How AI interprets · 6 min read
How AI interprets the still-life tradition.
A close reading of what happens when a contemporary image model is given the visual grammar of Northern Renaissance still life, and where its defaults fail.
Image models trained on broad image corpora reproduce the visible signatures of still-life painting (frontal staging, single-source light, pewter and fruit) but consistently miss the genre's deeper conventions: the meaning of lemon peel as a symbol, the spatial logic of a vanitas, the function of damage in Heda. Directed work fixes this by specifying conventions, not just subjects.
Give a contemporary text-to-image model the prompt "Dutch still life" and you will get a reasonable approximation: a dark background, a pewter cup, some fruit, a single source of warm light. Look at it for thirty seconds and a few things are wrong, in ways that matter.
What the model gets right
The visible signatures. Models trained on broad corpora — which now include every public-domain Heda, Claesz, and Kalf — have learned the visual fingerprint of the genre very well. The dark ground, the raking light from canvas-left, the heavy textiles, the half-peeled lemon: all of it shows up reliably.
What it gets wrong
The conventions. A few examples:
- The peeled lemon. In a Heda still life, the spiral of lemon peel is a symbol — bitter on the outside, sweet within. A model will render the spiral as decoration; a directed studio will frame it as the painting's subject.
- The vanitas object set. A skull, a candle, a watch, a book. Models will scatter these freely; the genre has a strict spatial grammar (skull at left, candle at right, book between).
- Damage and oxidation. Heda's pewter looks the way it does because the artist showed every dent and every tarnish line. Models smooth pewter into uniformity unless explicitly told not to.
- The composition triangle. Almost every great still life uses a tight asymmetric triangle running diagonally across the frame. Models default to centered staging.
Why this matters
If you generate still lifes without specifying the genre's conventions, the output reads as competent but unmemorable — a stock image of the idea of a still life. If you specify the conventions, you get something closer to the actual tradition.
This is the work that 'directed AI art' is doing. The curator's brief for Decayed Renaissance, the studio's third collection, lists every Heda convention as a requirement and every model default as a thing to fight against. The decay — the chromatic separation and scanline drift — is added not as a stylistic move but as a way to mark the image as a contemporary artifact reaching backward, not a pastiche.
What buyers should look for
When evaluating any AI-generated still life:
- Does the composition use an asymmetric triangle, not centered staging?
- Is there a single, identifiable light source from one direction?
- Are the objects symbolic — does the lemon, the cup, the book have an argument?
- Is damage rendered (tarnish, scuff, oxidation), or is everything aggressively smooth?
- Does the palette stay inside the genre (umber, sepia, pewter, gold leaf) or wander into modern oversaturation?