Reproduce “CineEvolve: An Agentic Self-Evolution Framework for Cinematic Multi-Shot Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingzhe Zhang, Peixuan Zhang, Genying Li, Bing Ma, Shuchen Weng, Si Li, Boxin Shi
CineEvolve trains a vision-language policy to plan and revise storyboards as Director and diagnose defects as Critic. Alternating reinforcement learning updates the roles, while E-GRPO adds revision safeguarding and historical anchoring; image and video backbones remain frozen. On held-out CineTraj, Cross-shot Consistency is 0.8893 versus ShotStream's 0.8467; CineEvolve leads on eight of nine metrics, while MultiShotMaster leads Background Consistency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.