BookAgent: Multi-Agent Storybook Generation

• BookAgent is a multi-agent framework that transforms rough drafts into aligned textual and visual storybooks while enforcing child safety.
• It employs a three-stage process—VAS, ICR, and TCC—to iteratively refine page scripts and illustrations for coherence and consistency.
• Empirical evaluations confirm significant improvements in text-image fidelity, character consistency, and overall narrative quality.

BookAgent is a safety-aware multi-agent collaboration framework for end-to-end illustrated storybook generation that transforms a rough user draft into a multi-page book of page scripts and page illustrations while explicitly enforcing cross-modal coherence, long-range consistency, and child-oriented safety (Gao et al., 17 Apr 2026). It is designed for the setting in which prior story visualization systems either assume that the story text is already fixed or decompose planning and illustration into separate one-way stages with weak feedback. BookAgent instead treats storybook creation as a closed-loop cognitive process in which planning, writing, illustration, checking, and repair interact through structured intermediate representations and iterative verification (Gao et al., 17 Apr 2026).

1. Task definition and problem formulation

BookAgent formalizes illustrated storybook synthesis as a mapping from a coarse specification to a sequence of aligned textual and visual pages. The input consists of a user draft $x$, an optional inspiration image $I_0$, a target page count $K$, and a global style descriptor $s$. The output is a storybook

$\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$

where $t_i$ is the script for page $i$ and $y_i$ is the corresponding illustration (Gao et al., 17 Apr 2026).

The framework is motivated by three coupled difficulties in long-form multimodal generation. First, it requires page-level multimodal grounding: the script and illustration on each page must describe the same events, entities, and attributes. Second, it requires long-horizon consistency: recurring characters, props, colors, counts, and event order must remain stable across the sequence. Third, in the child-oriented setting, safety cannot be treated as a generic post-hoc NSFW filter, because unsafe trajectories can originate in narrative planning, script generation, image generation, or later sequence composition (Gao et al., 17 Apr 2026).

The system is posed as a constrained optimization problem: $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$ Here $\hat{x}$ is the refined draft, $I_0$0 is page-level text-image faithfulness, $I_0$1 is identity consistency relative to character reference sheets, and $I_0$2 is global sequence coherence. The hard constraints $I_0$3 and $I_0$4 encode text and image safety decisions (Gao et al., 17 Apr 2026).

2. Three-stage architecture and role-based decomposition

BookAgent is organized as a closed-loop three-stage system: Value-Aligned Storyboarding (VAS), Iterative Cross-modal Refinement (ICR), and Temporal Cognitive Calibration (TCC). Its engineering premise is that different failure modes require specialized role-based agents with fixed input-output contracts rather than a single monolithic generator (Gao et al., 17 Apr 2026).

Stage	Primary modules	Function
VAS	Reviewer–Refiner, Text Safety Auditor, Character Extractor, Character Sheet Renderer, Page Planner	Refine the draft, enforce child safety, extract recurring characters, generate page scripts and initial prompts
ICR	Image Generator, Frame Director, Identity Director, Image Safety Auditor	Generate page images, score faithfulness and identity consistency, reject unsafe candidates, revise prompts
TCC	Sequence Director	Audit the full book, detect global inconsistencies, localize faulty pages, trigger selective repair

Within VAS, the Reviewer–Refiner $I_0$5 rewrites the raw user draft so that it fits the target page count, improves coherence, reduces ambiguity, and keeps the recurring cast manageable; the role table specifies at most 5 recurring characters. The Text Safety Auditor $I_0$6 checks whether the refined draft is child-safe and returns both a binary decision and a natural-language rationale if rewriting is required. The Character Extractor $I_0$7 extracts a character set $I_0$8 with $I_0$9, and the Character Sheet Renderer $K$0 converts each descriptor into a canonical neutral-background reference image $K$1 (Gao et al., 17 Apr 2026).

The Page Planner $K$2 then produces

$K$3

where each $K$4 is page-level text and each $K$5 is an initial illustration prompt encoding local semantics and global style. ICR uses an Image Generator $K$6 to propose page illustrations, a Frame Director $K$7 to evaluate semantic alignment, an Identity Director $K$8 to detect character drift, and an Image Safety Auditor $K$9 to reject unsafe images. TCC adds a Sequence Director $s$0 that reviews the provisional full book for character drift, plot contradictions, prop inconsistency, style discontinuity, and temporal logic errors (Gao et al., 17 Apr 2026).

A distinctive feature is the use of structured intermediate representations rather than unconstrained free-form messages. These include the refined draft $s$1, character set $s$2, reference sheets $s$3, page plan $s$4, page-level scores and critiques $s$5, identity scores and critiques $s$6, safety labels and rationales $s$7 and $s$8, sequence-level outputs $s$9, and local page memory $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$0 (Gao et al., 17 Apr 2026).

3. Multi-agent cognitive calibration

“Multi-agent cognitive calibration” is the paper’s term for using multiple specialized agents to repeatedly inspect, score, critique, and revise generated content so that the evolving storybook remains aligned at both local and global levels (Gao et al., 17 Apr 2026).

At page level, ICR instantiates a generate–verify–revise loop. For page $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$1, BookAgent first gathers the relevant reference set

$\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$2

At refinement attempt $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$3, it generates an illustration

$\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$4

The Frame Director produces a faithfulness score $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$5 and semantic issues $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$6, while the Identity Director produces an identity score $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$7 and identity issues $\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$8. Prompt revision follows the rule

$\mathcal{B} \triangleq \{(t_i, y_i)\}_{i=1}^{K},$9

where $t_i$0 converts image-safety rationale into negative prompt constraints and $t_i$1 aggregates semantic and identity critiques with page memory $t_i$2 to avoid regression (Gao et al., 17 Apr 2026).

A page candidate is accepted only if it is safe, sufficiently faithful, and sufficiently identity-consistent, expressed by thresholds $t_i$3 and $t_i$4. If retries are exhausted, the system keeps the best safe candidate: $t_i$5 This mechanism turns prompt engineering into an explicit corrective process mediated by multimodal critics rather than a single-shot generation step (Gao et al., 17 Apr 2026).

At sequence level, TCC performs verify-then-rectify calibration over the provisional book. The Sequence Director computes

$t_i$6

where $t_i$7 is a global consistency score, $t_i$8 is a set of sequence-level critiques, and $t_i$9 is the set of inconsistent page indices. If $i$0, the system selectively repairs pages $i$1 rather than regenerating the entire book. Those pages re-enter the ICR loop with prompt updates derived from the global critique set $i$2 and stricter reference conditioning (Gao et al., 17 Apr 2026).

4. Grounding, identity, and child-oriented safety

BookAgent treats consistency as a multi-scale constraint. Local grounding is enforced by jointly checking page text $i$3, initial prompt $i$4, page-specific references $i$5, frame-level faithfulness $i$6, identity-level fidelity $i$7, and local memory $i$8. Character consistency is achieved by extracting recurring characters during VAS, rendering canonical character sheets $i$9, conditioning the image generator on relevant references, and using the Identity Director to detect changes in species, color, clothing, missing characters, or extra characters (Gao et al., 17 Apr 2026).

Global continuity is handled by TCC and the Sequence Director. The benchmark design makes explicit that the framework is intended to preserve exact count invariants, spatial relations, color-shape bindings, temporal order, map continuity, inventory tracking, route order, and object placement. The paper describes benchmark-level consistency as the fraction of all rule constraints satisfied by a generated narrative, operationalizing sequence coherence through rule satisfaction over extracted attributes such as counts, spatial relations, and identities (Gao et al., 17 Apr 2026).

Safety is embedded as a structural constraint rather than a terminal filter. The safety auditors are defined as

$y_i$0

where $y_i$1 is a binary safe/unsafe decision and $y_i$2 is a natural-language rationale such as “violent content detected” (Gao et al., 17 Apr 2026).

The framework inserts safety checks at four points. In narrative planning safety, the raw draft is rewritten into $y_i$3 and audited before page planning begins. In script generation safety, the Page Planner only decomposes a safety-validated story. In image generation safety, each page candidate is audited, and unsafe outputs are rejected while the prompt is hardened through $y_i$4. In later verification, multimodal safety auditing remains part of iterative refinement and sequence-level calibration. The content categories explicitly named include NSFW content, toxic or age-inappropriate content, violent content, sexually explicit material, nudity, and harmful or sensitive concepts unsuitable for children; the paper also cites accidental nudity in a child-oriented context as a motivating failure case (Gao et al., 17 Apr 2026).

5. Implementation, benchmarks, and empirical results

BookAgent is implemented as an inference-time orchestration framework rather than as a newly trained unified model. The reported instantiation uses Google Gemini 3.0 for reasoning and Nano-Banana for generation. The framework uses role-specific prompts and structured outputs, reference-conditioned image generation, and a local page memory $y_i$5 that stores prior semantic and identity critiques during ICR. The appendix reports a frame-level verification threshold $y_i$6, maximum frame retry attempts of 3, a sequence-level verification threshold $y_i$7, and maximum sequence retry attempts of 1; the paper also reports loose, default, and strict settings with $y_i$8 and $y_i$9 varied accordingly (Gao et al., 17 Apr 2026).

The evaluation benchmark comprises 16 multi-page stories spanning 5–20 pages, with 170 scene-level evaluation units, 40+ distinct named characters, 60+ object categories, and 40+ rule groups. Automatic evaluation uses ensembles of large multimodal models that rate each story on a 1–5 scale for Image-Text Consistency, Cross-Frame Character Consistency, and Safety. Human evaluation asks parents of children aged 4–10 to rate overall preference on a 1–10 scale (Gao et al., 17 Apr 2026).

Method	Image-Text Consistency	Cross-Frame Character Consistency	Safety
StoryGPT-V	3.1	2.4	4.5
MovieAgent	2.8	2.1	3.6
StoryGen	2.5	1.9	4.4
BookAgent	4.6	4.7	4.8

These results show the largest gain in cross-frame consistency: BookAgent improves over StoryGPT-V by +2.3 on Cross-Frame Character Consistency, while also improving by +1.5 on Image-Text Consistency and +0.3 on Safety. The paper identifies MovieAgent as the most comparable baseline because it also uses multi-agent generation, although the comparison remains structurally imperfect because many prior systems assume fixed text rather than co-generating text and images (Gao et al., 17 Apr 2026).

Configuration	Img–Txt	Cross-Frame	Safety
Baseline (no VAS, ICR, TCC)	2.7	2.0	4.2
+ VAS	2.8	2.1	4.8
+ VAS + ICR	4.2	2.4	4.8
+ VAS + ICR + TCC	4.6	4.7	4.8

The progressive ablation isolates the contribution of each stage. VAS primarily boosts safety from 4.2 to 4.8; ICR primarily boosts local text-image grounding from 2.8 to 4.2; TCC primarily boosts global cross-frame consistency from 2.4 to 4.7. Qualitative cases labeled Milo and Rowan are used to illustrate accessory drift, object drift, and failures on exact symbolic constraints such as button counts; the appendix stress test also highlights preservation of fixed clothing attributes, stage-right placement of a prize chest, exact ticket counts, and reward mappings across a long narrative including an intentional violation and correction event (Gao et al., 17 Apr 2026).

The reported runtime scales roughly linearly with page count. Approximate costs under default retries are $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$0K tokens and $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$1 minutes for 5 pages, $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$2K tokens and $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$3 minutes for 10 pages, and $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$4K tokens and $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$5 minutes for 20 pages. The default style is a whimsical, soft-color children's picture-book style, shared across text and images for cross-modal consistency (Gao et al., 17 Apr 2026).

6. Limitations and position in the agent-systems literature

The paper explicitly acknowledges three main limitations: dependence on the underlying foundation models, best performance when tracking a relatively small set of characters and objects, and extra computational cost due to iterative verification and repair. It further notes that scaling to larger casts is challenging because of memory interference and calibration complexity (Gao et al., 17 Apr 2026).

Within the broader 2026 agent literature, BookAgent belongs to a class of systems that treat agentic orchestration as a first-class design object. AgentSelect formalizes narrative query-to-agent recommendation over capability profiles $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$6, where $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$7 is a backbone LLM and $$\max_{\hat{x},\,\{y_i\}_{i=1}^{K}} \sum_{i=1}^{K}\left[\alpha(t_i,y_i)+\eta(y_i,\{r_j\})\right]+\lambda\beta(\mathcal{B}) \quad \text{s.t.}\quad \mathcal{S}_T(\hat{x})=1,\ \mathcal{S}_I(y_i)=1,\ \forall i=1,\dots,K.$$8 is a set of external tools; its results emphasize that in sparse long-tail agent catalogs, content-aware capability matching is more reliable than ID-centric collaborative methods (Shi et al., 4 Mar 2026). That formulation is relevant to any setting in which frameworks such as BookAgent are cataloged as deployable configurations rather than treated merely as prompts or model variants.

AgentFlow, by contrast, addresses governance and security for source-code agents through a framework-agnostic Agent Dependency Graph whose typed nodes represent agents, prompts, models, capabilities, memory states, and control policies, and whose analyses include Agent BOM generation and prompt-to-tool risk detection (Wang et al., 2 Jul 2026). For systems with BookAgent-like orchestration, such static analysis is directly pertinent because BookAgent’s behavior depends on bindings among prompts, reference assets, memory, critics, and repair policies rather than only on ordinary host-language control flow.

In that landscape, BookAgent is most precisely understood as an inference-time orchestration framework for long-form multimodal narrative generation, centered on explicit planning, identity anchoring, iterative page-level calibration, sequence-level selective repair, and child-safety guardrails. Its primary contribution is not a new generator backbone, but a closed-loop architecture that reframes storybook synthesis from one-shot generation into a sequence of generation, verification, calibration, and repair (Gao et al., 17 Apr 2026).