“Upscale without losing quality” promises more than any software can deliver. An AI upscaler cannot recover detail that was never captured. What it can do is make a plausible, sharper enlargement than simple resizing. We wrote this from our own upscaler's engineering notes so it can say what the tool actually does, what it costs in time and size, and where its limits are, including that it enlarges by 2× only.
What upscaling is and is not
Ordinary resizing (bicubic and similar) stretches the pixels you have and smooths between them, which is why an enlarged photo looks soft. An AI upscaler has learned from many photographs what edges, hair, foliage and skin tend to look like, and it draws a sharper version of the picture at the larger size. The extra detail is inferred, not recovered. It usually looks right on natural photos, and it can look wrong where the guess is wrong, for example in tiny text or fine patterns.
The model our upscaler uses was trained on real, compressed photographs (faces, hair, foliage, buildings) rather than on clean computer-shrunk images, which is why it copes with the compressed pictures people actually upload. It cannot rebuild a heavily blurred or tiny source. Starting from a 100-pixel thumbnail gives you a bigger, smoother 200-pixel thumbnail.
Size, time and limits
The numbers that decide whether your photo will work:
| Limit | Value | What it means for you |
|---|---|---|
| Scale | 2× only | Width and height double; pixels quadruple |
| Upload size | 3 MB | The page refuses bigger files before uploading |
| Input dimensions | 2048 px on the long side | Larger uploads are rejected, not quietly shrunk |
| Output size | 16 megapixels | 2× works for inputs up to about 4 megapixels, such as 2000 × 2000 |
| Time | About 19 seconds per input megapixel | Measured on our 2-core server; the queue adds time when busy |
The time figure explains something that surprises people: the model runs at the input resolution and enlarges in its final step, so asking for 2× or 4× costs the same, and what you pay for is the size of the file you send. A 0.5-megapixel photo takes about 10 seconds, a 1-megapixel photo about 19 and a 2-megapixel photo about 38, before any waiting in the queue. If the upscaled result is for print or the web, it is better to send a tight crop of the part you need than the whole frame.
How big can you print the result?
Print size is pixels divided by the resolution you print at, commonly 300 pixels per inch for sharp prints. This is arithmetic from the pixel counts, so it shows how far 2× takes you:
| Your photo | After 2× | Print at 300 PPI |
|---|---|---|
| 800 × 600 px | 1600 × 1200 px | 5.3 × 4.0 in |
| 1000 × 1000 px | 2000 × 2000 px | 6.7 × 6.7 in |
| 1500 × 1000 px | 3000 × 2000 px | 10 × 6.7 in |
| 2000 × 1500 px | 4000 × 3000 px | 13.3 × 10 in |
For large prints that need exact pixels per inch, the Enlarge Photo for Print tool sets the size and resolution, and our DPI guide explains why changing the DPI label alone adds no detail.
Why you should look, not trust a number
When we developed the upscaler we compared it with plain bicubic enlargement on the standard quality score, PSNR, which measures how close the pixels are to a reference. Plain enlargement scored higher (27.8 dB against 26.5 to 27.2 dB for the AI models) and looked obviously worse: its sharpness measure was about 4 against roughly 150. The score rewards smoothness, the opposite of what you want from an upscaler. So the only dependable test is your own eyes at the size you will use. The tool's before and after slider is there for that.
What file you get back
The output format follows the input. A picture with transparency comes back as a PNG, a PNG upload comes back as a PNG, and a photograph comes back as a high-quality JPEG (quality 95, higher than the usual 85 to 90 because the encoder should not throw away the detail the model just added). The file is bigger than the original, which is the trade. If it must stay under a size limit, use the Image Compressor afterwards, and see our compression guide for what each setting costs.
Small product images and text
A common case is a small supplier photo that a shop needs larger. Upscaling helps with soft edges and compression blockiness, and it will not turn a 300-pixel thumbnail into a marketplace-quality image. Check the marketplace's own minimum size first, and treat the result as an improvement, not a replacement for a proper photograph. Be especially careful with text, labels and barcodes: an upscaler redraws them from a guess, so a number or a barcode in the result may not match the original. Never rely on an upscaled barcode or nutrition panel without checking it against the real product.
Step by step
- If your photo is over 3 MB or 2048 px on the long side, reduce it first with the Image Compressor or Image Resizer.
- Crop to the part you need so the upload is as small as possible. Time grows with the megapixels you send.
- Open the Image Upscaler and add the photo.
- Wait for the result, then drag the before and after slider and zoom in on faces, text and edges.
- Download it. If the result looks wrong in one area, try a different crop or a cleaner original.
The tool runs on our server because the model is too large for a browser. Uploads are deleted within 1 hour and results within 6 hours, as the privacy policy describes.
What this guide does not claim
We describe how our own tool behaves and the limits we know about. We have not published a comparison with other upscalers, and the speed figure is from our server, so it can change. Results depend heavily on the source, so test a typical photo of yours before you rely on the tool for a batch.
