Scientific figures compress complex pipelines into a single canvas, but understanding them requires paper-grounded, step-by-step narration aligned with visual highlights—a capability missing from current video generation systems and benchmarks. To address this, we introduce paper-grounded figure-to-video generation: generating narrated, region-grounded walkthrough videos from a figure and its paper. Second, we develop a multi-agent pipeline MINARD, which first generates paper-grounded narration and then ground the figure regions through staged extraction and clustering. Third, we release FigTalk, a benchmark derived from Paper2Video with new sequential and component-level grounding metrics that correlate strongly with human judgments. On FigTalk, MINARD produces human-like narrations grounded in paper evidence, and MINARD-Grounding substantially outperforms existing VLM grounding, animation-based, and diffusion-style video generation methods in both automatic and human evaluation.