MegaPotato AI upscaler benchmark 2026: the data
================================================

The data behind "The Best AI Image Upscalers of 2026-2027: Benchmark With Reproducible Metrics and Prices"
by the MegaPotato Research Team (MegaPotato, an AI image upscaler, megapotato.app):
https://megapotato.app/blog/best-ai-image-upscalers-2026

Version 1.0, 14 October 2026. Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
Please link to the article when you use the data.

FILES
photos.csv  The 137 photos of the two tests: kind, source, licence and how the small input was made.
            image_url is the file itself where the host gives a stable link (Pexels, Unsplash, Wikimedia
            Commons), Flickr's original-size page, the video for the frames, and our own FLUX images for
            the AI set (free to use, CC BY 4.0 like the data); blank for the Library of Congress (the
            item page lists the files). Every other photo stays under its own licence at its source.
scores.csv  One row per photo and tool: the judges' scores and the image measures.
faces.csv   One row per face and tool: the face judge, the same-person score and face quality.

TESTS
core   16 photos, 18 tools and settings, each made 4x bigger. Every photo has its original.
       "Plain resize" and "Real photo" are references and are not ranked.
full   135 photos, three tools (MegaPotato, Topaz Wonder 3.5, Crystal), 4x. 43 photos have an original.
       The others were judged against the input enlarged by bicubic resizing.

COLUMNS
scores.csv  opus_judge          Claude Opus 5.5, blind, 0-10
            qwen_zoomed_areas   Qwen3.8 on zoomed areas, 0-10
            lpips, a_fine       closeness to the original, lower is better
            musiq               looks good, higher is better
            niqe                looks good, lower is better
            dists, dreamsim     closeness, lower is better
            dinov2, topiq_fr    closeness, higher is better
            colour_delta_e      CIEDE2000 colour difference, lower is better
            maniqa, clip_iqa_plus, topiq   looks good, higher is better
            Measures that need an original are empty where there is none.
faces.csv   face                the face's number within the photo
            face_width_px       its width in the input
            qwen_face_score     Qwen3.8 on the face, 0-10
            qwen_different_person   1 = the Qwen3.8 face judge said it is not the same person
            same_person_score   0-1, the mean of AdaFace, ArcFace, AuraFace and (where run) MagFace;
                                only where there is an original
            face_quality        eDifFIQA, without the original

THE TOTAL (0-100)
1. For each tool, average every column over the photos of a test. The face columns are averaged over the
   faces; "different person" is the share of 1s, in %.
2. The judges' scores are multiplied by 10. Every other part is scaled so that the best of the 18 tools of
   the core test gets 100 and the worst 0, then clipped to 0-100. The full test uses the core test's scale.
3. Total = 28 % Opus + 8 % each for the Qwen3.8 zoomed areas, the Qwen3.8 faces, the same-person score,
   different person, face quality, LPIPS, A-FINE, MUSIQ and NIQE. When a part has no values for a set of
   photos (no faces, no original), the other weights are scaled up to add to 100 %.
These steps give every number in the article's tables.