Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

StoryState: Agent-Based State Control for Consistent and Editable Storybooks

Ayushman Sarkar, Zhenyu Yu, Wei Tang, Chu Chen, Kangning Cui, Mohd Yamani Idna Idris

TL;DR

System-level experiments on multi-page editing tasks show that StoryState enables localized page edits, improves cross-page consistency, and reduces unintended changes, interaction turns, and editing time compared to 1Prompt1Story, while approaching the one-shot consistency of Gemini Storybook.

Abstract

Large multimodal models have enabled one-click storybook generation, where users provide a short description and receive a multi-page illustrated story. However, the underlying story state, such as characters, world settings, and page-level objects, remains implicit, making edits coarse-grained and often breaking visual consistency. We present StoryState, an agent-based orchestration layer that introduces an explicit and editable story state on top of training-free text-to-image generation. StoryState represents each story as a structured object composed of a character sheet, global settings, and per-page scene constraints, and employs a small set of LLM agents to maintain this state and derive 1Prompt1Story-style prompts for generation and editing. Operating purely through prompts, StoryState is model-agnostic and compatible with diverse generation backends. System-level experiments on multi-page editing tasks show that StoryState enables localized page edits, improves cross-page consistency, and reduces unintended changes, interaction turns, and editing time compared to 1Prompt1Story, while approaching the one-shot consistency of Gemini Storybook. Code is available at https://github.com/YuZhenyuLindy/StoryState

StoryState: Agent-Based State Control for Consistent and Editable Storybooks

TL;DR

System-level experiments on multi-page editing tasks show that StoryState enables localized page edits, improves cross-page consistency, and reduces unintended changes, interaction turns, and editing time compared to 1Prompt1Story, while approaching the one-shot consistency of Gemini Storybook.

Abstract

Large multimodal models have enabled one-click storybook generation, where users provide a short description and receive a multi-page illustrated story. However, the underlying story state, such as characters, world settings, and page-level objects, remains implicit, making edits coarse-grained and often breaking visual consistency. We present StoryState, an agent-based orchestration layer that introduces an explicit and editable story state on top of training-free text-to-image generation. StoryState represents each story as a structured object composed of a character sheet, global settings, and per-page scene constraints, and employs a small set of LLM agents to maintain this state and derive 1Prompt1Story-style prompts for generation and editing. Operating purely through prompts, StoryState is model-agnostic and compatible with diverse generation backends. System-level experiments on multi-page editing tasks show that StoryState enables localized page edits, improves cross-page consistency, and reduces unintended changes, interaction turns, and editing time compared to 1Prompt1Story, while approaching the one-shot consistency of Gemini Storybook. Code is available at https://github.com/YuZhenyuLindy/StoryState
Paper Structure (17 sections, 1 equation, 5 figures, 4 tables)

This paper contains 17 sections, 1 equation, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. StoryState: Agent-Based State Control
  4. StoryState Construction
  5. Generation Pipeline
  6. Editing and Consistency Check
  7. Experiments
  8. Implementation Details
  9. Comparison with Baselines
  10. User Study
  11. Ablation Study
  12. Discussion
  13. Conclusion
  14. Storybook Dataset
  15. Prompt of Comparison Story
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Overview of the StoryState workflow. The system begins with a user prompt and constructs an explicit story state that comprises a character sheet, global world settings, and per-page scene descriptions. A team of LLM agents iteratively builds and updates this state, which drives both text and image generation through structured prompts. By maintaining this state persistently, StoryState enables localized editing and cross-page consistency without modifying the underlying models.
  • Figure 2: Qualitative comparison across story editing frameworks. Each row shows 10 pages generated by Gemini Storybook, 1Prompt1Story, or StoryState. Gemini requires full regeneration on edits; 1Prompt1Story improves consistency but limits diversity in actions and poses; StoryState enables localized edits while preserving identity and visual variation across pages.
  • Figure 3: User study comparing consistency and controllability across three story editing systems. StoryState is preferred for controllability (48%) and achieves the highest consistency preference (36%). See Table \ref{['tab:user']} for detailed results.
  • Figure A.4: Examples of StoryState generated results.
  • Figure A.5: Examples of StoryState generated results.