We stopped trusting
our own eyes.
Every studio argument about a layout ends in the same place: someone squints at the screen and declares the hierarchy broken. The squint test is the oldest arbiter in graphic design, and it is completely unfalsifiable — because the person squinting already knows what the layout is supposed to say.
So we went looking for an eye with no opinion. Drosophila melanogaster sees through roughly eight hundred facets and has almost no colour vision. It resolves masses, edges and motion, and it physically cannot read a headline. It is a squint test that cannot cheat.
We ran 126 of our own concepts through flyvis — a model of the fly's optic lobes built under the constraints of the connectome and trained on natural video. Every frame passes through 45,669 simulated neurons across 65 cell types. This is not a filter applied to an image. It is a nervous system responding to one.
Drag to lose
your resolution.
On the left, the frame exactly as the model received it. On the right, the same frame after the eye: 721 ommatidia, no colour, no type. Whatever survives the crossing is what carries a layout across a room — or down a feed.
Four layers,
lighting up.
The optic lobe is built as a retinotopic stack: one column per facet, four neuropils in depth. Every column below carries the activity our simulation actually computed for the concept on screen — contrast first, motion last.
Schematic geometry on the real hexagonal lattice — a diagram of measured activity, not a microscope image. Cell types are assigned to their known neuropil.
Four steps, and
the trap inside each.
The CMS poster is frame one, often an empty preloader. We locate the hero by edge density multiplied by motion stillness, discarding the first 15%. The median lands at 72% of the runtime.
The frame is fitted to the visual field and resampled into 721 ommatidia. Padding replicates edge pixels: a black letterbox would hand the fly a contrast edge stronger than anything in the design.
The network integrates over time steps. The first twelve frames are an onset transient — response spikes to 3.8 against a plateau of 1.4 — and are discarded.
We record L2 for contrast and T4/T5 for motion across the inner facets, dropping three outer rings where detectors have no neighbours. For films, across the whole timeline.
Then we tried to break
our own findings.
The flyvis ensemble holds fifty networks trained from different initialisations — different individuals of one species. A metric only means something if separate individuals agree on it; otherwise it describes one network's quirks, not your layout.
We had seven plausible metrics. We put every one of them through the same test: does a second fly rank the concepts the same way? Spearman correlation between independently trained networks.
▏ vertical rule marks ρ = 0.70, our threshold for reporting a metric
Driven by the stimulus
Cuts, motion, temporal structure. The response is dictated by the footage itself and every network sees it the same way. ρ between 0.82 and 0.98.
Contrast — borderline
ρ = 0.737 on average, 0.567 for the worst pair. Reported, flagged, not leaned on. Removing our own false letterbox edge lowered this from 0.805 — part of the earlier agreement was our padding.
Requiring interpretation
Attention concentration, response force, gaze direction. Here the network must infer spatial structure inside one frame and each does it differently. Directional selectivity scored ρ = 0.002 — pure noise.
Where a film stops
feeding the eye.
Of everything that survived the audit, one measurement changes how we cut. Tracking the response across a whole film shows the exact stretch where a concept opens strong and then offers the eye nothing. Shaded band = middle third.
Watch it react,
frame by frame.
Pick one of our concepts on the left. The film plays on the right, and underneath it the measured response runs in step with it — the playhead sits on the curve the simulation produced for these exact frames.
Live analysis of an uploaded layout or URL needs the model running on a server — PyTorch does not fit in a serverless function. The compute service is not wired up yet, so this panel is inactive rather than faked.
What this study
does not claim.
Nothing about colour
The model is monochrome. No palette conclusion can be drawn from it.
Nothing about meaning
This is early vision. Whether an idea is apt, on-brand or well written lies entirely outside the model.
High response ≠ good
Response level over a film mostly measures cut density. Ranking a portfolio by it would be a mistake.
Five networks of fifty
Static metrics used 5 ensemble members, curves used 2. A full run would tighten every number here.
One frame stands for a film
Static metrics are read from a single heuristically chosen hero frame, not the whole edit.
No human control group
The comparison against human votes is not done: too few votes collected so far.
We make websites.
This is how we argue
about them.
The finding is a boundary, not a ranking. Metrics anchored in the stimulus survive; metrics that ask a network to interpret space inside a frame do not. That line is worth more than any leaderboard, because it tells you which machine measurements of design you can trust at all.
And one methodological result we would rather have not needed. A pretty ordering produced by a neural network is worth nothing until a second network of the same kind produces the same ordering. We had a sensible-looking ranking of attention concentration. We threw it away: two flies agreed on it at ρ = −0.23.
None of this replaces judgement. It just moves one argument — does this hero hold at a glance — from taste to measurement. That is the part of our craft we can actually check, so we checked it on everything we have made.