Where the words come from
Our current source is the English Speller Database (SCOWL), a word-list project used for spelling dictionaries. We use the small US-English selection, with spelling variants limited. It is not a complete list of the English language.
Each new prompt draws four to six words from the filtered list, without repeating a word within that draw. Useful three-letter words stay; one- and two-letter entries are excluded. Earlier experiments used a noisier 10,000-word list and sometimes longer prompts.
We exclude words used in the other pools, plus color names, related presentation terms, common filler and an explicit-content ban list. Lowercase-only entries help reduce names and abbreviations. Filters can miss ambiguous words, so we still review the draw. The usable word count can change as the source and exclusions change.
Different choices, different jobs
Content words
Four to six words supply the starting ideas. They can be concrete, abstract, ordinary or unexpected.
24 art styles
One style shapes the appearance: for example, gouache, cyanotype, cut-paper collage or clay animation. The prompt gives style priority.
20 image types
A scene, diagram, repeating pattern, collection of objects and other formats suggest how to organize the image.
An image type for every run
Every new prompt draws one of 20 image types with equal probability. Scene draws also include lighting and composition; other types omit those instructions. The free-choice option is currently paused.
8 lighting options
Only when the image type is a scene, draw one lighting option, such as overcast daylight, moonlight or candlelight.
8 composition options
Scenes also draw a composition, such as overhead, wide shot or off-center. Other image types leave these instructions out.
Every prompt asks for no visible text and a portrait 4:5 format. These are generation instructions; an image may interpret them imperfectly.
The words are the title
A collection is named after its drawn words, in their original order. No invented theme title sits between the prompt and the images. Early collections retain their original longer word lists.
Within a comparison, we change one thing: regenerate the same prompt, shuffle the same words, switch the style, or switch the image type. The collection keeps its original title when the words are shuffled. If it is a scene comparison, lighting and composition stay fixed unless they are the thing being tested.
Some collections contain a single study; others hold several related images. The image viewer shows the input words, selected settings and a note about the comparison.
An exploration, not a guarantee
Random inputs do not make the output uniformly random. The model brings its own recurring interpretations, and rerunning the same prompt can change an image even when the words do not change. A different result after shuffling is interesting, but does not by itself prove that word order caused it.
We generate and review the images, choose what to keep, and add collections manually. This website is a catalog of those explorations, not an automatic image generator or a daily publishing schedule.
Browse the collections ↗