A chalkboard in an American diner carrying exactly nine lines of given text: a heading and eight dishes with prices. The score is how many characters came out wrong.
A flat map of the 48 contiguous states, each its own color, no labels. Every American has seen this map, and generated maps miscount states, merge them and invent coastlines.
One sentence, no details: a crawfish boil in a Louisiana backyard. Everything else is the model's own knowledge of the event, and a crawfish boil has a well-known look that a checklist can score.
A horse riding an astronaut, on the Moon. In 2022 this prompt was a standard counterexample: image models drew an astronaut riding a horse every time, because that is the picture the words usually describe. One sentence, one check.
An Outer Banks beach at sunset with no people, no birds, no boats and no footprints. Naming a thing in a prompt tends to put it in the picture, so four absences test negation four times. The setting adds a geography check: the Outer Banks' ocean beaches face east.
An open carton of twelve eggs, three of them brown, one white egg cracked. Three numbers to get right in one top-down photo.
One sentence: a man in an empty Phoenix parking lot at 12:30 p.m. in late June. The prompt never mentions the sun or a shadow. At that time the sun is about 80° up, so his shadow is a small pool at his feet. Models like long, dramatic shadows.
A wine glass filled with red wine to the very rim, about to spill. For years image models could not draw this: almost every wine glass photo shows a glass about one-third full, and models drew that whatever the prompt said.
The 32-panel black-and-white soccer ball: 12 pentagons and 20 hexagons in a fixed arrangement. Models draw the look of a soccer ball and get the tiling wrong.
A station clock showing 4:37. Watch and clock advertising almost always shows 10:10, and image models learned that time from it. One look at the hands scores it.
Overhead shot of a child's right hand playing a C major chord. Three tests in one frame: the black keys must come in groups of two and three, the hand must hold up, and the fingers must be on C, E and G.
A couple's hands with fingers interlaced on a café table, and a wedding band on one specific finger of one specific hand. Interlaced fingers are where generated hands merge, and the ring tests whether the model knows which finger is the ring finger and which hand is the left one.
Three kids on a lake dock, shot from water level so six bare feet fill the frame. Image models still miscount toes and mirror feet; this puts thirty toes and their reflections in close-up.
One reference picture of a character and one fixed scene: that same character, in that same style, sitting on their living room sofa stroking their cat, seen head to feet. The prompt never says what the reference is, so the model has to read the style itself — and then invent an entire body, outfit and room that the reference never showed. The failure that makes it worth recording is the model that draws it beautifully instead of faithfully.
Testing product design for different demographics
Test for generating school images.
Comparing school images
New York hedge fund CFO prompt - concise
Generate profile pictures for typical hedge fund manager CFOs
black-and-white, hand-drawn illustration (playful), suitable for YouTube videos and the like. thick black ink lines with a natural hand-drawn wobble and occasional unevenness.