pith. sign in

arxiv: 2604.16541 · v1 · submitted 2026-04-17 · 💻 cs.CV

BOOKAGENT: Orchestrating Safety-Aware Visual Narratives via Multi-Agent Cognitive Calibration

Bo Gao , Chang Liu , Yuyang Miao , Siyuan Ma , Ser-Nam Lim This is my paper

Pith reviewed 2026-05-10 09:40 UTC · model grok-4.3

classification 💻 cs.CV
keywords multi-agent frameworkvisual narrative generationstorybook synthesissafety alignmentnarrative coherencemulti-modal calibrationconsistency verificationillustrated stories
0
0 comments X

The pith

BookAgent generates illustrated storybooks from user drafts by having agents jointly plan, script, illustrate, and repair inconsistencies while enforcing safety.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that storybook creation can be handled as one connected process instead of separate stages of writing and drawing. It does this through a system where multiple agents coordinate to keep text and images aligned on each page and to catch problems in how characters and plot hold together across the whole story. Safety rules specific to children are built into the planning and checking steps so the output stays appropriate. A reader would care because many current tools still produce stories that jump around or include unsafe elements, and an integrated approach could make reliable visual storytelling more practical.

Core claim

BookAgent is a safety-aware multi-agent collaboration framework for end-to-end storybook synthesis from a user draft. It jointly performs planning, scripting, illustrating, and global repair of inconsistencies. Dynamic page-level alignment calibration matches textual scripts to visual layouts, while temporal verification rectifies inconsistencies in character identity and storytelling logic. Experiments show it outperforms current methods in narrative coherence, visual consistency, and safety compliance.

What carries the argument

The multi-agent collaboration with dynamic page-level alignment calibration and temporal global inconsistency verification that keeps text-image matches and story logic intact throughout the narrative.

If this is right

  • Storybook generation no longer needs to start from a fixed storyline sequence supplied in advance.
  • Child-specific safety constraints can be enforced inside narrative planning and sequence-level verification rather than added afterward.
  • Inconsistencies in character identity and storytelling logic can be detected and corrected at the global level after initial generation.
  • Multi-modal creation tasks become more reliable when agents handle planning, illustration, and repair in one loop.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same style of per-page and global calibration could be adapted to keep consistency in longer video sequences or interactive story systems.
  • Embedding safety checks throughout the pipeline rather than at the end might reduce the need for heavy post-filtering in other child-oriented content generators.
  • Public release of the code makes it possible for others to measure whether the calibration steps hold up when users provide very different draft styles or lengths.

Load-bearing premise

The assumption that dynamic page-level alignment calibration and global inconsistency verification can be implemented in a way that consistently produces the claimed gains in coherence and safety.

What would settle it

A replication experiment using the same evaluation metrics that finds no significant improvement over prior staged methods in narrative coherence, visual consistency, or safety compliance.

Figures

Figures reproduced from arXiv: 2604.16541 by Bo Gao, Chang Liu, Ser-Nam Lim, Siyuan Ma, Yuyang Miao.

Figure 1
Figure 1. Figure 1: Teaser: Long-horizon story consistency re￾quires collaboration. Given the same multi-step story prompt with strict ordering and counting constraints, a single-pass baseline generation fails to preserve charac￾ter identity and temporal consistency across pages (top). In contrast, BOOKAGENT leverages multi-agent col￾laboration to maintain stable characters, correct event order, and consistent visual attribut… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of BOOKAGENT. The framework follows a closed-loop, multi-agent architecture with three mechanisms. Stage 1: Value-Aligned Storyboarding (VAS) audits the input story against safety guardrails and structures it into a page plan with extracted characters and a reusable character sheet. Stage 2: Iterative Cross-modal Refinement (ICR) iteratively refines page prompts and generates candidate images, gui… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on character and object consistency (Milo). drift and object hallucination. In contrast, our method successfully anchors character identity and props throughout the sequence. This advantage is further pronounced in the Rowan case ( [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on hard attribute constraints (Rowan). proximation of human qualitative judgment while reducing individual evaluator bias. We conduct a small-scale user study to evalu￾ate overall preference for generated visual stories. For each prompt, participants viewed anonymized visual stories generated by different methods and were asked to rate their overall preference on a 1- to-10 scale, wh… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study of Iterative Cross-modal Refinement (ICR) and Temporal Cognitive Calibration (TCC), where inconsistency and the corresponding correct ones are highlighted in red and green boxes, respectively. StoryGPT-V MovieAgent StoryGen Ours 0 2 4 6 8 10 Average Preference Score (1 10) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: User study results showing average preference [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example structured feedback produced during [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Overview of our interactive storybook gen [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional visualizations. (Top) Example 0. (Bottom) Example 1. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional visualizations. (Top) Example 2. (Bottom) Example 3. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Additional visualizations. (Top) Example 4. (Bottom) Example 5. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Representative visualizations of the expert-level long narrative stress test. Each panel corresponds to a [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: A single-page story card used in our long-horizon constraint stress test. Highlighted phrases indicate [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
read the original abstract

Recent advancements in Large Generative Models (LGMs) have revolutionized multi-modal generation. However, generating illustrated storybooks remains an open challenge, where prior works mainly decompose this task into separate stages, and thus, holistic multi-modal grounding remains limited. Besides, while safety alignment is studied for text- or image-only generation, existing works rarely integrate child-specific safety constraints into narrative planning and sequence-level multi-modal verification. To address these limitations, we propose BookAgent, a safety-aware multi-agent collaboration framework designed for high-quality, safety-aware visual narratives. Different from prior story visualization models that assume a fixed storyline sequence, BookAgent targets end-to-end storybook synthesis from a user draft by jointly planning, scripting, illustrating, and globally repairing inconsistencies. To ensure precise multi-modal grounding, BookAgent dynamically calibrates page-level alignment between textual scripts and visual layouts. Furthermore, BookAgent calibrates holistic consistency from the temporal dimension, by verifying-then-rectifying global inconsistencies in character identity and storytelling logic. Extensive experiments demonstrate that BookAgent significantly outperforms current methods in narrative coherence, visual consistency, and safety compliance, offering a robust paradigm for reliable agents in complex multi-modal creation. The implementation will be publicly released at https://github.com/bogao-code/BookAgent/tree/main.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes BookAgent, a multi-agent collaboration framework for end-to-end generation of illustrated storybooks from user drafts. Unlike prior staged approaches, it jointly handles narrative planning, scripting, illustration, and global inconsistency repair. Key mechanisms include dynamic page-level alignment calibration between textual scripts and visual layouts, plus temporal calibration via verification and rectification of character identity and storytelling logic inconsistencies. The central claim is that extensive experiments demonstrate significant outperformance over current methods in narrative coherence, visual consistency, and safety compliance, providing a robust paradigm for safety-aware multi-modal creation.

Significance. If the claimed improvements from the calibration and multi-agent repair mechanisms are quantitatively validated, the work would advance reliable multi-modal narrative generation by addressing holistic grounding limitations and incorporating child-specific safety constraints, which remain underexplored in existing literature on story visualization and generative models.

major comments (2)
  1. [Abstract] Abstract: The assertion that 'extensive experiments demonstrate' significant outperformance in narrative coherence, visual consistency, and safety compliance is load-bearing for the central claim, yet no metrics, baselines, ablation studies, dataset details, or failure-mode analysis are supplied. This prevents assessment of whether the page-level alignment calibration and global inconsistency verification actually close the grounding gap or yield measurable gains over staged baselines.
  2. [Method] Method description (inferred from abstract and §3): The dynamic calibration and multi-agent repair loop are presented as key innovations, but without concrete implementation details, convergence guarantees, or pseudocode, it is unclear how these steps reliably produce the claimed holistic consistency improvements.
minor comments (2)
  1. [Abstract] The abstract references prior works on story visualization and safety alignment only in general terms; specific citations to representative baselines would improve context.
  2. [Abstract] The GitHub link is provided but the manuscript does not indicate whether code, models, or datasets will be released to support reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, providing clarifications and indicating revisions where the manuscript will be updated to improve clarity and completeness.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The assertion that 'extensive experiments demonstrate' significant outperformance in narrative coherence, visual consistency, and safety compliance is load-bearing for the central claim, yet no metrics, baselines, ablation studies, dataset details, or failure-mode analysis are supplied. This prevents assessment of whether the page-level alignment calibration and global inconsistency verification actually close the grounding gap or yield measurable gains over staged baselines.

    Authors: We acknowledge that the abstract is a high-level summary and does not include quantitative details. The full experimental results, including specific metrics (e.g., coherence, consistency, and safety scores), baselines (comparisons to staged story visualization methods), ablation studies on the calibration components, dataset specifications, and failure-mode analysis, are presented in Sections 4 and 5 with supporting tables and figures. To directly address the concern, we have revised the abstract to include key quantitative highlights of the improvements from the calibration and repair mechanisms. revision: yes

  2. Referee: [Method] Method description (inferred from abstract and §3): The dynamic calibration and multi-agent repair loop are presented as key innovations, but without concrete implementation details, convergence guarantees, or pseudocode, it is unclear how these steps reliably produce the claimed holistic consistency improvements.

    Authors: We agree that more explicit implementation details are warranted. In the revised manuscript, we have added pseudocode for the dynamic page-level alignment calibration and the multi-agent temporal repair loop, along with a description of the iterative verification-rectification process and observed empirical convergence (typically within a small number of iterations). Hyperparameters and coordination logic are now detailed in the updated Section 3. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper presents a procedural multi-agent framework for storybook generation with components like dynamic page-level alignment calibration and global inconsistency verification. No equations, mathematical derivations, fitted parameters, or first-principles predictions appear in the provided text. Claims of outperformance rest on experimental results rather than any analytical chain that could reduce to self-definition or fitted inputs by construction. The description remains self-contained without load-bearing self-citations or ansatzes that collapse the central result.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are specified in the abstract; the framework description relies on high-level concepts of multi-agent collaboration and calibration whose internal mechanisms are not detailed.

pith-pipeline@v0.9.0 · 5537 in / 1049 out tokens · 41558 ms · 2026-05-10T09:40:25.971557+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

6 extracted references · 6 canonical work pages

  1. [1]

    Talecrafter: Interactive story visualization with multiple characters

    TaleCrafter: Interactive Story Visualization with Multiple Characters.CoRR, abs/2305.18247. Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2024. CRITIC: large language models can self-correct with tool-interactive critiquing. InICLR. Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Z...

    work page arXiv 2024

  2. [2]

    The Park Bench Placement Challenge

    StoryDALL-E: Adapting Pretrained Text- to-Image Transformers for Story Continuation. In ECCV, pages 70–87. Mir Tafseer Nayeem and Davood Rafiei. 2024. KidLM: Advancing Language Models for Children-Early In- sights and Future Directions.CoRR, abs/2410.03884. Joon Sung Park, Joseph C. O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael...

    work page arXiv 2024

  3. [3]

    The yellow balloon must be **above** the green box

    work page

  4. [4]

    The purple ball must be **to the left of** the green box

    work page

  5. [5]

    There is a tree next to the bench, and the green box must be **to the right of** the tree

    work page

  6. [6]

    The Labels That Lie

    For the photo, Pepe must stand **in front of** the bench, while a kitten named **Mimi** stands **behind** the bench. Pepe checks each rule carefully and fixes one mistake. When the photo is taken, every relation is correct. **“The Labels That Lie”** On a table are two boxes: a **red box** that contains only **blue triangles**, and a **blue box** that cont...

    work page