[4Xin2] Poster session 2

[4Xin2-20] Narrative-to-image: Automatic Generation of Images Matching Narrative

〇Sho Iwasaki¹, Shuji Awai¹, Toshihiko Aoki¹, Shohei Ishizuka¹, Takeshi Konno¹ (1.Fujitsu Limited)

Keywords:Narrative, Image Generation, Generative AI

This paper proposes a novel method for automatically generating images that align with narratives. In recent years, "narrative" has gained attention for its ability to make complex information more understandable and emotionally engaging. Given that human information processing is visually dominant, the combined use of narratives and images is expected to be effective in information transmission. Therefore, our goal is to utilize text-to-image method to create images that align with narrative texts. Our method gradually concretizes the visualization of the narrative in step-by-step manner, thereby generating image generation prompts that align with the narrative's theme. Moreover, multiple generated images are evaluated using a Vision Language Model, allowing for the selection and output of the most fitting images. Subjective evaluations of the generated images suggest that the proposed method is successful in creating images that are consistent with the narrative.