ViMix-14M: A Curated Multi-Source Video-Text Dataset

Dataset Description

ViMix-14M is a large-scale video-text dataset containing ~14 million video-text pairs with multi-granularity captions, designed to address the data bottleneck in text-to-video generation.

Text-to-video generation has surged in interest since Sora, yet open-source models still face a data bottleneck: there is no large, high-quality, easily obtainable video–text corpus. Existing public datasets typically require manual YouTube crawling, which yields low usable volume due to link rot and access limits, and raises licensing uncertainty. This work addresses this challenge by introducing ViMix-14M, a curated multi-source video–text dataset of around 14 million pairs that provides crawl-free, download-ready access and long-form, high-quality captions tightly aligned to video. ViMix-14M is built by merging diverse open video sources, followed by unified de-duplication and quality filtering, and a multi-granularity, ground-truth-guided re-captioning pipeline that refines descriptions to better match actions, scenes, and temporal structure. We evaluate the dataset by multimodal retrieval, text-to-video generation, and video question answering tasks, observing consistent improvements over counterpart datasets. We hope this work can help removing the key barrier to training and fine-tuning open-source video foundation models, and provide insights of building high-quality and generalizable video-text datasets.

Links

• Paper: https://arxiv.org/abs/2511.18382
• Code: https://github.com/yangtiming/ViMix-14M

Dataset Statistics

Metric	Value
Total Videos	13.7M
Total Duration	22.8K hours
Average Duration	6.0 seconds
Caption Lengths	Short: 14.4 words Middle: 50.6 words Long: 109.8 words

Source Datasets

Dataset	Videos	Duration
InternVid-10M-FLT	10.6M	16.3K hrs
VideoUFO	1.09M	2.12K hrs
VidGen-1M	1.00M	2.26K hrs
Kinetics-700	0.63M	1.58K hrs
Something-Something V2	0.22M	234 hrs
OpenVideo	0.11M	331 hrs
UCF-101	0.01M	26.6 hrs

Dataset Structure

Data Fields

{
  "segment_id": "unique_video_identifier",
  "dataset_name": "source_dataset_name",
  "split": "train/val/test",
  "duration": 6.5,
  "use_gt_label": true,
  "start_time": "00:00:00.000",
  "end_time": "00:00:06.500",
  "caption_short_en": "Brief 10-20 word description",
  "caption_middle_en": "Detailed 40-60 word description with colors, background, style, and actions",
  "caption_long_en": "Comprehensive 80-130 word description with spatial relations, attributes, and reasoning",
  "vbench_scores": {
    "subject_consistency": 0.84,
    "background_consistency": 0.87,
    "temporal_flickering": 0.88,
    "aesthetic_quality": 0.37,
    "imaging_quality": 63.81
  }
}

Citation

If you use ViMix-14M in your research, please cite:

@article{yang2025vimix,
  title={ViMix-14M: A Curated Multi-Source Video-Text Dataset with Long-Form, High-Quality Captions and Crawl-Free Access},
  author={Yang, Timing and Ren, Sucheng and Yuille, Alan and Wang, Feng},
  journal={arXiv preprint arXiv:2511.18382},
  year={2025}
}

Acknowledgements

We thank the creators of: InternVid, VideoUFO, VidGen-1M, Kinetics-700, Something-Something V2, OpenVideo, UCF-101, Qwen2.5-VL, and VBench.