Please review the MM-Dia EULA, submit this access request, and email the signed EULA to jinzeyu23@mails.tsinghua.edu.cn. Access will be granted after manual review.

MM-Dia

MM-Dia is a dataset for expressive multimodal dialogue generation introduced in the ICLR 2026 paper “From Natural Alignment to Conditional Controllability in Multimodal Dialogue”. Curated from movies and TV series, it is the first dataset centered on dialogue-level expressiveness across modalities, with synchronized text, speech, visual context, and hierarchical style annotations.

Comparison of MM-Dia with existing dialogue-related datasets

In our paper, we explore three representative applications of MM-Dia: 1. style-controllable dialogue speech synthesis, 2. vision-conditioned dialogue speech synthesis, and 3. speech-driven dialogue video generation. The dataset may also support broader research in multimodal dialogue modeling, conversational speech, affective computing, and audio-visual learning.

Overview of MM-Dia and the representative tasks explored in the paper

Relationship and interaction distribution in MM-Dia

Dataset at a Glance

Statistic	MM-Dia
Dialogues	54,700
Speaking turns	449,138
Total duration	360.26 hours
Average speakers per dialogue	2.29
Average duration per dialogue	23.71 seconds
Average turns per dialogue	8.21
Speaker visibility	Partial

Released Content

Due to copyright restrictions on the source movies and TV series, the MM-Dia release contains:

JSON annotations: transcripts, timestamps, speaker information, style descriptions, Affective Triplets, expressiveness scores, and visual alignment metadata.
DAC-encoded audio: dialogue audio encoded at 44.1 kHz for local reconstruction as WAV.
Reconstruction tools: code for aligning and extracting dialogue video clips from legally obtained source media.

Original movie and TV video files are not distributed with MM-Dia. Researchers requiring visual clips must obtain legal copies of the corresponding source media and reconstruct the clips locally. The provided pipeline can also be used to extract and annotate dialogue clips from users' own new video and subtitle sources.

File Structure

MM-Dia/
├── json/
│   └── <batch>/
│       └── <movie_or_episode>.json
├── audio_enc/
│   └── <batch>/
│       └── <movie_or_episode>/
│           └── <clip>/
│               ├── <clip>.dac             # Denoised audio encoded at 24 kHz
│               ├── <clip>_44k.dac         # Denoised audio encoded at 44.1 kHz
│               ├── <clip>_no_vocals.dac   # Encoded non-vocal track with background music and ambient sounds
│               └── <clip>_orig.dac        # Encoded original audio with volume normalization
└── character/
    ├── movie_character/
    └── TV_character/
        └── <series_name>/
            └── <character_name>.<jpg|png>

After local reconstruction, decoded audio and extracted video follow the same hierarchy:

audio/<batch>/<movie_or_episode>/<clip>/*.wav
video/<batch>/<movie_or_episode>/<clip>.mp4

Corresponding JSON, audio, and reconstructed video entries refer to the same dialogue segment.

Character Image Gallery

The character/ directory is a reference image gallery for recurring movie and TV characters. It supports character recognition and speaker attribution in the MM-Dia curation pipeline. Movie entries are organized by IMDb ID and include a cast index, while TV entries are grouped by series.

Annotations

MM-Dia provides utterance-level and dialogue-level information, including:

utterance text, timestamps, and speaker identity;
dialogue style descriptions;
Affective Triplets: relationship, interaction mode, and emotional tone;
emotion intensity and emotion-flow volatility;
speaker visibility and visual alignment metadata.

Representative JSON fields include utterances, affective_triplet, description, emotion_intensity, emotion_volatility, and speaker_visibility.

What to Download

Research goal	JSON	Encoded audio	Source media required	Additional preparation
Conversational speech modeling	✓	✓	—	Decode DAC audio to WAV
Speech and affect understanding	✓	✓	—	Decode DAC audio to WAV
Audio-visual dialogue modeling	✓	✓	✓	Decode audio and reconstruct video
Speech-driven dialogue video generation	✓	✓	✓	Decode audio and reconstruct video

For direct benchmark evaluation with distributed WAV and video files, see MM-Dia-Bench.

Access and EULA

MM-Dia is gated and uses manual approval. It is available for non-commercial research and educational use under the MM-Dia EULA.

To request access:

Read and sign MM-DIA_EULA.pdf.
Submit the access request on this Hugging Face page using the same identity and email address.
Email the signed EULA to jinzeyu23@mails.tsinghua.edu.cn.
Access will be granted after the submitted information and signed agreement are reviewed.

Please do not redistribute the dataset or use it for commercial purposes. Users are responsible for obtaining lawful access to any source media used for video reconstruction.

Download

After approval, authenticate with Hugging Face and download the repository:

hf auth login
hf download jessyjin/MM-Dia \
  --repo-type dataset \
  --local-dir MM-Dia

The processing and reconstruction tools are maintained in the GitHub repository:

git clone https://github.com/jessyjinzy/MM-Dia.git
cd MM-Dia
pip install -r requirements.txt

Extract the Packaged Audio

The encoded audio is distributed as a Zstandard-compressed tar archive split into multiple parts. Install zstd, then run the following command from the downloaded dataset directory:

tar -xf MM_DIA-json.tar
tar -xf MM_DIA-character.tar

cat MM_DIA-audio_enc.tar.zst.part-* \
  | zstd -d -c \
  | tar -xf -

This streams the ordered parts directly into tar and extracts the audio_enc/ directory without creating an additional full-size archive. Keep all parts in the same directory and ensure that the destination has sufficient free space before extraction.

Decode the Audio

python utils/audio_codec_utils.py decode \
  --batch_input_dir /path/to/MM-Dia/audio_enc/movie \
  --batch_output_dir /path/to/output/audio/movie

The decoder supports batch processing and produces WAV files for training, evaluation, or listening.

Reconstruct Video

Original video is not included. After obtaining legal source media, reconstruct aligned dialogue clips with:

python av_align_and_extract.py \
  --batch_json_dir /path/to/MM-Dia/json/movie \
  --batch_audio_dir /path/to/output/audio/movie \
  --source_dir /path/to/legal/source/media \
  --batch_output_dir /path/to/output/video/movie

See the GitHub documentation for setup details and troubleshooting.

Limitations and Responsible Use

Copyright in the underlying movies and TV series remains with their respective rights holders. We do not claim ownership of the source media. If you are a rights holder and believe that any released material should be removed, please contact jinzeyu23@mails.tsinghua.edu.cn with the relevant details. We will review the request and remove the affected content where appropriate.

The dataset must not be used for identity impersonation, deceptive media generation, surveillance, harassment, or other harmful applications.

Citation

@inproceedings{jin2026mmdia,
  title     = {From Natural Alignment to Conditional Controllability in Multimodal Dialogue},
  author    = {Jin, Zeyu and Zhou, Songtao and Wang, Haoyu and Tian, Minghao and Yun, Kaifeng and Chen, Zhuo and Qin, Xiaoyu and Jia, Jia},
  booktitle = {The International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

Downloads last month: 27

Total file size:

8.11 GB