The original title is "Character Consistency Across 24 Pages of a Generated Picture Book"

Short summary

A practical deep-dive on keeping character identity and style consistent across 24 AI-generated picture-book illustrations. The author separates identity consistency (harder, driven by reference embeddings) from style consistency (easier, fixed via seeds and prompt fragments) and recommends regenerating entire books rather than single pages. Concrete metrics like pairwise face-embedding distance and palette distance are proposed to catch drift that per-image quality scores miss.

• •Identity and style consistency are separate problems requiring different fixes — lock style via seeds and prompts, weight identity conditioning heavily
• •Regenerate the whole book as a unit; single-page fixes break neighbour consistency
• •Use pairwise face-embedding distance and palette distance across all page pairs, not per-image metrics, to detect drift

Generated with AI, which can make mistakes.