Maintaining visual and character consistency in AI storybooks
Executive Summary
Multi-chapter illustrated story generation often suffers from character visual drift and plot incoherence across pages. We developed a hybrid cross-attention memory model that extracts invariant character feature vectors from initial input sketches. Evaluated on 10,000+ illustrated storybooks in Pictales, our method improves multi-chapter character retention by 68% over standard prompt-chaining baselines.
1. The Character Drift Challenge in Story Generation
When creating multi-chapter children's books or graphic novels, young readers depend heavily on immediate character recognition. Standard generative image diffusion models treat every page as an isolated generation, causing protagonist clothing, facial features, and companions to change unpredictably from chapter to chapter.
2. Invariant Visual Feature Memory
Our approach separates the protagonist's intrinsic identity from background environments and episodic actions:
Persistent Identity Token Vectors
Encodes distinct hair, costume, and silhouette features into a compact cross-attention memory key that is injected across all subsequent story chapters.
Adaptive Pacing & Narrative Logic
Ensures story arc progression (introduction, tension, resolution) with consistent age-graded vocabulary tailored for 4–12-year-old readers.
Cite This Work & Download Research Note
Download the complete research note PDF or test the engine live on Pictales:
@article{kites2025pictales,
title={Maintaining Visual and Character Consistency in AI Storybooks},
author={Kites Narrative AI Group},
journal={Kites Technical Reports},
volume={TR-2025-09},
year={2025},
url={https://kites.dev/research/pictales-character-consistency}
}