PhotoAIBench

Independent tests, public data

AI photo tools, measured

Measured tests of AI upscalers, background removers and photo restoration, on public-domain photos with published data and code, plus free image tools that run in your browser and never upload your photos. Written by a computer-vision engineer.

See the upscaler baselines Compare two images

By Updated

Kodachrome portrait of a weathered, smiling man in a striped knitted cap and an embroidered vest
Guide at Little Norway, 1942 Arthur Rothstein, FSA/OWI · Public domain

Free tools

  • Free tool

    Image resizer and compressor

    Shrink a batch of photos for the web or email, choose JPEG, WebP or AVIF, see the real file size, and remove camera and GPS metadata. Nothing is uploaded.

  • Free tool

    PSNR and SSIM calculator

    Compare an original with an upscaled, compressed or edited copy. PSNR, SSIM and a heat map of the differences, computed on your device.

From the baselines

The same lynx fur, enlarged four times, five ways

A 128 × 128 pixel region of the lynx photo: the original, the 32 × 32 pixels each method started from, and what bicubic, Lanczos and three Real-ESRGAN models made of them. Interpolation is soft but invents nothing. The generative model draws sharp, convincing fur that is not exactly where the real fur was.

That gap, between looking right and being right, is what the scores on this site measure.

See all 12 photos and scores

Crops of the same region of the lynx photo: the original, the 4x-reduced input, and the outputs of bicubic, Lanczos, Real-ESRNet x4plus, Real-ESRGAN x4plus and Real-ESRGAN general v3
Plate · Carpathian lynx, crops at 2x, input at 8x

The test set

12 photos, every one public domain or CC0

Portraits, landscapes, lettering and fine texture: the subjects where upscalers differ most. Each frame opens its crops and scores.

  1. A Carpathian lynx facing the camera, its ear tufts and thick spotted fur sharp against a dark, blurred background01 Carpathian lynx Wilfredor · CC0
  2. A small palm-covered island on a turquoise sea, with two white-sailed boats in front of it under scattered clouds02 Sunset Key, Florida Julian Lupyan · CC0
  3. Kodachrome portrait of a weathered, smiling man in a striped knitted cap and an embroidered vest03 Guide at Little Norway, 1942 Arthur Rothstein, FSA/OWI · Public domain
  4. An ornate hanging inn sign reading Hotel Post in blackletter, topped by a gilded carriage and horses, beside red geraniums04 Hotel Post, Murnau Jebulon · CC0
  5. Olive groves and green fields of the Messara plain stretching to a snow-capped mountain range under white clouds05 Mount Ida, Crete Jebulon · CC0
  6. Sepia studio portrait of Alexander Graham Bell, white-bearded, in a dark three-piece suit06 Alexander Graham Bell, c. 1895 Unknown photographer · CC0
  7. The gilded baroque interior of the Golden Chapel in Recife, every wall and arch covered in carved gold and painted panels07 Golden Chapel, Recife Wilfredor · CC0
  8. Rio de Janeiro at golden hour: mountains in haze, the city along the bay, and boats moored in the water below08 Rio de Janeiro at sunset Wilfredor · CC0
  9. Kodachrome portrait of a grinning young mechanic in a plaid shirt and cap, standing in front of a garage sign09 Garage mechanic, 1943 Marjory Collins, OWI · Public domain
  10. An orange Posted Nature Sanctuary sign nailed to a tree trunk in a bare spring forest10 Posted sign, Capen Hill Peter Cooper Jr. · CC0
  11. A rocky river running through a green valley toward a forested mountain under a blue sky11 Jacques-Cartier river, Quebec Wilfredor · CC0
  12. Black-and-white portrait of Walter White in a dark suit and tie, seated with his hands folded12 Walter White, c. 1950 Clara Sipprell · CC0

What is here

Two free tools that never upload your photos. The resizer and compressor shrinks a batch of photos for the web or email, shows the real file size for each quality setting, and removes the camera and GPS metadata. The PSNR and SSIM calculator compares an original with an edited, compressed or upscaled copy and shows where they differ. Both run entirely in your browser.

Measured upscaler baselines. 12 public-domain and CC0 photos, reduced 4x and enlarged again by five free methods, from plain bicubic to Real-ESRGAN, scored against the originals with PSNR, SSIM and LPIPS, with side-by-side crops. It is the yardstick commercial tools have to beat: upscaler baselines.

The method, in public. How photos are chosen and licensed, how a tool is scored, what the numbers can and can't tell you, and how commercial tools are tested within their own terms: methodology.

A buyer's guide grounded in the test. What actually differs between AI upscalers, and how to test one on your own photos before you pay: how to choose an AI upscaler.

How the site works

Free-to-use test photos. Every test photo is public domain or CC0, with its source and license on the methodology page, and no photo of an identifiable living person. Anyone can download the same files and check our numbers.

Scores from code, crops for your eyes. PSNR and SSIM measure how close an output is to the original, pixel for pixel. They don't measure how good it looks, and AI upscalers are exactly where the two part ways. So every score comes with crops of the same region from every method.

Commercial tools within their terms. We read each tool's terms of service before using it, use only free tiers that need no account, and never work around a limit or a captcha. A vendor sees its results seven days before they are published. A pilot round has been measured and will be published once the vendor has had that week.

Background removal and restoration are next. Their test methods are written down on the methodology page; nothing is published until it has been measured.

Paid links are labeled. The site may earn affiliate commissions from some photo-software companies, labeled "(paid link)" wherever they appear. None are active today. See the affiliate disclosure.

Also available as Markdown.