Compare your variants. The attention model ranks them.
Heatpoints Championship compares two or more variants of an ad, a thumbnail or a landing page and ranks them by predicted attention. Every variant goes through the same UNISAL saliency model and the same public score formula as the single tool, so the order comes from the pixels, not from an opinion. Free, no account.
Heatpoints Lab. Updated 2026-09-03.
Drop two or more variants
PNG, JPG or WebP, up to 10 MB each. One credit per variant, free, no account.
How does the championship rank variants?
Each variant is scored by the server with the formula the public scanner uses: 0.35 × attention intensity (×4, capped at 100) + 0.35 × hierarchy + 0.30 × focus clarity. Variants are sorted by that score. Equal scores keep the upload order and are flagged as a tie.
One score, one formula
0.35 × attention intensity (×4, capped) + 0.35 × hierarchy + 0.30 × focus clarity. The public scanner's formula, applied to every variant by the server, so the tool, the extension and the API agree.
Same pixels, same rank
The map is deterministic. Drop the same file twice and it lands in the same place with the same numbers. Identical files in one run are analyzed once and cost one credit.
Measured, never generated
The ranking and the gaps come from the maps. No language model writes the verdict, so nothing on this page is made up. Export the leaderboard as PNG or the numbers as CSV.
What you get for each variant
- Score, 0 to 100
- The composite the whole site uses: 0.35 × attention intensity (×4, capped at 100) + 0.35 × hierarchy + 0.30 × focus clarity. It ranks the run.
- Attention intensity
- The mean intensity of the saliency map: how much predicted attention the image attracts overall. A flat, low-salience image scores low.
- Focus clarity
- How concentrated or scattered the attention is, from the spatial dispersion of the map. One sharp hotspot scores high; attention smeared everywhere scores low.
- Coverage
- The share of the image above 30 percent predicted attention. High coverage dilutes; very low coverage reads as empty.
- Hierarchy
- The ratio between the hottest of nine zones and the second hottest. A clear dominant zone scores high; nine similar zones score near zero.
- Hot zone
- Which cell of a 3 by 3 grid holds the most predicted attention, from top-left to bottom-right.
- Gaps against the winner
- For every variant below first place, the difference on the score and on each metric, so you see what the winner does better, in numbers.
- Exports
- The leaderboard as one PNG with every heatmap and score, or a CSV with rank, score, the four metrics, hot zone and the gaps.
The four metrics and the formula are documented on the science page.
Predictive pre-test or live A/B test?
A championship predicts where the eye goes in the first seconds, from the files alone. An A/B test measures what real people click or buy, with real traffic. They answer different questions, so the useful order is: pre-test to shortlist, A/B test to confirm.
| Question | Championship (predictive pre-test) | Live A/B test |
|---|---|---|
| What it needs | The image files | Traffic, a budget and days |
| Time to a result | About a second per variant | Days to weeks, until the result is significant |
| What it measures | Predicted first-glance attention on the creative | Measured clicks or conversions |
| Cost of the losing variant | One credit | The spend or the conversions it lost |
| Best for | Shortlisting candidates before launch | Confirming the finalists in the real feed |
Who compares variants this way
Media buyers
Rank the Meta or display creatives of a campaign before the first euro leaves, then put the heatmaps in the deck.
Ad score, with a feed mockupYouTube creators
Rank the thumbnail candidates of a video by where the eye lands, before the algorithm decides.
Thumbnail score, at sidebar sizeDesigners
Export two or three frames and let the model settle the layout debate with numbers, in the same run.
The Figma pluginGrowth teams
Compare landing page screenshots against each other, and against the benchmark of 93 well-known SaaS pages.
The studyWhat the numbers rest on
0.872
AUC-Judd of UNISAL on the MIT300 benchmark, with an NSS of 2.322, as reported in the paper (Droste et al., ECCV 2020). The model behind every map on this site, trained on human fixation data. Source: the science page
44.1
Average score of 93 well-known SaaS landing pages scanned with this exact formula in 2026; the median was 41 and no page reached 80. 81 percent had their hottest zone outside the hero. Source: the study
1×
One inference per distinct file. The map is deterministic and cached by content: the same file gives the same map, the same score and the same rank, on the web, in the extension and through the API.
Frequently asked questions
How many variants can I compare in one run?
Two without an account, three with a free account, and up to ten on the Pro plan for ad-makers, which is coming soon. Each variant costs one credit, the same credit as a single analysis, and identical files in one run are analyzed once.
Which file formats and sizes work?
PNG, JPG and WebP, up to 10 MB per image. Screenshots, exported ad creatives and thumbnails all work: drop the exact file you plan to publish.
Is the ranking the same every time?
Yes. The model is deterministic: the same pixels produce the same map and the same score, so the same set of files always ranks in the same order. Ties keep the upload order and are flagged as ties.
What does the score mean?
A composite from 0 to 100: 0.35 × attention intensity (×4, capped) + 0.35 × hierarchy + 0.30 × focus clarity, the formula of the public scanner. It describes how strongly and how clearly the image pulls the eye, not how it will convert. For scale, 93 well-known SaaS pages averaged 44.1 and none reached 80.
Does a championship replace an A/B test?
No. It predicts where the eye goes in the first seconds; an A/B test measures what people click or buy. Use the championship to shortlist candidates before launch, then confirm the finalists with real traffic.
Can I compare YouTube thumbnails or Meta ads?
Yes. Thumbnails, ad creatives, banners, landing page screenshots and Figma exports are all static images to the model. For a single thumbnail or ad shown inside a feed mockup, use the thumbnail score or the ad score pages.
Where do my images go?
They are sent to the server for inference, analyzed, then deleted. No account, no public gallery, no training on your visuals.
Can I export the ranking?
Yes. Download the leaderboard as a PNG with every heatmap and score, or the numbers as a CSV: rank, score, the four metrics, hot zone and the gaps against the winner.
One image at a time is enough? The free tool is on the homepage. Live pages? The Chrome extension overlays attention on any page you browse.