Orchestrating Safety-Aware Visual Narratives via Multi-Agent Cognitive Calibration
We introduce BookAgent, a safety-aware multi-agent framework for end-to-end visual storybook generation.
Unlike prior pipelines that decouple text and image generation, BookAgent jointly plans, generates, verifies, and repairs multi-modal narratives.
💡 Unlike stage-wise pipelines, BookAgent introduces a closed-loop cognitive generation paradigm for long-horizon multi-modal storytelling.
BookAgent is built upon a closed-loop multi-agent architecture with three key stages:
-
Value-Aligned Storyboarding (VAS)
Ensures safety and transforms raw drafts into structured story plans. -
Iterative Cross-modal Refinement (ICR)
A generate–verify–revise loop that enforces text-image grounding and identity consistency. -
Temporal Cognitive Calibration (TCC)
Performs global reasoning and selective repair to maintain long-horizon consistency.
We build a fully functional interactive system for storybook creation, supporting:
- ✏️ Story input and page control
- 🎨 Style customization
- 🔁 Iterative global refinement
- 🧩 Character consistency via reference sheets
- Node.js
-
Install dependencies:
npm install
-
Set your API key in
.env.local:GEMINI_API_KEY=your_api_key_here
-
Run the app:
npm run dev
- 🔁 Closed-loop generation (not one-shot)
- 🎭 Character identity consistency across pages
- 🧠 Multi-agent collaboration
- 🛡️ Child-safe content generation
- 📚 Long-horizon narrative reasoning
BookAgent significantly improves:
- 📖 Narrative coherence
- 🧍 Character consistency
- 🛡️ Safety compliance
compared to prior methods such as StoryGPT-V and MovieAgent.
We thank Google AI Studio for providing an intuitive platform for rapid prototyping and deployment of our interactive demo system.
@article{gao2026bookagent,
title={BookAgent: Orchestrating Safety-Aware Visual Narratives via Multi-Agent Cognitive Calibration},
author={Gao, Bo and Liu, Chang and Miao, Yuyang and Ma, Siyuan and Lim, Ser-Nam},
journal={ACL Findings},
year={2026}
}