FreeStory: Training-Free Character Consistency for Free-Form Visual Storytelling

A new training-free framework called FreeStory aims to solve a persistent problem in AI-driven visual storytelling: keeping characters visually consistent when prompts use free-form language instead of rigid, repeated descriptions, according to research posted on arXiv [1]. The work, submitted on June 23, 2026, reformulates character consistency as “entity-grounded feature reuse” [1]. In natural storytelling, a character is typically introduced once and later referenced with pronouns or type-based expressions. Existing training-free methods, which reuse attention features to maintain identity, depend on structured prompts where full character descriptions appear in every line — a condition that diverges from how people actually write narratives [1]. FreeStory addresses the gap by associating reference mentions with their corresponding character descriptions. It combines dynamic character masks, correspondence-aware feature matching, key-value injection, and query blending to preserve identity while retaining generation diversity [1]. The researchers also introduce FreeStoryBench, a benchmark that includes both single- and multi-character stories under free-form prompts [1]. Visual storytelling sits at the intersection of text-to-video generation and narrative theory. Text-to-video models, driven largely by video diffusion models, have advanced rapidly during the 2020s, producing high-quality, text-conditioned videos [2]. The narrative paradigm, a communication theory articulated by scholar Walter Fisher, holds that all meaningful communication occurs through storytelling and that humans naturally understand complex information through narrative [4]. Computational creativity research, meanwhile, seeks to model or simulate human-like creative processes, with one applied branch known as media synthesis [5]. The FreeStory paper appears on arXiv, an open-access repository that hosts preprints across mathematics, physics, computer science, and related fields [9]. As of November 2024, the repository was receiving about 24,000 submissions per month [9]. The platform also supports arXivLabs, a framework enabling community collaborators to build tools such as bibliographic explorers and code-finding services that sit alongside paper abstracts [7][8]. Experiments reported in the paper show that FreeStory achieves state-of-the-art performance among training-free methods on structured benchmarks and delivers stronger overall consistency than baselines when tested under free-form prompts [1]. The authors did not provide numerical metrics in the abstract, and no external validation of the results was available at the time of this report.