A self-directed test of how far one trained style reference can carry. A single sref was fixed, then fed nothing but bare nouns, rocket, tower, car, roman edifice, to see whether the model would return them all in the same retro-futurist illustration language. The hard part was proving it was a system, not a fluke: that the look holds no matter how unrelated the subject.
- 01
Searched for and locked one style reference whose palette, linework and retro-futurist mood were strong enough to override the subject every time.
- 02
Reduced the prompts to almost nothing, *one noun plus the fixed sref*, so any consistency had to come from the trained style, not from prompt engineering.
- 03
Ran a wide spread of unrelated nouns through the exact same reference and kept the cleanest expression of each, the way you would build out an illustrated catalogue.
- 04
Held the reference constant across the whole set, so the collection reads as one illustrator's hand applied to a deliberately random brief.
A consistent illustrated world built from one reference and single-word prompts, kept as a style-system proof: evidence that a well-chosen sref can become a reusable house style on demand.
