How to choose an AI upscaler (and test one on your own photos first)

An upscaler can't recover detail that isn't in the small image. It guesses. Every choice below comes down to how it guesses, how much of that you want, and what you give up to get it. This guide is based on our measured baselines and on reading the terms and free tiers of the web upscalers we considered for testing; it does not rank products.

1. Decide what you need the bigger image for

• Print or a large display of a photo you care about (a family photo, a portrait, a landscape for the wall): you want believable texture in faces, hair and foliage, and no invented features. Look at faces at 100% before you print.
• A product photo, a document, a sign or a screenshot: you want faithful edges and correct text. Invented detail is a defect here, not a feature.
• Web use at a modest size: a 2x enlargement shown at screen size hides most differences. Plain resizing may be enough; try it first.
• An old, damaged photo: you need restoration (scratches, cracks, fading), not just upscaling. Most "enhancers" do both at once and redraw faces while they are at it; decide whether you are comfortable with that.

2. Understand the one trade-off that matters

AI upscalers sit between two poles:

• Faithful models output the average of all plausible answers. Edges are cleaner than plain resizing, but fine texture stays soft. They score well on pixel-accuracy metrics (PSNR, SSIM).
• Generative models (GAN or diffusion based) draw sharp, plausible texture: pores, fur, leaves, brick. It looks like a real high-resolution photo, but the detail is invented, and some of it is wrong.

Our baselines show the split clearly. On 12 public-domain photos enlarged 4x, the generative Real-ESRGAN x4plus beat plain bicubic resizing on the perceptual metric (LPIPS) on 12 of 12 photos, and lost to it on pixel accuracy (PSNR) on 12 of 12. The same network trained without the generative loss (Real-ESRNet) was the most faithful on most photos and looked softer. On a sign's small print, the generative model wrote crisp letters, and got one of them wrong (see the crop).

Many paid tools offer both behaviours as modes ("faithful", "creative", "standard", "high fidelity") or a slider. That control is worth more than a model name.

3. Test it on your own photos before paying (ten minutes)

Marketing examples are chosen to flatter. You can run the same test we do, on the kind of photos you care about:

1. Take two or three sharp photos of your typical subject at full resolution. Reduce each to a quarter of its width and height in any editor, or with our image resizer (choose PNG so no compression is added). Keep the originals.
2. Upload the small copies to the upscaler you are considering, at 4x. Download the results.
3. Load each original and its upscaled result into the PSNR and SSIM calculator. The numbers tell you how faithful the result is; the heat map shows where it went wrong.
4. Then look, at 100%, at faces, hands, text and repeating patterns. Read the text. Check that eyes, teeth and jewellery are the shape they were.
5. Run the same small copies through a free baseline (plain resizing, or a free open-source upscaler, below). If you can't see the difference, you don't need to pay.

The scores won't tell you which result you prefer. They will tell you which tool changed your photo more, and the heat map shows where to look.

4. Read the free tier and the terms, not just the price

What we found reading the free tiers and terms of service of web upscalers (the full list goes with our pilot test, once the vendors have reviewed it):

• Many free tiers need an account, and several cap you at one to three images a day, at 2x only, or put a watermark on the result. Check that the free result is full resolution before judging quality.
• Check the output format. At least one web tool in our pilot returned JPEG files, adding compression, even for PNG uploads.
• Read the content license. Some services' terms give the company a broad, perpetual license to what you upload; others limit it to providing the service, and some delete uploads within a day. For private photos, that clause matters more than the price.
• Some terms forbid benchmarking or automated use. That is the vendor's right, and we respect it by not testing those tools. It also means you will not find independent measurements of them, only the vendor's own examples.

5. Consider running it on your own computer

• Free and open source: the Real-ESRGAN models in our baselines are free (BSD-3-Clause) and run on any computer; a graphics card makes them fast. Upscayl is a free desktop app (AGPL-3.0) that runs models from the same family without the command line. Your photos never leave your machine.
• Paid desktop software such as Topaz Gigapixel also runs locally and adds more models and controls. Topaz's documentation says its free trial offers every feature except exporting the result, so you can't take trial outputs into our calculator; judge on screen at 100%. We have not measured it.
• Web tools are the easiest for a handful of images and need no hardware. Mind the upload terms above.

A checklist

• [ ] I know whether I need faithful detail (documents, products, text) or believable texture (portraits, landscapes).
• [ ] I tested two or three of my own photos with the reduce-and-enlarge method and looked at faces and text at 100%.
• [ ] I compared against a free baseline (plain resizing or an open-source model).
• [ ] The free result I judged was full resolution, with no watermark.
• [ ] I read what the service's terms say about my uploaded photos.

How we will test commercial upscalers

The method, the photos and the scoring code are already public on the methodology page. Commercial tools are tested only through free tiers that need no account and only where their terms allow it; a vendor sees its results before they are published. Results will appear here, with crops, when that review is done.

Also available as Markdown.