UpscaleBench

About

UpscaleBench is a benchmark for AI upscaling models, judged blind by video professionals. It's run by Topaz Labs, an Adobe company.

This first edition benchmarks 4 video upscaling models from 4 companies, plus bicubic as a reference. We ran the evaluations in October 2026 and published the results in November 2026.

We recruited 30 video professionals to evaluate upscales of 20 diverse videos (AI-generated, live action, and archival). In each blind A/B comparison, they watched two upscaled versions of the same video and picked the one they preferred.

See Methodology for how we ran it, and the paper for the full method and analysis.

November 2026 edition

Video Upscaling Leaderboard

Matchups

Evaluators made 260 blind comparisons. For each one, they watched the original and two upscaled versions (A vs. B) and picked the one they preferred, or called it a tie. The comparisons fall into two types of matchup: model vs. model, and model vs. the bicubic reference. Each row below shows how many times evaluators preferred each side, and how many ties there were.

Model vs. model

The 4 upscaling models met in a round robin: 6 matchups, each judged 30 times (180 comparisons).

Left preferredTieRight preferred

Model vs. bicubic

Each model also met bicubic, judged 20 times per matchup (80 comparisons).

Model preferredTieBicubic preferred

Videos

These are the same videos evaluators compared, re-encoded for the web. Evaluators saw them blind, labeled only A and B. Select the label at the top left or right of the video to switch versions.

Video info
Preference
This video is not available yet.
0:00 - 0:10

All videos

AI generated videos

Each video started as still images made with Nano Banana Pro, GPT Image 2.5 or Nano Banana 2.1. Seedance 2.5 then turned the stills into a 10-second video with 3 shots.

Live action, old and archival videos

These come from the Internet Archive, Artlist (licensed stock video) and Wikimedia Commons.

Evaluators

We recruited 30 video professionals who upscale video in their paid work: 15 who create with AI and 15 in professional post, matching the 10 AI and 10 real-footage videos. Every comparison was blind, and none of the evaluators works for a company whose model we benchmarked.

AI creatives 15

AI filmmakers, creative technologists and creators whose professional work focuses on generating and upscaling AI video.

Professional post 15

Colorists, editors, VFX and compositing artists, and archival and restoration specialists working across film, streaming, TV and archives.

Methodology

How we ran this edition.

Models

We picked 4 companies that make video upscalers and benchmarked each company's top model. We plan to expand to more companies and more models.

To keep the comparison like for like, we only included precise upscaling models which aim to stay faithful to the original video. Therefore, we intentionally did not include creative upscaling models, which reimagine details with more creative liberty.

We also scored bicubic, the standard non-AI resize, as a reference. It shows whether each model actually improves on simply enlarging the original.

Starlight 2.6Topaz Labs
Magnific PrecisionMagnific
FLUX 3 PreciseBlack Forest Labs
vCube ProByteDance
BicubicFFmpeg

Price

What it costs to upscale 1 minute of 24 fps video, to 4K and to 1080p, through each company's own API at the lowest self-serve price it lists, checked on 5 November 2026. We left out app plans, enterprise deals and resellers.

Price is not reflected in the preference score.

Speed

How long it takes to upscale that minute, using the median of our October 2026 renders. We count render time only, not time waiting in a queue. Magnific doesn't report render time, so its figure is an estimate, marked ~.

Speed is not reflected in the preference score.

Test videos

We picked 20 videos to cover a wide range of what people upscale: AI-generated clips, live action, archival footage and a classic cartoon. They don't cover every use case, but they span the main ones.

10 are AI generated and 10 are live action, old or archival, the same split as the evaluator groups. Each has something hard to upscale well: faces, fur, small lettering, foliage, water, fast action or low light.

The originals are SD, 720p or 1080p, at 24 fps and up to 10 seconds. You can watch them all under Videos.

Upscaling

Every model got the exact same original file for each video.

  • Same starting point: we set every original to 24 fps without adding frames. The 3 interlaced videos, an old TV format, were deinterlaced once, before any model got them.
  • One run per model: 720p and 1080p originals went to 4K, SD originals to 1080p. Each model ran on all 20 videos on 19 October 2026, at its default settings except where listed below.
  • Kept as delivered: we didn't resize, crop or color correct any result. If a model changed the size or shape of the picture, evaluators saw that too.
How we ran each model
ModelCompanyWhere we ran itSettings
Starlight 2.6Topaz LabsTopaz for WebSharpness 5, other settings default
Magnific PrecisionMagnificMagnific web appMagnific (New), Precision, strength 60, sharpen 0, smart grain 0, FPS boost off
FLUX 3 PreciseBlack Forest LabsBFL APICreativity 0 (precise), upscale factor set per video to reach the target size
vCube ProByteDancefalPro, 4K or 1080p preset, target 24 fps
BicubicReferenceffmpegExact scale factor, keeping the original's shape

The copies on this site are all re-encoded the same way (H.264, up to 18 Mbps), the originals included, at each model's output size and shape.

Evaluators

Preference is subjective, so we wanted trained eyes: professionals who assess upscaled video in their own work. We split them into two groups to represent the main people who upscale video, from AI-generated content to camera and archival footage.

All 30 have paid video work in the past year: 15 AI creatives and 15 in professional post. They are independent: we paid each of them the same, whatever they picked, and no one who works for Topaz Labs, Adobe, Black Forest Labs, Magnific or ByteDance took part. See all 30 under Evaluators.

Blind evaluation

Every comparison was judged blind, so each pick reflects only what evaluators saw on screen, not what they knew about a model or its company.

In each comparison, evaluators watched the original and two upscaled versions, labeled A and B, and told us which one they preferred, or that it was a tie, and why. Nobody told them which model made which version, or what to look for.

How we ran it

  • Plan registered first: before the first session, on 25 October 2026, we registered our plan on OSF, a public research registry. That fixed the test and the scoring before we saw any answers.
  • Names hidden: we removed everything in the files that could reveal a model or its company, without changing the picture, and renamed the files A and B.
  • Moderated sessions: each evaluator joined one moderated video call of about an hour, between 26 and 30 October 2026.

Scoring

Each comparison is a win for one model and a loss for the other. A tie counts as half a win for each. We turn all 260 comparisons into one score per model with the Bradley-Terry model, a standard way to rank from head-to-head results. It uses every comparison, ties included, to put all models on one scale.

The score is the chance, out of 100, that evaluators prefer a model when it is compared with the other models on the board. The interval under it shows where the score would likely land if we ran the test again with different evaluators and videos.

A gap between two models counts only if it passes a statistical test. Because we test all 10 pairs at once, we made that test stricter (a Holm correction). The paper shows the result for every pair.

Limitations

What we tested

  • Blind comparisons by 30 working video professionals.
  • 20 short videos. 10 are AI generated, 6 are live action and 4 are old or archival.
  • SD, 720p and 1080p originals, upscaled to 1080p or 4K.
  • Each model as it was on 19 October 2026, at the settings listed under Upscaling. Each video was upscaled once.

What we didn't test

  • Other settings. We didn't tune any model.
  • Vertical, HDR or 4K to 8K video. Every video is horizontal, standard dynamic range and 1080p or smaller.
  • Home users and hobbyists. The panel is all professionals.

Keep in mind

  • 30 evaluators and 20 videos is a small sample, so the intervals under the scores are wide. Where two intervals overlap, don't read much into the gap.
  • We didn't control how people watched. Evaluators used the screens and players they work with.
  • All 10 AI videos come from one generator, Seedance 2.5. ByteDance makes Seedance and also makes vCube Pro.
  • FLUX 3 Precise returned 3 of the SD videos 12.5% too wide, because it ignored a flag telling players to show them at 4:3. Without those 3 videos, it moves from third to second (52.1, against 49.9 for Magnific Precision).
  • These scores only hold within this edition. Don't compare them with other editions or with image results.

Paper

The paper has the full method and analysis. It also lists every check we ran on the result and every change from the registered plan. Individual answers, recordings and transcripts stay private, as we promised the evaluators.

arXiv paper

The link goes live when the paper is posted.

Pre-registration on OSF

Registered on 25 October 2026, before the first session.

Released with the paper

Which evaluator compared which pair, a list of every test file with its SHA-256 hash, and the analysis code. Also the original and upscaled videos, where rights allow.