How the fly sees your ad

The fly on this site is not a real insect and not a cartoon. It is flyvis, a computer model of the fruit fly (Drosophila melanogaster) visual system published by Lappalainen, Tschopp, Turaga and colleagues in Nature in 2024. Its wiring is taken from the fly's connectome, the measured map of which neuron connects to which. Training on natural video then tuned it until its neurons predict how real fly neurons respond to motion.

1. The compound eye

A fruit fly has no single lens. Each eye is a honeycomb of roughly 700–800 facets called ommatidia, each sampling one small patch of the world. The model uses 721 of them. When your ad airs, we slide and pulse it across this hexagonal grid, then average each facet's patch to one brightness value. That coarse mosaic, shown in the Compound eye view, is the only picture the fly gets. Fine print disappears, while big shapes and strong contrast survive.

2. ON and OFF pathways

Light hits photoreceptors (R1–R6), which feed lamina neurons such as L1 and L2. These split the signal into "getting brighter" (ON) and "getting darker" (OFF) channels, much like your own retina does. The L1 and Mi1 views show neurons along the ON pathway.

3. Motion detectors

Flies are built to catch movement. They have to dodge swatters and land on fruit at speed. Cells called T4 (ON edges) and T5 (OFF edges) each come in four types (a, b, c, d), one for each cardinal direction of motion. In the Motion neurons view, brightness is how hard the strongest of them is firing and colour is the winning direction. The "motion detectors" meter is the average activity on screen.

What it is not

This is a model of one part of the fly's brain, the visual motion pathway, not a whole fly, and it has no opinions about your product. Real flies see ultraviolet, cover nearly 360° and detect flicker far faster than a monitor refreshes; the model leaves all of that out. Treat the views as a scientifically grounded approximation of a fly's early vision, not a literal recording.

How a slot is made

When an ad is bought, a computer in Oregon renders it once: 30 seconds of the ad drifting across the eye, simulated at 50 steps per second through about 45,000 model neurons. The responses are compressed into a small clip that every viewer's browser replays in sync. The same moment in the rotation is the same moment for everyone.

Further reading