How do you teach AI to paint like it paid attention?

Background removal for the CoNoggin Oat ground

Two techniques for making illustrations emerge from the page: BiRefNet subject extraction for levels 1–3 (floating subjects), edge-dissolve for levels 4–5 (organic brushstroke edges fade to transparent). Tested with diplomatic etiquette scenes on CoNoggin Oat (#FBF9F6).

Four AI models compete on the painting-time dial

Z-Image-Turbo ($0.005/image), Seedream 5.0 Lite ($0.035), Flux Dev + watercolour LoRA, and Flux Schnell — same prompts, same palette, different architectures. Side-by-side reveals that no single model wins on every axis. The answer is tiered: fast model for previews, quality model for production.

Teaching AI to understand how long an artist spent

We mapped the real watercolour process — from 30-second sketch to afternoon masterwork — onto a 5-level intensity dial. Each level adds the next layer of technique, colour density, and compositional complexity. The result: a controllable gradient from whisper to poster.

Same subject, same palette — six visual dialects

Six distinct watercolour approaches with the same teal palette and subject. From East Asian ink wash to storybook atmospheric, each tradition brings a different emotional language. The experiment revealed which styles Flux handles natively and which trigger artefacts.

How we learned to stop naming genres and start describing physics

Genre labels like "East Asian ink wash" or "impressionist watercolour" cause Flux to add stamps, seals, and signatures from training data. We discovered that "concept art" framing eliminates these artefacts entirely — anchoring the model in modern creative industry context where those conventions don't exist.