Gallery · rendered and measured
The same drawing, by every model
Five visual tasks, every model's actual output side by side. Nothing here is judged on vibes alone: hands are measured by angle, wedges by colour-sampling, bands and bars by pixels — the render is the ground truth. The one subjective call (is the koala recognisable?) goes to a vision judge.
An analog clock showing 3:45
Hands measured by rendered angle — at 3:45 the hour hand must sit ¾ between 3 and 4, not on the 3.loading renders…
A pie chart of 450 / 300 / 150 / 100
Wedge angles colour-sampled off the render — 162°, 108°, 54°, 36°, or it fails.loading renders…
The flag of France
The prompt never names the colours. Band count, hoist-to-fly order and equal thirds, from pixels.loading renders…
A koala in a gum tree
Structure checked in code; recognisability judged by a vision model reading the render.loading renders…
A served bar-chart page (Jan–Jun sales)
Deployed live, bars measured in a real browser against the data. Screenshot taken before teardown.loading renders…
A rotating, lit 3D cube (Three.js)
Real WebGL, deployed live: frames must change over time and the faces must shade differently — a flat fake can't. One frame captured before teardown.loading renders…
How to read it. Green = every measured check passed. Red names the actual miss — hour on 3, wrong sizes, bars wrong — the kind of error that looks fine at a glance and lies. SVG cells are the model's file rendered directly; the bar-chart cells are screenshots of the model's live deployed page. Click any drawing to enlarge it.