Paper2Video

PaperTalker: a multi-agent framework that generates complete academic presentation videos from a paper — slides, subtitles, cursor grounding, TTS speech, and a talking head — with tree-search visual refinement for layout, plus a benchmark of 101 papers paired with author-made presentations and tailored metrics (Meta Similarity, PresentArena, PresentQuiz, IP Memory). 2.3K★ within months.

From Mike Zheng Shou's ShowLab; sibling of Code2Video (ICML 2026, 1.9K★), the lab's code-centric agentic paradigm for educational video generation.