How AI interprets · 4 min read
How AI interprets botanical illustration.
A close look at what contemporary image models do well and badly with botanical illustration, the field Karl Blossfeldt defined.
Image models trained on broad corpora reliably reproduce the surface conventions of botanical specimen illustration (frontal staging, scientific isolation, plate-numbered composition) but consistently fail at the genre's structural conventions: accurate plant anatomy, scale relationship between specimen and frame, and the discipline of single-specimen-per-plate. Directed work fixes this by specifying conventions, not just subjects.
Karl Blossfeldt's 1928 specimen plates remain the dominant reference point for botanical illustration. Every contemporary "specimen plate" is in dialogue with Urformen der Kunst, including AI-generated ones.
What the model gets right
The conventions visible on the surface: a single subject, frontal staging, the specimen isolated against a neutral ground, the visual feel of plate-numbered scientific illustration. Models have seen thousands of Blossfeldts and have learned the look.
What it gets wrong
- Plant anatomy: models will produce flowers with the wrong number of petals, leaves with implausible vein structures, seed heads that mix two species
- Scale: a specimen plate has the subject occupying 30-60% of the frame. Models default to either subject-fills-frame or subject-too-small
- Single specimen discipline: models try to be helpful by adding companion subjects (a smaller leaf, a stem)
- Lighting: Blossfeldt used diffuse soft light from a single direction. Models often add multiple light sources for "appeal"
How Brutalist Florals fixes this
The collection's brief is restrictive in exactly the ways above. Each plate must show one subject (rejected if companion plant appears). Lighting must be single-source from upper-left. Subject scale must fall in the 30-60% band. Anatomy is checked against reference photography before survival.
What this means for buyers
When evaluating any AI-generated botanical work, look at scale, count the petals, check whether companion subjects intrude, look at the light. If those four are wrong, the work is decoration, not study.