Reproduce “MEVG: Multi-event Video Generation with Text-to-Video Models” 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
Gyeongrok Oh, Jaehwan Jeong, Sieun Kim, Wonmin Byeon, Jinkyu Kim, Sungwoong Kim, Sangpil Kim
MEVG (Multi-event Video Generation) introduces a training-free diffusion-based method from Korea University that generates continuous videos depicting multiple events from sequential text prompts. It achieves superior temporal consistency, semantic alignment, and realism in multi-event narratives by leveraging last frame-aware latent initialization and structure-guided sampling, alongside an LLM-powered prompt generator.
Key insights
- Achieving superior temporal and visual coherence across distinct events is critical for generating natural and engaging multi-event video narratives.
- Training-free methods, leveraging existing pre-trained T2V diffusion models with intelligent latent space manipulation and guided sampling, can efficiently generate complex narratives.
- An LLM can effectively bridge the gap between complex human language and specific T2V model requirements by intelligently parsing multi-event descriptions into optimal single-event prompts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.