Helping Figures Tell their Story! Paper-Grounded Video Generation Explaining Complex Scientific Figures

A new pipeline called MINARD can automatically generate narrated walkthrough videos that explain complex scientific figures by grounding each step of the narration in specific regions of the figure, drawing directly from the accompanying research paper [1]. The system, formally named Multimodal Interpretation of Narrated Architecture via Region Decomposition, addresses a gap in current video generation tools, which have lacked the ability to produce paper-grounded, step-by-step visual explanations aligned with figure highlights [1][2]. MINARD generates these narrations and sequentially links them to components within a figure, creating a guided tour of the visual data [2]. To evaluate its performance, the researchers also introduced FigTalk, a benchmark that includes new metrics for measuring sequential and component-level grounding [1][2]. In tests on this benchmark, MINARD produced narrations that were rated as humanlike and faithful to the source paper, outperforming existing approaches in both automatic metrics and human evaluations [2]. The work enters a growing field of systems that use large language models to translate dense academic content into dynamic visual formats. Another system, manimator, takes a research paper or natural language prompt and generates explanatory animations using the Manim engine, employing an agentic architecture where one model creates a structured scene description and another translates it into executable code [3]. A separate pipeline, CogniPresent, frames the generation of academic presentation videos as a multimedia learning problem, using the Cognitive Theory of Multimedia Learning to guide slide creation, narration, and audio pacing so that narration complements rather than repeats on-screen text [4]. A broader effort to automate full presentation videos highlights the coordination challenges involved, including slide design, synchronized subtitles and speech, a talking-head presenter, and a cursor indicator to anchor audience attention [5]. The MINARD system focuses more narrowly on the figure itself, treating it as a canvas that compresses a complex pipeline and requires a dedicated narrative to unpack [1][2].