Storycraft: Short Story Generation Using Deep Learning

This research paper sheds light on the development of a multimodal story generator known as StoryCraft, that uses advanced deep-learning techniques to produce coherent and engaging narratives from visual inputs according to different genres. Building upon the capabilities of large language models such as the LLaVa model for generating captions from the visual inputs, the GPT2 model for generating small texts from the captions and lastly the LLaMa model for generating contextually and grammatically correct stories from the small texts tailored to various genres including horror, comedy, romance, children’s, adventure, mystery and fantasy. The proposed model integrates visual and textual data to increase storytelling quality, overcoming restrictions of existing models that struggle with generating grammatically correct narratives.Apart from overcoming the limits of the previous models it also makes use of transformers rather than the traditional CNN and RNN system leading to the generation of not-so-short stories. The analysis of the generated stories is conducted through an amalgamation of automated metrics, such as ROUGE-L and BLEU scores and human evaluation metric as well, alongside a comparative analysis with different models to prove which model performs better and gives desired output. Results depict that the Storycraft model (LLaMA) outperforms previous state-of-the-art systems, demonstrating valuable enhancements in narrative quality and user satisfaction.