NeurIPS 2026

MEV: A Multi-Event Video Dataset for Long-Take Generation

Peiyuan Zhu1 Shaoan Xie2 Yifan Shen1 Wen Tian2 Zijian Li1,2 Guangyi Chen1,2 Kun Zhang2,1
1 Mohamed bin Zayed University of AI 2 Carnegie Mellon University

Long-take videos with precise event boundaries and hierarchical captions, camera-aware descriptions, and generation-oriented quality scores.

Paper Data Code BibTeX

TL;DR

Existing video-text datasets often provide short clips or coarse captions, which weakens supervision for long-take temporal structure. MEV instead decomposes real-world videos into reproducible atomic events with explicit boundaries, hierarchical captions, camera-aware event descriptions, and quality profiling. This makes temporally localized conditioning easier to study and improves multi-event long-video adaptation in downstream generation experiments.

Dataset Overview

Designed around temporal structure rather than clip-level slogans

MEV pairs long-take source videos with event boundaries, hierarchical captions, camera motion, semantic tags, and quality scores.

MEV overview figure showing event-level captions and global video annotations
227K raw candidate clips from MixKit and YouTube
720p-4K source quality range used during collection
8,016 / 25,262 long-take videos and event-level video segments

Showcase

Data Showcase

Long-take examples pair playback with event captions, camera motion, semantic attributes, and quality scores.

Annotation Taxonomy

A schema built for controllable generation and structured analysis

Each record combines video-level scene context with event timestamps, captions, camera labels, semantic attributes, and VBench quality scores.

Level Contents Research Use
Global fields Background, theme, style, shot type, camera movement, lighting, atmosphere, subject profile Whole-video semantics and controllable prompting
Global summaries Short, middle, and long captions Prompt granularity control and retrieval evaluation
Event descriptors Start and end timestamps, event caption, camera label, camera motion description Temporally localized conditioning and event reasoning
Semantic attributes Appearance, disappearance, multi-person, lighting change, animal, cooking, sports, travel, child, environmental change Subset curation and challenge-focused analysis
Quality profile Global- and event-level VBench scores: temporal flickering, subject consistency, background consistency, aesthetic quality, imaging quality, motion smoothness Quality-aware filtering and benchmark diagnostics

Schema Sketch

A representative record layout

{
  "global_prompt": {
    "background": "...",
    "theme": "...",
    "style": "...",
    "shot_type": "...",
    "camera_movement": "...",
    "lighting": "...",
    "atmosphere": "...",
    "subject_profile": "...",
    "short_caption": "...",
    "middle_caption": "...",
    "long_caption": "..."
  },
  "events": [
    {
      "start_time": 0.0,
      "end_time": 4.2,
      "caption": "...",
      "camera_motion_label": "pan-right",
      "camera_motion_desc": "...",
      "focus_shift": null,
      "has_appearance": false,
      "has_cooking": false,
      "has_sports": false,
      "has_travel": true,
      "has_child": false
    }
  ],
  "vbench_scores": {
    "subject_consistency": 0.0,
    "background_consistency": 0.0,
    "motion_smoothness": 0.0
  }
}

Statistics and Analysis

Multiple views of temporal, linguistic, and camera structure

Duration, event-count, caption-length, camera-motion, and semantic-attribute distributions expose both structure and diversity in MEV.

Statistical analysis figure for the MEV dataset
Statistical distributions of MEV from temporal, caption, camera, and semantic perspectives.

Observed Patterns

What the statistics emphasize

  • Covered durations cluster around tens of seconds while keeping a long tail.
  • Total covered duration and event count are related but remain diverse.
  • Non-static camera motion is profiled explicitly rather than hidden in captions.
  • Short, middle, and long captions show clear length separation.
  • Appearance and disappearance attributes expose event-local transition-heavy subsets.
  • Word distributions confirm that captions are action-centered rather than purely scenic.

Experiments

MEV is evaluated as both data and supervision

MEV-Wan is measured with event-centric MEV-Eval scores and generic VBench quality metrics against Wan2.1-1.3B, LongLive, and Ultra-Wan.

MEV-Eval

Event-centric generation scores

EEA, EOA, TGA, and BTC evaluate execution, ordering, temporal grounding, and transition coherence.

Method EEA EOA TGA BTC
Wan2.1-1.3B 0.42 0.20 0.33 0.41
LongLive 0.47 0.39 0.48 0.58
Ultra-Wan 0.46 0.33 0.42 0.50
MEV-Wan 0.61 0.45 0.60 0.64

VBench

Generic video quality scores

Average over five random 128-prompt draws from the MEV test set.

Method Subj. Back. Flick. Smooth. Aesth. Image
Wan2.1-1.3B 0.954 0.949 0.989 0.991 0.585 61.674
LongLive 0.949 0.950 0.986 0.990 0.504 61.953
Ultra-Wan 0.936 0.956 0.990 0.987 0.557 63.328
MEV-Wan 0.960 0.952 0.990 0.991 0.590 64.924
Comparison of caption granularity in video generation
Higher caption granularity improves alignment with ground-truth event progression, while current generators miss hard local events.
Qualitative comparison of multi-event generation results
MEV-Wan shows smoother transitions and stronger temporal consistency than Wan2.1, LongLive, and the Ultra-Wan baseline.

Citation

BibTeX 
@misc{zhu2026mevmultieventvideodataset,
  title = {MEV: A Multi-Event Video Dataset for Long-Take Generation},
  author = {Peiyuan Zhu and Shaoan Xie and Yifan Shen and Wen Tian and
            Zijian Li and Guangyi Chen and Kun Zhang},
  year = {2026},
  eprint = {2610.38241},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2610.38241}
}