Story visualization is an under-explored task that falls at the intersection\nof many important research directions in both computer vision and natural\nlanguage processing. In this task, given a series of natural language captions\nwhich compose a story, an agent must generate a sequence of images that\ncorrespond to the captions. Prior work has introduced recurrent generative\nmodels which outperform text-to-image synthesis models on this task. However,\nthere is room for improvement of generated images in terms of visual quality,\ncoherence and relevance. We present a number of improvements to prior modeling\napproaches, including (1) the addition of a dual learning framework that\nutilizes video captioning to reinforce the semantic alignment between the story\nand generated images, (2) a copy-transform mechanism for\nsequentially-consistent story visualization, and (3) MART-based transformers to\nmodel complex interactions between frames. We present ablation studies to\ndemonstrate the effect of each of these techniques on the generative power of\nthe model for both individual images as well as the entire narrative.\nFurthermore, due to the complexity and generative nature of the task, standard\nevaluation metrics do not accurately reflect performance. Therefore, we also\nprovide an exploration of evaluation metrics for the model, focused on aspects\nof the generated frames such as the presence/quality of generated characters,\nthe relevance to captions, and the diversity of the generated images. We also\npresent correlation experiments of our proposed automated metrics with human\nevaluations. Code and data available at:\nhttps://github.com/adymaharana/StoryViz\n