THE MODEL
UNISAL — how our attention model works.
No black box, no magic accuracy percentage. Here is exactly what runs when you scan a page: the model, the data it learned from, the formula we compute on top — and the things it cannot do.
1. What UNISAL is
UNISAL (UNified Image and video SALiency) is a saliency model published by Droste et al. at ECCV 2020 — peer-reviewed, open source, and still one of the strongest architectures for predicting where human eyes land on an image. 'Saliency' is the computer-vision term for visual attention: given pixels, predict the probability that a human fixates each region in the first seconds of viewing.
Its distinctive trait is unified training: one compact network (a MobileNetV2 backbone with domain-adaptive layers) trained jointly on image AND video gaze data. That joint training matters for web pages, which sit between the two worlds: static layout, but scanned by the eye in a sequence.
2. What it learned from
The model never saw your page before. What it carries is the statistics of thousands of real human gazes, recorded in controlled studies and published as public research datasets: SALICON (10,000 images with attention annotations), MIT1003/MIT300 (the canonical eye-tracking benchmark), CAT2000, and video datasets (DHF1K, Hollywood-2, UCF-Sports) recorded with real eye trackers.
From those recordings it learned the regularities of human attention: contrast pulls, faces capture, large text weighs, isolated elements win, centers bias. Your scan applies those regularities to your pixels.
3. What we compute on top
UNISAL outputs a gray probability map. Everything else is our layer, and it's deliberately simple and public: attention intensity (mean saliency — typical pages sit at 5-25%), focus clarity (does attention pool on one point or smear?), visual hierarchy (can the eye rank elements?), coverage (how much surface engages the eye), and the primary gaze zone on a 3×3 grid.
The score you see everywhere is: 0.35 × attention intensity (×4, capped at 100) + 0.35 × hierarchy + 0.30 × focus clarity. No hidden weights, no per-customer tuning. The AI-written parts of the report (audit, copy analysis) interpret these numbers — they never change them.
4. Why determinism matters
The same screenshot always produces the same map and the same score. That sounds obvious; it isn't — LLM-judged 'design scores' drift between runs. Determinism is what makes the score usable as an instrument: fix your page, re-scan, and the delta measures YOUR change. It's also why the score works in agent loops and CI (see our MCP server and API).
5. What it cannot do (honestly)
It predicts the first seconds of attention of a first-time viewer — not minutes of task-driven browsing, not brand familiarity, not your specific audience's habits. It reads pixels: text is shape and contrast to it (that's why our Words Map runs a separate language-level analysis). And a great attention score doesn't guarantee conversion — it guarantees the eye finds your message; the message still has to be good.
We deliberately don't market an accuracy percentage. Saliency benchmarks are public (MIT300 et al.) and UNISAL's numbers are in the paper; on YOUR site, the honest verification is our Pulse module — it measures real visitors cookieless and confronts prediction with reality, page by page.
PUBLISHED BENCHMARKS — FROM THE PAPER, NOT FROM US
These are UNISAL's exact numbers as published by Droste et al. (ECCV 2020) on the standard public saliency benchmarks. AUC-J and NSS are the two metrics the field ranks models by — higher is better.
| BENCHMARK | AUC-J | SIM | s-AUC | CC | NSS |
|---|---|---|---|---|---|
| SALICON (images) | 0.864 | 0.775 | 0.739 | 0.879 | 1.952 |
| MIT300 (images) | 0.872 | 0.674 | 0.743 | 0.784 | 2.322 |
| DHF1K (video, test) | 0.901 | 0.390 | 0.691 | 0.490 | 2.776 |
Two honest caveats. First: these benchmarks measure free-viewing gaze on natural images, not task-driven browsing of web pages — the prediction is a pre-publish signal, not a replacement for real analytics (that's exactly why Pulse exists). Second: newer models exist — so we benchmarked against them ourselves, below.
Model size: 15.5 MB · ~0.009 s/frame. Source: arxiv.org/abs/2003.05477
OUR OWN BENCHMARK — RUN BY US, PUBLISHED WHATEVER IT SAYS
We ran our production model head-to-head against DeepGaze IIE (2021, the reference high-accuracy model) on the Toronto dataset — 120 images with real human fixations, chosen because it sits in the training set of NONE of the compared models (MIT1003, SALICON, CAT2000 and OSIE all do, for one model or another). Same images, same metrics, same code.
| MODEL | NSS | AUC-J | CC | SIM | s/img |
|---|---|---|---|---|---|
| UNISAL (ours, production weights) | 2.310 | 0.882 | 0.803 | 0.672 | 0.06 |
| DeepGaze IIE (2021) | 2.341 | 0.883 | 0.836 | 0.706 | 0.82 |
| SUM (WACV 2025) | pending — requires a CUDA run; same protocol, published as soon as it's done | ||||
The honest reading: DeepGaze IIE wins every accuracy metric — by ~1% on AUC-J and NSS, by 3-5% on CC and SIM — while being 27× larger and 14× slower. That trade is exactly why we run UNISAL: near-parity accuracy at a speed that makes free instant scans, CI checks and agent loops economically possible. Toronto: Bruce & Tsotsos, 120 images, 20 observers. Run 2026-07-13, harness and raw results in the repo.
WHAT WE KEEP, AND FOR HOW LONG
Anonymous scans from the free score tool are automatically purged after 30 days (unless you published them yourself in a roast or a battle). Reports and their screenshots belong to your account: they stay until you delete them or close the account. Uploaded images outside a saved report are never stored after processing.
SOURCES & FURTHER READING
The best way to evaluate a model: run it.
Scan any page free — the map, the score and the formula above, in 30 seconds.
Score my page →