THE SCIENCE
Built on published research. Honest about its limits.
Every Heatpoints prediction comes out of a public, peer-reviewed academic model — not a proprietary black box. Here is exactly what runs under the hood, how the saliency map becomes your scores, and where prediction stops.
THE MODEL
UniSAL — Unified Image and Video Saliency Modeling
Droste, Jiao & Noble — ECCV 2020, University of Oxford
UniSAL is a single neural network that predicts visual saliency — where human eyes go first — for both images and video. It was trained on public datasets of real human eye-tracking: SALICON, with over 10,000 images annotated with actual attention data, plus DHF1K, Hollywood-2 and UCF-Sports for video.
It has been evaluated on the standard academic benchmarks for the field — SALICON, MIT300 and DHF1K. The paper, the code and the benchmark results are all public. You don't have to take our word for anything: you can read the study and inspect the model yourself.
FROM SALIENCY TO YOUR SCORES
The model outputs a saliency map — a per-pixel prediction of attention. We reduce that map to four metrics with fixed, deterministic formulas. The same screenshot always produces the same scores, whether you analyze it on the web, in the Figma plugin or with the Chrome extension.
ATTENTION
01
Attention
The mean intensity of the saliency map. How much predicted attention your page attracts overall — a page full of flat, low-salience areas scores low.
FOCUS
02
Focus clarity
How concentrated or scattered the attention is, measured from the spatial dispersion (standard deviation) of the map. One sharp hotspot scores high; attention smeared everywhere scores low.
COVERAGE
03
Coverage
The percentage of the page surface above the attention threshold. Too little and most of your layout is invisible; too much and nothing stands out.
HIERARCHY
04
Hierarchy
The page is split into a 3×3 grid and the dominant zone is compared to the runner-up. A strong ratio means one clear entry point; a weak one means competing focal points.
THE HONEST PART
What prediction can and cannot see
A saliency model predicts the first seconds of visual attention from pixels alone. That's a powerful signal — and a bounded one. Anyone selling it as a crystal ball is overselling.
WHAT IT SEES
- Contrast — strong luminance and color differences that pull the eye
- Faces — one of the most reliable attractors of human attention
- Size and scale — large elements dominating smaller ones
- Isolation — an element surrounded by whitespace stands out
- Visual hierarchy — how elements compete for the first glance
- Patterns learned from real human eye-tracking data, not hand-written rules
WHAT IT CANNOT SEE
- Your brand — a logo people recognize behaves differently from one they don't
- Visitor motivation — someone hunting for a price ignores your hero
- The deeper meaning of your copy — the model sees text as shapes, not arguments
- Personalization — every visitor gets the same prediction
- What happens after the first glance — scrolling, reading, deciding
This gap is exactly why we built Pulse: real-visitor measurement on your live page, so the prediction gets confronted with actual behavior instead of standing unchallenged. See how Pulse closes the loop →
QUESTIONS WE ACTUALLY GET
Is this eye tracking?
No. Eye tracking measures where real people look, with hardware, on recruited participants. Heatpoints predicts where people will look, using a model trained and validated on large public eye-tracking datasets. You get the signal in seconds instead of weeks — and Pulse exists precisely to check the prediction against real behavior.
Why don't you claim an accuracy percentage?
Because saliency models aren't measured that way. Academic benchmarks report precise metrics like AUC and NSS on specific datasets — numbers that mean something to researchers but turn into marketing when compressed into a single "93% accurate" badge. We prefer to cite the paper and let you read the actual results.
Do my images train the model?
No. UniSAL was trained by its authors on public academic datasets. Your screenshots are analyzed and scored — they are never used as training data.
Judge it on your own page
The fastest way to evaluate a prediction model is to run it on a page you know by heart — and see if it surprises you.