AttentionAI
AttentionAI predicts visual attention areas from a single image, enabling design optimization before production.
Example output
Examples of attention maps generated from marketing visuals and web interfaces.
What this project illustrates
-
Turning research work into a directly usable tool
-
Designing models focused on usability and interpretability
-
Balancing scientific rigor with real-world product constraints
-
Bridging scientific foundations and UX decision-making
Context & intent
Context
Visual attention studies (eye-tracking) are costly and difficult to integrate into fast product cycles.
On landing pages or advertisements, key elements (CTAs, main messages) often go unnoticed.
Approach
AttentionAI relies on a deep learning model trained on human attention data to automatically generate predictive heatmaps, highlighting the areas most likely to attract visual attention.
Value
-
Evaluate a visual in just a few seconds
-
Identify the areas that are actually visible within an interface
-
Quickly compare multiple design variants
-
Help prioritize CTAs, messages, and key elements
-
Accelerate UX and marketing decision-making
Project management & key decisions
-
Choosing a generalizable model rather than one optimized for a single domain, in order to cover a wide range of use cases
-
Favoring a direct visual output (heatmap) over a single score, to remain actionable for designers and product managers
-
Intentionally limiting the scope to attention prediction to keep the signal clear and robust
-
Cost-controlled inference strategy: ~10 min first run, ~5 min subsequent runs; 15 predictions/day cap
Tech stack
-
Frontend : React, Netlify
-
Backend : FastAPI, Python
-
ML : Keras, Vertex AI
-
Infra : Docker, Artifact Registry, GCS
Scientific background
AttentionAI builds on work in cognitive science and computer vision showing that human visual attention is guided by both visual factors (contrast, saliency, structure) and cognitive factors.
Saliency prediction models make it possible to approximate these mechanisms and generate attention maps close to those obtained through eye-tracking.