MEV dataset targets multi-event, long-take text-to-video generation
Original title: MEV: A Multi-Event Video Dataset for Long-Take Generation
Fine-tuning with temporally grounded event conditioning improved long-take transitions and content coherence.
The authors introduce MEV, a multi-event video dataset meant to address short, single-event clips and coarse captions in text-to-video data. Videos show subjects performing sequences of distinct, non-overlapping events, with temporal boundaries and procedural annotations. A pipeline segments videos into grounded events and writes global and event-level captions. Under an event-centric protocol, fine-tuning on MEV with temporally grounded event conditioning produced better long-take transitions and more coherent content. The dataset is described as a public resource for multi-event fine-tuning and evaluation.
- Existing datasets are described as mostly short clips of one event or coarse captions.
- Event units are non-overlapping and intended to support localized conditioning.
- Captions are generated at both the full-video level and the individual-event level.
- The paper does not state dataset size, clip length, or model names.
From the original article
1 min read at arxiv.orgCurrent datasets for text-to-video generation are largely limited to short clips depicting a single event or coarse-grained captions. As a result, they provide limited supervision for modeling long-take temporal dependencies, often leading to temporally inconsistent generations when describing multiple events. In this paper, we introduce MEV, a multi-event video dataset designed for long-take video generation.
By Peiyuan Zhu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.