How Do Different Components Affect Story Illustration? From Narrative Processing to Prompt Repair

Illustrated stories support children’s learning. They also aid digital publishing and preserve regional narratives. However, manual illustration is costly and difficult to scale. Diffusion models offer an alternative. Yet, they often omit narrative details. They may also alter recurring characters across scenes. This work tries to find answers to three research questions concerning diffusion-based story visualization. First, we study how scene segmentation and part-of-speech filtering influence narrative alignment. Second, we evaluate IP-Adapter and consistent self-attention for cross-scene identity preservation and examine their trade-offs with scene diversity and narrative fidelity. Finally, we propose an attention-guided prompt-repair method that measures the grounding of required entities, attributes, relations, and character states and then iteratively revises prompts. Our experiments indicate that image conditioning improves character continuity but can restrict scene variation. The proposed framework converts diffusion attention from a post hoc explanation into an automatic signal for diagnosing and repairing story prompts.