FAQ

1. What does PhotoSalience do?

PhotoSalience analyses a photograph and predicts where attention is likely to gather as a heatmap, and how the eye is likely to move through the frame as a scanpath.

2. Is PhotoSalience free to use?

Yes. PhotoSalience is free to use, browser-based, and requires no signup or account.

3. Are my uploaded photos stored?

No. Uploaded images are processed only to generate your heatmap and scanpath, then removed from server memory after a short period. See the Privacy page for more detail.

4. How accurate are the predictions?

The results are indicative, not measurements of real viewer behaviour. They suggest likely attention patterns based on vision-science research, but every viewer is different. Think of them as a compositional guide, not a verdict.

5. What models power the analysis?

Heatmaps are generated with UNISAL, a vision-science model trained on human attention data.

Scanpaths combine a custom algorithm built by the PhotoSalience team with DeepGaze III, an autoregressive eye-fixation model. Together they predict the likely eye-movement sequence through the frame.

See the Research Credits section for the underlying papers and model references.

6. Can I use the results commercially?

Yes. You retain ownership of the images you upload, and of the heatmap and scanpath outputs generated from them. See the Terms page for more details.

7. Why does the heatmap sometimes seem to ignore my main subject?

The heatmap predicts where visual attention is most likely to be drawn, often by contrast, edges, brightness, isolation, or other salient features in the frame. That does not always match the subject you intended. When it does not, the result can still be useful: it may reveal a distraction, a competing point of interest, or a gap between intention and attention.

8. Why might the scanpath look different at different resolutions of the same photograph?

The models analyse the pixels you send. A higher-resolution source carries more visual information than a lower-resolution one — slightly different fine detail, slightly different contrast, slightly different edges. Both are valid views of the same scene, but the eye-fixation model is sensitive enough to those small differences that the predicted scanpath can shift, especially at later fixations where each step depends on the previous ones.

9. Why does the first fixation land on that position?

Eye-tracking research has consistently shown a “central fixation bias” — viewers’ first fixation tends to land at or near the centre of an image, regardless of where the actual subject sits and regardless of reading direction. PhotoSalience reflects this finding without applying it blindly: rather than dropping the first fixation at the geometric centre regardless of content, the algorithm looks for the most salient feature inside a centred region of the frame.

10. How many scanpath points should I choose?

It depends on the photograph. A scene with one clear subject often tells its story in six or seven scanpath points; a busier composition with several competing elements may need more before the pattern becomes meaningful. As you raise the count, the scanpath explores progressively less prominent parts of the frame — so once the important areas are covered, additional points tend to land on quieter, less revealing regions.

A good way to find the right number for your image is to slide the count upward and watch how the path evolves. When new points stop adding insight, you’ve found the sweet spot for that photograph. The slider currently caps at 12 points to keep processing times reasonable.

11. Are the scanpath dots all the eye sees?

No. The scanpath traces how attention is likely to move through the frame — not a literal list of pixels the eye registers. Each dot marks an estimated fixation point, but the eye continuously sweeps between them, and peripheral vision picks up everything along and around the line connecting consecutive dots. If a salient detail (a bird in the sky, a face at the edge of the frame) sits close to the path but not on a dot, it’s likely still noticed as the gaze passes near it.