SocialDirector: Training-Free Social Interaction Control for Multi-Person Video Generation
Paper • 2605.10079 • Published
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is available for non-commercial academic research only. The video content remains the property of its original copyright holders and source datasets (MELD, Ego4D, SocialGesture / YouTube). By requesting access you confirm that you will use the dataset solely for non-commercial academic research and will comply with the terms of the original sources.
Log in or Sign Up to review the conditions and access this dataset content.
SocialDirector Dataset
Evaluation dataset of SocialDirector: Training-Free Social Interaction Control for Multi-Person Video Generation — 149 five-second multi-person clips with structured social-interaction annotations: who does what, when, and toward whom.
Contents
├── annotations.json # 149 records: prompt, boxes, text spans, timestamped events
├── first_frames/ # 149 reference images (generation input), {dataset}_{id:04d}.jpg
└── videos/ # 149 ground-truth 5s clips, {dataset}_{id:04d}.mp4
| Source | Clips | Global id range |
Domain |
|---|---|---|---|
| MELD | 19 | 0–18 | TV-series multi-party conversations |
| MMSI | 50 | 19–68 | Real-world social interactions (Ego4D & YouTube) |
| SocialGesture | 80 | 69–148 | Social game recordings |
Statistics: 674 annotated persons (1–8 per clip, mean 4.5), 479 timestamped events (299 with an explicit interaction target).
Annotation format
Each record in annotations.json:
| Field | Description |
|---|---|
id |
Global sample id (0–148) |
dataset |
Source dataset (meld / mmsi / socialgesture) |
video |
Relative path to the 5s ground-truth clip |
first_frame |
Relative path to the reference image (first frame) |
prompt |
Structured generation prompt describing the scene and all events |
n_speakers |
Number of annotated persons |
speaker_boxes |
[x1, y1, x2, y2] boxes, normalized to [0, 1], indexed by speaker id (mostly left-to-right; the prompt gives each speaker's position naming) |
speaker_text_ids |
Per-person [char_start, char_end) span of their action sentence(s) in prompt |
speaker_scene_ids |
Per-person [char_start, char_end) span of their mention in the scene description |
events |
List of {start, end, speaker, event, target} — timestamped social events with actor and interaction target |
interactions |
Directed actor→target pairs in index form with normalized time windows |
Intended use
Evaluation dataset for controllable multi-person social interaction video generation:
generate a 5s video from first_frame + prompt (+ speaker_boxes / events
for layout-conditioned methods), then evaluate whether each person performs the
specified action within the specified time window toward the specified target.
The videos/ clips are the real ground-truth references (e.g., for FVD / LPIPS).
License
This dataset is released for non-commercial academic research use only.
- Annotations (
annotations.json), created by the authors, are provided for academic research use only. - Videos and first frames are curated from MELD, Ego4D, and SocialGesture / YouTube recordings. The video content remains the property of its original copyright holders and source datasets; we claim no rights over it. By using this dataset you agree to comply with the terms and licenses of the original sources (including the MELD, Ego4D, and SocialGesture license agreements) in addition to the terms above.
If you are a copyright holder and believe any content should be removed, please contact us (oyly@iis.u-tokyo.ac.jp) and we will promptly take it down.
Citation
@article{ouyang2026socialdirector,
title = {SocialDirector: Training-Free Social Interaction Control for Multi-Person Video Generation},
author = {Ouyang, Liangyang and Liu, Ruicong and Kang, Caixin and Huang, Yifei and Sato, Yoichi},
journal = {arXiv preprint arXiv:2605.10079},
year = {2026}
}
- Downloads last month
- 20