Storytelling is a powerful educational tool, especially for young learners, as it fosters engagement and supports language development. The addition of images increases attention, reinforces meaning, and enhances comprehension. However, creating high-quality, coherent visuals for educational stories is often costly, time-consuming, and challenging. Generative AI (GenAI) offers a promising low-cost alternative through text-to-image generation. However generated images frequently exhibit issues such as semantic inconsistencies, inappropriate content, and a lack of visual coherence across scenes, undermining their educational value. This methodological paper investigates how to address this image-generation challenge. Our research explores strategies to support non-technical users in generating coherent and high-quality image sequences that align with educational goals. Central to this approach is Descriptor, a ChatGPT-based assistant that supports iterative prompt refinement. Drawing on a large-scale storytelling project for primary education, we present practical techniques and a preliminary qualitative case study showing how Descriptor-generated prompts improve image consistency and support effective visual storytelling for educational purposes.