Best free attention heatmap tools for your next review
For a free attention heatmap, start with Glance if you want a local design workflow. UNISAL and DeepGaze IIE provide research code; pysaliency helps evaluate predictions; the manual squint test checks composition without software. These options solve different jobs. Paid platforms add hosted workflows and support.
Choose by the job you need done
“Best” depends on whether you need a design review, research implementation, metric evaluation or team service. This is not a measured performance ranking. Free code does not automatically grant rights to model weights or datasets.
| Option | Useful for | Limit to check |
|---|---|---|
| Glance | Free, open-source local attention maps and design comparisons; no Glance account. | In development. MIT app code; model assets separately licensed. Verify downloads and supported models. |
| DeepGaze IIE code | Research implementation of learned saliency prediction. | Research only here: no reuse license established for this recommendation. Public code is not permission for commercial reuse or redistribution; request terms. |
| UNISAL | Researchers needing image and video saliency code. | Apache-2.0 code; archived repository. Check weights, dependencies and dataset terms. Requires a research/developer setup. |
| pysaliency | Developers evaluating models and saliency metrics. | A Python evaluation framework, not a polished screenshot-to-heatmap product. Check license and dataset terms. |
| Squint test by hand | Anyone checking large light/dark groups immediately. | No learned gaze model, measured gaze or accuracy score. |
What each free route gives you
Glance is the design-workflow route: keep files local, inspect a prediction and follow the model citation. A browser check makes the first image easy to try; desktop inference can run offline after setup. Check the download page before relying on a particular release. Browser runtime and model caching have separate availability constraints.
DeepGaze IIE and UNISAL are research routes. Expect to prepare inputs, manage dependencies and interpret model outputs. The DeepGaze IIE paper and UNISAL paper explain their methods. They do not establish that every conversion or export reproduces the published benchmark.
Use pysaliency when your question is how to evaluate a model. Use squinting when your question is whether the composition reads as a few clear shapes. Neither belongs in a feature checklist as though it were a hosted design review suite.
Paid options for team workflows
Choose a paid service when its collaboration, reporting or support removes work your team would otherwise do manually. Bring the same image and a clear reporting requirement to each trial or demo.
- Attention Insight comparison: PRO €119/month; Team €199; Business €399; Agency €799. Dated snapshot; check current plans and billing. Hosted collaboration, reports and API access. Prices checked 11 Nov 2026 — verify on their site: pricing / commercial enquiries.
- Neurons Predict comparison: No public price in 11 Nov review; demo / sales quote required. Customer success support and brand workflow. Prices checked 11 Nov 2026 — verify on their site: pricing / commercial enquiries.
- Brainsight comparison: Team €199/month; Business €399; Agency €799. Dated snapshot; confirm annual billing and tax. Shared workspaces, collaboration and email support. Prices checked 11 Nov 2026 — verify on their site: pricing / commercial enquiries.
- Feng-GUI comparison: No public pricing found in 11 Nov review; request current commercial terms. Hosted reports across media and API workflow. Prices checked 11 Nov 2026 — verify on their site: pricing / commercial enquiries.
12 Nov 2026 recheck: Feng-GUI has a public pricing page at the linked URL. The earlier “not found” observation is historical. These are dated price observations, not current quotes. Live plan names and monthly/annual billing can differ. No shared test establishes which product predicts your audience better.
Read the map before acting
Predicted attention is not observed gaze, comprehension or conversion. Keep model, crop and region boundaries fixed when comparing variants. Read the accuracy guide, then follow the landing-page workflow to choose one question for a real user test.