Donfa1con/R-DGG

Reference-Based Directed Granger Gain for Evaluating Speech-Conditioned Listener Motion

Evaluation dataset for the paper. The evaluation code is in the GitHub repository.

What this is

This dataset holds the material to score dyadic listener generation on the SI-184 benchmark. SI-184 is a set of 184 dyadic conversation clips. Seven talking-head systems each generate a listener video for every clip. This dataset gives you those generated videos. It also gives you the motion features and the audio features that the metrics read.

Two motion encoders run on every video: EMOCA (FLAME expression and pose codes) and LivePortrait (implicit keypoint motion). Two audio features come from the ground-truth side: HuBERT speech features and VAD voice-activity. With these files you compute the R-DGG score and the classic dyadic metrics. You do not need to generate the videos again. You do not need to run the feature extractors again.

The 184 clip identifiers are listed in pairs184.txt. Column 1 gives the listener stem. Column 2 gives the partner (speaker) stem.

Systems

The dataset covers eight sources. GT is the ground truth. It carries features only, not video. The other seven are the generated systems. Each generated system carries video plus features.

source	role	video	features
GT	ground truth	no	emoca, liveportrait, hubert, vad
AVTR-1	generated (dyadic)	yes	emoca, liveportrait
AvatarForcing	generated (dyadic)	yes	emoca, liveportrait
dystream	generated (dyadic)	yes	emoca, liveportrait
FLOAT	generated (lip-sync)	yes	emoca, liveportrait
ditto	generated (lip-sync)	yes	emoca, liveportrait
SoulX_FlashHead_Lite	generated (lip-sync)	yes	emoca, liveportrait
SoulX_FlashHead_Pro	generated (lip-sync)	yes	emoca, liveportrait

Layout

pairs184.txt                                   # 184 lines: <listener>.wav <speaker>.wav
README.md                                      # this card
GT/
  LS_AL_emoca/<stem>.npz                        # 184  EMOCA codes
  LS_AL_liveportrait/<stem>.npz                 # 184  LivePortrait motion
  LS_AL_hubert/<stem>.npz                       # 184  HuBERT audio features
  LS_AL_vad/<stem>.npz                          # 184  voice-activity
<model>/                                        # each of the 7 generated systems
  LS_AL/<stem>.mp4                              # 184  generated listener video (25 fps)
  LS_AL_emoca/<stem>.npz                        # 184  EMOCA codes
  LS_AL_liveportrait/<stem>.npz                 # 184  LivePortrait motion

Every directory holds exactly 184 files. Every stem matches column 1 of pairs184.txt.

Licenses

Use this dataset for non-commercial purposes only.

The dataset as a whole is licensed CC BY-NC 4.0.

The GT/* features are derived from the Seamless Interaction dataset (Meta / FAIR). They are attributed to Seamless Interaction and are covered by CC BY-NC 4.0.

Each generated system's videos stay under that system's own upstream license. The list below states each license as declared by its upstream project:

source	upstream license of the generated videos
ditto	Apache License 2.0
FLOAT	CC BY-NC 4.0 (Attribution-NonCommercial 4.0 International)
SoulX_FlashHead_Lite	Apache License 2.0
SoulX_FlashHead_Pro	Apache License 2.0
AvatarForcing	CC BY-NC 4.0 (Attribution-NonCommercial 4.0 International)
dystream	No license file is declared by the upstream project. Treat the videos as research use only. All rights stay with the original authors.
AVTR-1	Non-commercial research only. The model weights are under a non-commercial community license. The inference and renderer code are under the PolyForm Noncommercial License 1.0.0.

The strictest term applies to any part you reuse. Several systems are non-commercial or research-only, so the combined dataset is non-commercial.

Not included

The dataset does not include the ground-truth raw video or the ground-truth audio. Those come from the Seamless Interaction release, and you must accept its access terms to obtain them.

Fetch the ground-truth video and audio with the code repository script data/download_si184.py. That script writes:

GT/LS_AL/<stem>.mp4        # ground-truth video (both dyad roles)
GT/LS_AL_wav/<stem>.wav    # driving / partner audio (both dyad roles)

The reference portraits (data/references/<stem>.png) ship with the code repository, not with this dataset.

Usage

1. Get the R-DGG code repository.

2. Copy this whole tree into the repository at data/. Merge it with the existing data/ content (the pairs184.txt, references/, and the download_si184.py script).

3. Run data/download_si184.py to add the ground-truth video and audio (see "Not included").

4. Compute the R-DGG score:

   bash pipeline/4_rdgg.sh
   

5. Compute the classic dyadic metrics:

   bash pipeline/5_classic.sh
   

Both scripts read the features and videos in this tree. Neither script re-generates the videos. Neither script re-extracts the features.

Citation

@inproceedings{kravtsov2026referencebased,
  title={Reference-Based Directed Granger Gain for Evaluating Speech-Conditioned Listener Motion},
  author={Artem Kravtsov and Dmitrii Ziganshin and Vsevolod Poletaev and Anastasia Tikhonova},
  booktitle={NeurIPS 2026 Workshop Real-Time Conversational Agents: Toward Natural Multimodal Interaction},
  year={2026},
  url={https://openreview.net/forum?id=uZdMPMmHxa}
}