CineDance_01 is a gated, non-commercial research release. The information below is collected to review and administer dataset access. By submitting a request, you agree to comply with the dataset license, applicable laws, and third-party rights.

CineDance-1M — Release Part 1 of 4

CineDance: Towards Next-Generation Multi-Shot Long-Form Cinematic Audio-Video Generation

Project page · Paper · PDF · Code and release updates

Overview

CineDance-1M is a large-scale research dataset introduced for multi-shot, long-form joint audio-video generation. The accompanying work combines the dataset with CineBench, a human-aligned benchmark, and the CineDance generation model to study cinematic narrative generation in open research.

The paper reports that the complete CineDance-1M collection averages 92.8 seconds and 24.2 continuous shots per sequence. Its data-governance pipeline has three stages: diverse collection and cleaning, film-theory-inspired narrative parsing, and hierarchical dual-modality annotation.

This repository, CineDance_01, is the first of four planned release parts. It contains approximately one quarter of the planned CineDance-1M collection. This release currently includes the video archives only. The structured annotations described in the paper are not included in this repository at this stage.

Release snapshot

Item	CineDance_01
Release position	Part 1 of 4
Video clips	240,488
TAR archives	150
Archive names	`cinedance-video-part0000.tar`–`cinedance-video-part0149.tar`
Total archive size	5,829,783,020,544 bytes (5.83 TB)
Current payload	Video only
Access	Gated, manual review

Each TAR is a packaging shard rather than a semantic or train/validation/test split. Archives have a flat layout: extracting an archive places the MP4 files directly into the target directory. Released filenames use opaque UUID-form identifiers instead of source-title-based names. Native audio is retained in the video containers where available.

The remaining three release parts are still being prepared and will be announced through the project page and code repository.

Download

Access must first be requested through this dataset page and approved. Then authenticate with a Hugging Face user token:

hf auth login

Because the complete current release is 5.83 TB, downloading only the required shards is recommended:

hf download CineDance/CineDance_01 \
  --repo-type dataset \
  --include "video/cinedance-video-part0000.tar" \
  --local-dir ./CineDance_01

To download all currently released video shards, change the include pattern to video/*.tar. Ensure that the destination has sufficient capacity for both the archives and any extracted videos.

Extract one shard with:

mkdir -p ./CineDance_01/videos
tar -xf ./CineDance_01/video/cinedance-video-part0000.tar \
  -C ./CineDance_01/videos

Intended use

CineDance_01 is intended for non-commercial research and education concerning, for example:

multi-shot and long-form video generation;
joint audio-video modeling and temporal coherence;
cinematic video understanding, retrieval, and representation learning; and
evaluation and analysis of long-form generative systems.

Gated access does not transfer ownership of, or grant additional rights in, third-party material that may appear in the videos. Users are responsible for complying with the dataset license, applicable laws, and third-party rights. The dataset must not be used for unlawful, privacy-invasive, harmful, or commercial purposes.

Limitations and release notes

This is a partial release: only the first of four planned parts is available.
This repository currently contains video only; annotations are not included.
No official train/validation/test split is defined by the packaging shards.
Cinematic data can reflect geographic, cultural, genre, language, and production biases present in its sources.
Automated and manual curation cannot guarantee removal of every low-quality, sensitive, or otherwise undesirable example. Users should perform checks appropriate to their application.

Paper

The CineDance paper introduces three connected components:

CineDance-1M, a large-scale cinematic dataset for multi-shot, long-form text-to-audio-video research;
CineBench, a 1,000-case benchmark with six human-aligned evaluation dimensions; and
CineDance, an audio-video generation model adapted from LTX-2.3 that is used to validate the curation strategy and dataset quality.

For implementation updates, evaluation assets, and future model releases, see the official code repository.

Authors

Yuheng Chen — Shanghai Jiao Tong University; equal contribution
Teng Hu — Shanghai Jiao Tong University; equal contribution
Yuji Wang — Shanghai Jiao Tong University
Qingdong He — University of Electronic Science and Technology of China
Zhucun Xue — Zhejiang University
Qianyu Zhou — The University of Tokyo
Xiangtai Li — Nanyang Technological University
Lizhuang Ma — Shanghai Jiao Tong University; corresponding author
Jiangning Zhang — Zhejiang University; project lead
Dacheng Tao — Nanyang Technological University

The author list above follows the arXiv record for the paper.

License and access

This dataset repository is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license.

Access requests are reviewed manually. Hugging Face supplies the requester's username and account email by default; the additional form collects full name, country, affiliation, position, intended use, and a personal or institutional homepage. Supplying information does not guarantee approval, and granted access may be withdrawn if the dataset terms are violated.

For questions and release updates, use the CineDance GitHub repository or the project page.

Citation

If you use CineDance-1M or CineDance_01 in your research, please cite:

@misc{chen2026cinedancenextgenerationmultishotlongform,
  title        = {CineDance: Towards Next-Generation Multi-Shot Long-Form Cinematic Audio-Video Generation},
  author       = {Yuheng Chen and Teng Hu and Yuji Wang and Qingdong He and Zhucun Xue and Qianyu Zhou and Xiangtai Li and Lizhuang Ma and Jiangning Zhang and Dacheng Tao},
  year         = {2026},
  eprint       = {2606.09639},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2606.09639}
}

Downloads last month: 79

Total file size:

5.83 TB

Paper for CineDance/CineDance_01

CineDance: Towards Next-Generation Multi-Shot Long-Form Cinematic Audio-Video Generation

Paper • 2606.09639 • Published Jun 11