Multimodal generative models for storytelling

Abstract

Storytelling is an open-ended task that entails creative thinking and requires a constant flow of ideas. Generative models have recently gained momentum thanks to their ability to identify complex data's inner structure and learn efficiently from unlabeled data [34]. Natural language generation (NLG) for storytelling is especially challenging because it requires the generated text to follow an overall theme while remaining creative and diverse to engage the reader [26]. Competitive story generation models still suffer from repetition [19], are unable to consistently condition on a theme [51] and struggle to produce a grounded, evolving storyboard [43]. Published story visualization architectures that generate images require a descriptive text to depict the scene to illustrate [30]. Therefore, it seems promising to evaluate an interactive multimodal generative platform that collaborates with writers to face the complex story-generation task. With co-creation, writers contribute their creative thinking, while generative models contribute to their constant workflow. In this work, we introduce a system and a web-based demo, FairyTailor¹, for machine-in-the-loop visual story co-creation. Users can create a cohesive children's story by weaving generated texts and retrieved images with their input. FairyTailor adds another modality and modifies the text generation process to produce a coherent and creative sequence of text and images. To our knowledge, this is the first dynamic tool for multimodal story generation that allows interactive co-creation of both texts and images. It allows users to give feedback on co-created stories and share their results. We release the demo source code² for other researchers' use.