Upscale once, from the original. Enlarging an already enlarged image, or saving as JPG between rounds, adds softness and compression artefacts each time.
- Crop before you enlarge, not after, so the pixel budget goes to the part you keep.
- Export as PNG if you plan to edit the result further, and convert to JPG or WebP only at the end.
Troubleshooting
- Edges have a bright or dark outline: Sharpen is too high for this image. Lower it in steps of 10 percent until the halo disappears.
- The result looks too soft: raise Sharpen, or choose a smaller factor. The larger the factor, the softer the result looks at 100 percent zoom.
- The file is very large: PNG stores every pixel, and a 4× result has 16 times the pixels of the original. Export as JPG or WebP, or turn on Also limit file size.
- A transparent PNG got a white background: JPG cannot store transparency, so transparent areas are filled with the Background colour. Export as PNG or WebP to keep them.
- The browser tab crashes on a big image: a 64 megapixel result needs hundreds of megabytes of memory while it is processed. On phones especially, crop first, use a smaller factor, or upscale one image at a time.
- You need exact dimensions, such as 1920 × 1080: use Resize image instead, which takes a target width and height directly.
AI or Classic: which to choose
The AI method generates detail that looks plausible: crisper edges, cleaner text, firmer textures. The results usually look much better than resampling, but some of that detail is invented. Faces can change subtly, very small text can turn into convincing nonsense, and fine textures may be replaced with something that was never there. Unlike most AI upscalers, it runs on your device, so nothing is uploaded and there are no credits; the trade-off is time: on the processor a 1 megapixel photo takes a minute or two, while WebGPU is far faster.
Classic makes the opposite trade. It only works with the pixels you gave it, so the result is predictable and faithful to the original, but it cannot recover detail. It is instant and suits logos, screenshots and documents where invented detail would be wrong. Both methods work offline in the installed app with Pro once the model has been downloaded. The free plan accepts up to 50 images per batch and files up to 200 MB; Pro raises that to 500 images and 1 GB per file.