GFX-301b · Module 1

Reference Decomposition

3 min read

A single reference image is a starting point. Three to five references of the same style are a pattern library. The decomposition process: analyze each reference independently, then cross-reference the extractions to identify the invariant elements — the components that appear in every instance.

Invariant elements are the style. Variant elements are the content. If three references all use the same color palette, that is style. If they use different subject matter, that is content. The decomposition separates the two with surgical precision. The research agent produces an invariant report: "Across 4 references: palette is invariant (same 5 colors within 3% hue variance). Composition is semi-invariant (all use a centered focal point, but grid density varies). Typography is invariant (Exo 2, same weight distribution). Lighting is invariant (top-left directional, warm). Texture varies (2 use photographic, 2 use illustrated)." The semi-invariant and variant elements need explicit decisions — do you want to lock them down or allow variation?

Do This

  • Use 3-5 references minimum — single references conflate style and content
  • Cross-reference extractions to identify invariants — invariants are the transferable style
  • Make explicit decisions about semi-invariant elements — do not let the model decide for you

Avoid This

  • Extract style from one image — you cannot distinguish style from content with a sample size of one
  • Assume everything in the reference is the style — the subject matter, layout, and topic are content, not style
  • Let the generator "interpret" the style — interpretation introduces drift that compounds across outputs