How AI interprets · 4 min read
How AI interprets color.
Why "muted ochre" comes out neon, and how restraint becomes a directable parameter.
Image models trained on broad image corpora over-saturate by default — most training images come from screens, where everything looks brighter. "Muted ochre" reliably comes out as neon mustard unless the brief enforces a saturation ceiling, names specific Pantones or hex values, and rejects out-of-palette candidates during curation. Color discipline is the most consistent gap in undirected AI art.
Color is the visible signal of curatorial discipline. Tight palettes read as considered. Loose palettes read as decoration.
Why models over-saturate
Training corpora skew toward screen-displayed images, which are typically saturation-boosted versus the physical world. The model's "neutral" sits two stops up from what a painter would call neutral. Asking for "muted" pulls back, but not far enough.
What works in a brief
- Saturation ceiling: "no saturation above 60%" — enforced visually during curation
- Specific hex values: "#6B1F1F oxblood, not red"
- Negative naming: "muted ochre, not yellow; warm grey, not silver"
- Reference photographs the model was likely trained on (Mapplethorpe's late flowers as a palette anchor)
- Hard rejection of candidates that exceed the saturation ceiling, even if compositionally strong
Where the studio fights the model
Almost every collection has a "no neon" rule. Almost every survival pass rejects candidates that returned with too much saturation, regardless of composition. The discipline is exhausting but visible in the finished work: a palette that does not move.