Pith. sign in

REVIEW 3 major objections 5 minor 48 references

REACT 2025: the Third Multiple Appropriate Facial Reaction Generation Challenge

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MARS is claimed to be the first dataset collected specifically for multiple appropriate facial reaction generation, with baseline models now benchmarked on it.

desk verdict Genuinely new MARS dataset earns the challenge a serious look, but 'appropriate' is defined by co-occurrence, which is a validity gap that needs an expert-rating check. read the letter →

arxiv 2505.17223 v1 pith:HJHUOH75 submitted 2025-05-22 cs.CV

Siyang Song , Micol Spitale , Xiangyu Kong , Hengde Zhu , Cheng Luo , Cristina Palmero , German Barquero , Sergio Escalera
show 7 more authors
Michel Valstar Mohamed Daoudi Tobias Baur Fabien Ringeval Andrew Howes Elisabeth Andre Hatice Gunes
This is my paper · ORCID
classification cs.CV
keywords facialreactiongenerationmultipleappropriatereactionsdyadicinteractionMARSdatasetdiffusionmodelsbenchmarkchallengeactionunitsaffectivecomputing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces the REACT 2025 challenge, whose core offering is the MARS dataset, described as the first multimodal dataset collected specifically for multiple appropriate facial reaction generation (MAFRG) in natural human-human interactions. The dataset contains 137 dyadic interactions split into 2856 sessions across five topics, with audio, video, EEG, transcripts, and frame-level facial descriptors. The paper claims that every facial reaction expressed by any listener in a session is an appropriate reaction to each speaker behaviour in that session, which turns co-occurrence into the ground-truth label. On this basis it reports baselines for offline and online generation, with the diffusion-based PerFRDiff reaching FRCorr of 0.58 offline and 0.55 online, establishing initial benchmarks. If the labelling premise holds, the field gains a common evaluation standard and a purpose-built training resource.

What carries the argument

The load-bearing machinery is the session-consistency labelling rule: within each of the 23 pre-defined sessions, all facial reactions expressed by different listeners under the same semantic context are treated as the set of appropriate facial reactions to each speaker behaviour expressed in that session. This rule converts co-occurrence into ground truth and defines the one-to-many target that MAFRG models must learn. A secondary mechanism is the Transformer-based variational encoder-decoder used for evaluation, which encodes real AFRs of arbitrary length into fixed-length token sequences so that variable-length generated reactions can be scored against all ground-truth reactions. The baselines then instantiate the one-to-many mapping by learning a distribution (Trans-VAE and PerFRDiff) or a Gaussian mixture graph distribution (REGNN) over plausible reactions.

What would settle it

Take a random set of listener-reaction clips labelled as appropriate in MARS and have independent human raters judge whether each reaction is a contextually appropriate response to the paired speaker clip; if agreement with the dataset labels is low, the ground-truth definition fails and all metric scores built on it would need re-interpretation.

Watch

Extended reading notes

Core claim

The central discovery is that a specifically collected dyadic corpus can turn the one-to-many mapping of facial reactions into a trainable task: MARS pairs 23 speakers with 137 listeners in controlled video-conference conversations, and labels every listener reaction to a given speaker behaviour within a session as an appropriate facial reaction. The paper argues this makes the dataset the first to offer natural human-human interaction scenarios and rigorous AFR labels for MAFRG, in contrast to earlier challenges that had to repurpose clips recorded for other goals. It further shows that variable-length speaker behaviours and variable-length appropriate reactions can be handled by a Transformer-based variational encoder-decoder that embeds arbitrary-length reactions into fixed-length token sequences, allowing the FRDist and FRCorr metrics to compare generated reactions against all real reactions of varied lengths. On this setup the baselines, especially the diffusion model PerFRDiff and the graph-based REGNN, outperform random and mean-prediction baselines, with PerFRDiff achieving the highest correlation (0.58 offline, 0.55 online) and the best diversity among the compared methods.

Load-bearing premise

The dataset labels treat every facial reaction that any listener happened to express during a session as an appropriate response to every speaker behaviour in that session, with no expert judgment that the reaction is actually appropriate.

Editorial extensions

If this is right

  • If the MARS labels are accepted, the REACT 2025 ranking becomes a common benchmark for appropriateness, diversity, realism, and synchrony in facial reaction generation.
  • Future models trained on MARS can be evaluated against a corpus collected specifically for the task, rather than repurposed from other datasets.
  • The reported baseline results imply that diffusion-based generation (PerFRDiff) currently offers the best correlation-diversity trade-off in both offline and online settings.
  • The challenge's variable-length setup means that winning systems must handle uneven speaker behaviour durations, an ability previous REACT challenges did not require.
  • The inclusion of EEG, personality traits, and 3DMM coefficients in MARS enables research on personalised and multimodal reaction generation beyond the challenge itself.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If co-occurrence is treated as appropriateness, evaluation metrics inherit that assumption; a human-perceptual validation study of sampled listener reactions would test whether the label rule measures what the challenge claims.
  • The B_Mime baseline's strong FRCorr (0.52 offline) suggests that simple speaker-mimicry may account for a large share of correlated facial behaviour, implying that 'appropriateness' as measured here may be confounded with facial mimicry.
  • The MARS label rule could potentially be extended to other modalities in the dataset (e.g., EEG or audio responses), offering a way to study multimodal listener entrainment without new annotation campaigns.
  • Because the same session label set is used for both training and evaluation, models that memorise session-level priors rather than speaker behaviour could inflate FRCorr; the challenge's held-out test split would be the place to check for this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces the REACT 2025 challenge on Multiple Appropriate Facial Reaction Generation (MAFRG), together with the MARS dataset, claimed to be the first specifically collected naturalistic human-human dyadic interaction dataset for MAFRG with rigorous ground-truth appropriateness labels. The challenge defines offline and online MAFRG sub-tasks, describes a set of evaluation metrics (FRCorr, FRDist, FRDiv, FRVar, FRRea, FRSyn), and reports baseline results for Trans-VAE, PerFRDiff, and REGNN against random, mean, and mimicry baselines on the test set. The authors state that learned baselines outperform the random and mean baselines and that PerFRDiff achieves the best FRCorr, with the dataset and baseline code made publicly available.

Significance. If the ground-truth labels are valid, MARS is a substantial contribution: it is larger and more naturalistic than previous dyadic reaction datasets, includes audio, visual, EEG, personality traits, and multiple annotation streams, and provides a common benchmark for a one-to-many generative task. The release of working baseline implementations (Trans-VAE, PerFRDiff, REGNN) and an evaluation protocol is a practical asset for the community. The paper also extends prior REACT challenges to variable-length speaker inputs, which better reflects real dyadic interaction. However, the significance is conditional on the validity of the appropriateness labels and on the evaluation protocol being well-defined and reproducible.

major comments (3)
  1. [Section 3, 'Ground-truth labels'] The definition of ground-truth appropriate facial reactions as 'all facial reactions expressed by different listeners under the same session' is a co-occurrence criterion, not an independently validated measure of appropriateness. No expert annotation, listener self-report, or inter-rater reliability analysis is provided. Since the same labels are used both to train baselines and to compute FRCorr and FRDist in Section 4, the evaluation is circular with respect to the paper's central claim that the benchmark measures appropriateness. The claim that MARS contains 'rigorous AFR labels' is therefore unsupported as stated. I ask the authors to either provide an external validation of the co-occurrence proxy (e.g., expert ratings of a sample of listener reactions, or evidence of cross-listener agreement beyond mere co-occurrence) or explicitly reframe the labels as consensus-based typical reactions rather than rigorously validated appropriate reactions.
  2. [Table 3, Section 6] The baseline results are reported without error bars, confidence intervals, or significance tests. This is consequential because the B_Mime baseline, which simply copies the speaker's facial behaviour, reaches FRCorr 0.52 compared to the best learned model's 0.58 offline, and B_Mime actually achieves a better FRDist than PerFRDiff (206.02 vs 217.32). The text acknowledges mimicry but still concludes that the learned models 'largely outperformed' the baselines. Without repeated runs or statistical comparisons, the claim that the learned models are meaningfully better than a trivial mimicry baseline, and that FRCorr isolates appropriateness rather than mimicry, is not established. Please report variance across training seeds and perform significance tests, or temper the claim accordingly.
  3. [Section 4, 'Evaluation metrics'] The evaluation protocol has two underspecified free parameters that affect all reported metrics. First, the number of generated reactions M in Tasks 1 and 2 is not fixed in the table; the 'GT 10' row suggests M=10, but this is never stated explicitly and no sensitivity analysis is provided. Second, FRCorr and FRDist are computed on fixed-length token sequences produced by a learned Transformer VAE tokenizer, yet the paper does not specify how the tokenizer was trained (e.g., on which split, with which labels), how token length is chosen, or how the metric is aggregated over tokens and over the M generated reactions. If the tokenizer is trained on the same co-occurrence labels used for the baselines, the evaluation is doubly circular. Please make the tokenizer and the metric computation fully concrete, and release the evaluation code with the baseline code.
minor comments (5)
  1. [Section 3, dataset description] The counts are inconsistent: the text says the dataset comprises 137 dyadic interaction clips but also reports 270 multi-modal recordings and 23 speakers with 137 listeners. Clarify the relationship between clips, recordings, and session pairs (2856 session pairs vs 5712 sessions).
  2. [Section 2, task definitions] There is a typo in the sentence 'The formal definition them is provided by [11]'; it should read 'The formal definition of them is provided by [11].'
  3. [Table 3] The table formatting is difficult to read: the row 'GT 10 0.00 0.1876 0.0669 – 48.66' appears to mix the value M=10 with the metric values, and the '10' is not explained in the caption or text.
  4. [Section 7] The sentence 'Our evaluation protocol strictly ranked all participant models under the same settings by evaluating two aspects' is followed by a list of four aspects (appropriateness, diversity, realism, synchrony); 'two' should be 'four'.
  5. [Section 1] The claim that MARS is 'the first specifically collected MAFRG dataset' should be qualified with 'to our knowledge' and ideally cite or discuss prior dyadic interaction datasets that were not purpose-built, to avoid overclaiming.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'rigorous AFR labels' claim and the FRCorr/FRDist benchmark reduce by construction to a co-occurrence definition: appropriateness is whatever any listener did in the same session, and the same pseudo-labels form the evaluation ground truth.

  1. self definitional [Section 3, Ground-truth labels]
    "for each speaker behaviour expressed in a specific session, we define all facial reactions expressed by different listeners under the same session to be appropriate facial reactions (i.e., ground-truth) for responding to it."

    The paper's central contribution is a dataset with 'rigorous AFR labels' and baselines evaluated for appropriateness via FRCorr/FRDist. But the ground-truth AFR label is defined as the set of all listener reactions co-occurring in the same session, with no expert annotation, listener self-report, or independent validation. Therefore the label is the definition, not a measured property: 'appropriate' means 'was expressed by some listener in that session'. Any evaluation against these labels (Sec. 4, Sec. 6) is a consistency check against the same co-occurrence set, so the benchmark cannot confirm that the labels capture appropriateness beyond the stipulation.

  2. other [Section 4, Evaluation Metrics]
    "we trained a Transformer-based variational encoder-decoder framework to summarise every arbitrary-length real AFRs into a fixed-length sequence of tokens, allowing to represent all variable-length real facial reactions with fixed-length token sequences in a latent space. This way, each generated AFR can be directly compared with all corresponding real AFRs of varied lengths in this latent space for computing FRDist and FRCorr metrics."

    The latent space in which FRDist and FRCorr are computed is itself trained on the same 'real AFRs' that were defined in Sec. 3 as all co-occurring listener reactions. The metric therefore scores generated reactions in a representational space optimised to reconstruct the very pseudo-labels it evaluates. This adds a second closed loop: the evaluation geometry is fit to the label set, so high FRCorr/FRDist scores partly reflect proximity to the training-label distribution rather than an externally verified notion of appropriateness.

full rationale

The circularity is concentrated in the label definition, not in the data collection or the baseline code, which are real and reproducible. Section 3 equates appropriateness with co-occurrence: every listener facial reaction in a session is declared an appropriate reaction to every speaker behaviour in that session. This definitional move directly feeds the evaluation metrics in Section 4 (FRCorr, FRDist) because those metrics compare generated reactions to the same 'GT real AFRs'. The learned tokenizer used for those metrics is also trained on the same GT reactions, further closing the loop. The paper's headline novelty, 'rigorous AFR labels', therefore has no independent empirical content: the labels are rigorous only relative to the stipulation that co-occurrence is appropriateness. This is not a case of a fitted parameter being renamed a prediction, nor of a load-bearing self-citation chain; the baselines are compared fairly and the dataset is a genuine new resource. However, the central evaluation claim does reduce by construction to the co-occurrence pseudo-label, so the score is 6 rather than 0-4. An expert-rating validation of a subset of reactions would break the circle.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central benchmark rests on the dataset's label definition (co-occurrence equals appropriateness), the comparability of the 23-session protocol, and the validity of adopted metrics and feature extractors. No free parameters are fit to derive a scientific constant; the main unspecified quantities are the number of generated reactions M and internal baseline hyperparameters, which are not disclosed.

free parameters (2)
  • Number of generated reactions M
    M is the count of facial reactions generated per speaker behaviour; it affects diversity scores but is not reported for the baselines in Table 3.
  • Evaluation latent tokenizer length
    The Transformer tokenizer used to compute FRCorr and FRDist has an undisclosed sequence length, which affects distance calculations and is chosen by hand.
assumptions (4)
  • domain assumption Every facial reaction expressed by any listener during a session is an appropriate reaction to each speaker behaviour in that session.
    Section 3, Ground-truth labels: the GT set is defined by co-occurrence across listeners without expert annotation.
  • domain assumption The fixed 23-session protocol gives comparable semantic context across different dyads.
    Section 3: cross-listener comparison as GT requires that session context is shared and stable.
  • domain assumption The metrics from [11,34] measure appropriateness, diversity, realism and synchrony as intended.
    Section 4 adopts FRCorr, FRDist, FRVar, FRDiv, FRRea, FRSyn without validating them on MARS.
  • domain assumption The automatic feature extractors (AU model, EmoNet, Faceverse, wav2vec 2.0) provide valid behaviour descriptors.
    Section 3 Provided data: all downstream labels and inputs rely on these pretrained tools.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REACT 2025: the Third Multiple Appropriate Facial Reaction Generation Challenge." pith.science (2026). https://pith.science/paper/HJHUOH75

@misc{pith2026250517223,
  author       = {Pith},
  title        = {Pith review of: REACT 2025: the Third Multiple Appropriate Facial Reaction Generation Challenge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJHUOH75}},
  note         = {Machine review of arXiv:2505.17223}
}
read the original abstract

In dyadic interactions, a broad spectrum of human facial reactions might be appropriate for responding to each human speaker behaviour. Following the successful organisation of the REACT 2023 and REACT 2024 challenges, we are proposing the REACT 2025 challenge encouraging the development and benchmarking of Machine Learning (ML) models that can be used to generate multiple appropriate, diverse, realistic and synchronised human-style facial reactions expressed by human listeners in response to an input stimulus (i.e., audio-visual behaviours expressed by their corresponding speakers). As a key of the challenge, we provide challenge participants with the first natural and large-scale multi-modal MAFRG dataset (called MARS) recording 137 human-human dyadic interactions containing a total of 2856 interaction sessions covering five different topics. In addition, this paper also presents the challenge guidelines and the performance of our baselines on the two proposed sub-challenges: Offline MAFRG and Online MAFRG, respectively. The challenge baseline code is publicly available at https://github.com/reactmultimodalchallenge/baseline_react2025

Figures

Figures reproduced from arXiv: 2505.17223 by the authors.

Figure 1
Figure 1. Illustration of the data collection scenario of the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Statistics of participants’ ethnic groups. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the Trans-VAE baseline [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Overview of the PerFRDiff baseline. 5 BASELINE MODELS To enable fair and easy comparison, we first re-employ an open￾source Trans-VAE designed for both online and offline MAFRG tasks in previous challenges [34, 35]. We also provide the generic MAFRG model of the state-…
Figure 5
Figure 5. Figure 5: Overview of the REGNN baseline. facial attribute time-series and raw 3DMM coefficients sequence, to a shared latent space, serving as conditional inputs to the diffu￾sion Transformer. Then, the model follows the standard denoising diffusion process to gradually denoise…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 38 canonical work pages

  1. [1]

    Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. 2022. Teach: Temporal action composition for 3d humans. In2022 International Conference on 3D Vision (3DV). IEEE, 414–423

    work page 2022

  2. [2]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Conference’25, , Song et al. Advances in neural information processing systems33 (2020), 12449–12460

    work page 2020

  3. [3]

    Quang Tien Dam, Tri Tung Nguyen Nguyen, Dinh Tuan Tran, and Joo-Ho Lee. 2024. Finite Scalar Quantization as Facial Tokenizer for Dyadic Reaction Generation. In2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG). IEEE, 1–5

    work page 2024

  4. [4]

    Kaijun Deng, Dezhi Zheng, Jindong Xie, Jinbao Wang, Weicheng Xie, Linlin Shen, and Siyang Song. 2025. DEGSTalk: Decomposed Per-Embedding Gaussian Fields for Hair-Preserving Talking Face Synthesis. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    work page 2025

  5. [5]

    Ulf Dimberg. 1982. Facial reactions to facial expressions.Psychophysiology19, 6 (1982), 643–647

    work page 1982

  6. [6]

    Cafaro et al. 2017. The NoXi database: multimodal recordings of mediated novice-expert interactions. InICMI 2017

    work page 2017

  7. [7]

    Evonne Ng et al. 2022. Learning to listen: Modeling non-deterministic dyadic facial motion. InIEEE/CVF CVPR 2022. 20395–20405

    work page 2022

  8. [8]

    Fabien Ringeval et al. 2013. Introducing the RECOLA multimodal corpus of remote collaborative and affective interactions. InIEEE FG 2013

    work page 2013

Show all 48 references
  1. [9]

    Palmero et al. 2021. Context-aware personality inference in dyadic scenarios: Introducing the udiva dataset. InIEEE/CVF 2021

    2021

  2. [10]

    Song et al. 2022. Learning Person-specific Cognition from Facial Reactions for Automatic Personality Recognition.IEEE Transactions on Affective Computing (2022)

    2022

  3. [11]

    Song et al. 2023. Multiple Appropriate Facial Reaction Generation in Dyadic In- teraction Settings: What, Why and How?https://arxiv.org/abs/2302.06514(2023)

    2023 arXiv

  4. [12]

    Yuchi et al. 2017. Dyadgan: Generating facial expressions in dyadic interactions. InIEEE CVPR Workshops 2017. 11–18

    2017

  5. [13]

    Ursula Hess, Pierre Phillippot, and Sylvie Blairy. 1998. Facial reactions to emo- tional facial expressions: Affect or cognition?Cognition & Emotion12, 4 (1998), 509–531

    1998

  6. [14]

    Ximi Hoque, Adamay Mann, Gulshan Sharma, and Abhinav Dhall. 2023. BEAMER: Behavioral Encoder to Generate Multiple Appropriate Facial Reactions. InProceedings of the ACM International Conference on Multimedia. 9536–9540

    2023

  7. [15]

    Guanyu Hu, Jie Wei, Siyang Song, Dimitrios Kollias, Xinyu Yang, Zhonglin Sun, and Odysseus Kaloidas. 2024. Robust Facial Reactions Generation: An Emotion- Aware Framework with Modality Compensation. In2024 IEEE International Joint Conference on Biometrics (IJCB). IEEE, 1–10

    2024

  8. [16]

    Yuchi Huang and Saad M Khan. 2018. Generating Photorealistic Facial Expres- sions in Dyadic Interactions. InBMVC. 201

    2018

  9. [17]

    Yuchi Huang and Saad M Khan. 2018. A generative approach for dynamically varying photorealistic facial expressions in human-agent interactions. InPro- ceedings of the 20th ACM International Conference on Multimodal Interaction. 437–445

    2018

  10. [18]

    Cong Liang, Jiahe Wang, Haofan Zhang, Bing Tang, Junshan Huang, Shangfei Wang, and Xiaoping Chen. 2023. UniFaRN: Unified Transformer for Facial Reaction Generation. InProceedings of the ACM International Conference on Multimedia. 9506–9510

    2023

  11. [19]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

    2024 arXiv

  12. [20]

    Zhenjie Liu, Cong Liang, Jiahe Wang, Haofan Zhang, Yadong Liu, Caichao Zhang, Jialin Gui, and Shangfei Wang. 2024. One-to-Many Appropriate Reaction Map- ping Modeling with Discrete Latent Variable. In2024 IEEE 18th International Conference on Automatic Face and Gesture Recogni...

    2024

  13. [21]

    Cheng Luo and et al. 2022. Learning Multi-dimensional Edge Feature- based AU Relation Graph for Facial Action Unit Recognition.arXiv preprint arXiv:2205.01782(2022)

    2022 arXiv

  14. [22]

    Cheng Luo, Siyang Song, Weicheng Xie, Micol Spitale, Zongyuan Ge, Linlin Shen, and Hatice Gunes. 2024. ReactFace: Online Multiple Appropriate Facial Reaction Generation in Dyadic Interactions.IEEE Transactions on Visualization and Computer Graphics(2024)

    2024

  15. [23]

    Qincheng Lv, Xiaofeng Liu, Jie Li, Rongrong Ni, Pujun Xue, and Siyang Song. 2025. Hierarchical Multimodal Decoupling-Fusion Framework for offline Multiple Appropriate Facial Reaction Generation. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal P...

    2025

  16. [24]

    1974.An approach to environmental psychology.the MIT Press

    Albert Mehrabian and James A Russell. 1974.An approach to environmental psychology.the MIT Press

    1974

  17. [25]

    Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. 2022. Learning To Listen: Modeling Non-Deterministic Dyadic Facial Motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 20395–20405

    2022

  18. [26]

    Dang-Khanh Nguyen, Prabesh Paudel, Seung-Won Kim, Ji-Eun Shin, Soo-Hyung Kim, and Hyung-Jeong Yang. 2024. Multiple Facial Reaction Generation Using Gaussian Mixture of Models and Multimodal Bottleneck Transformer. In2024 IEEE 18th International Conference on Automatic Face and...

    2024

  19. [27]

    Minh-Duc Nguyen, Hyung-Jeong Yang, Ngoc-Huynh Ho, Soo-Hyung Kim, Se- ungwon Kim, and Ji-Eun Shin. 2024. Vector Quantized Diffusion Models for Multiple Appropriate Reactions Generation. In2024 IEEE 18th International Con- ference on Automatic Face and Gesture Recognition (FG). ...

    2024

  20. [28]

    Minh-Duc Nguyen, Hyung-Jeong Yang, Soo-Hyung Kim, Ji-Eun Shin, and Seung- Won Kim. 2025. Latent behaviour Diffusion for Sequential Reaction Generation in Dyadic Setting. InInternational Conference on Pattern Recognition. Springer, 233–248

    2025

  21. [29]

    Behnaz Nojavanasghari, Yuchi Huang, and Saad Khan. 2018. Interactive genera- tive adversarial networks for facial expression generation in dyadic interactions. arXiv preprint arXiv:1801.09092(2018)

    2018 arXiv

  22. [30]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al . 2018. Improving language understanding by generative pre-training. (2018)

    2018

  23. [31]

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. 2019. Habitat: A platform for embodied ai research. InProceedings of the IEEE/CVF international conference on computer vis...

    2019

  24. [32]

    Zilong Shao, Siyang Song, Shashank Jaiswal, Linlin Shen, Michel Valstar, and Hatice Gunes. 2021. Personality recognition by modelling person-specific cog- nitive processes using graph representation. Inproceedings of the 29th ACM international conference on multimedia. 357–366

    2021

  25. [33]

    Siyang Song, Yuxin Song, Cheng Luo, Zhiyuan Song, Selim Kuzucu, Xi Jia, Zhijiang Guo, Weicheng Xie, Linlin Shen, and Hatice Gunes. 2022. Gratis: Deep learning graph representation with task-specific topology and multi-dimensional edge features.arXiv preprint arXiv:2211.12482(2022)

    2022 arXiv

  26. [34]

    Siyang Song, Micol Spitale, Cheng Luo, Germán Barquero, Cristina Palmero, Sergio Escalera, Michel Valstar, Tobias Baur, Fabien Ringeval, Elisabeth André, et al. 2023. React2023: The first multiple appropriate facial reaction generation challenge. InProceedings of the 31st ACM ...

    2023

  27. [35]

    Siyang Song, Micol Spitale, Cheng Luo, Cristina Palmero, German Barquero, Hengde Zhu, Sergio Escalera, Michel Valstar, Tobias Baur, Fabien Ringeval, et al

  28. [36]

    Mingze Sun, Chao Xu, Xinyu Jiang, Yang Liu, Baigui Sun, and Ruqi Huang. 2025. Beyond talking–generating holistic 3d human dyadic motion for communication. International Journal of Computer Vision133, 5 (2025), 2910–2926

    2025

  29. [37]

    Antoine Toisoul and et al. 2021. Estimation of continuous valence and arousal levels from faces in naturalistic conditions.Nature Machine Intelligence3, 1 (2021), 42–50

    2021

  30. [38]

    Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. 2024. DIM: Dyadic Interaction Modeling for Social Behavior Generation. InEuropean Con- ference on Computer Vision. Springer, 484–503

    2024

  31. [39]

    Isaac Wang and Jaime Ruiz. 2021. Examining the use of nonverbal communication in virtual agents.International Journal of Human-Computer Interaction37, 17 (2021), 1648–1673

    2021

  32. [40]

    Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu

  33. [41]

    Tong Xu, Micol Spitale, Hao Tang, Lu Liu, Hatice Gunes, and Siyang Song. 2023. Reversible Graph Neural Network-based Reaction Distribution Learning for Mul- tiple Appropriate Facial Reactions Generation.arXiv preprint arXiv:2305.15270 (2023)

    2023 arXiv

  34. [42]

    Jun Yu, Ji Zhao, Guochen Xie, Fengxin Chen, Ye Yu, Liang Peng, Minglei Li, and Zonghong Dai. 2023. Leveraging the Latent Diffusion Models for Offline Facial Multiple Appropriate Reactions Generation. InProceedings of the ACM International Conference on Multimedia. 9561–9565

    2023

  35. [43]

    Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. 2021. Facial: Synthesizing dynamic talking face with implicit attribute learning. InProceedings of the IEEE/CVF international conference on computer vision. 3867–3876

    2021

  36. [44]

    Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. 2022. Responsive listening head generation: a benchmark dataset and baseline. In European conference on computer vision. Springer, 124–142

    2022

  37. [45]

    Hengde Zhu, Xiangyu Kong, Weicheng Xie, Xin Huang, Xilin He, Lu Liu, Linlin Shen, Wei Zhang, Hatice Gunes, and Siyang Song. 2025. PerReactor: Offline Personalised Multiple Appropriate Facial Reaction Generation. InProceedings of the AAAI Conference on Artificial Intelligence, ...

    2025

  38. [46]

    Hengde Zhu, Xiangyu Kong, Weicheng Xie, Xin Huang, Linlin Shen, Lu Liu, Hatice Gunes, and Siyang Song. 2024. Perfrdiff: Personalised weight editing for multiple appropriate facial reaction generation. InProceedings of the 32nd ACM International Conference on Multimedia. 9495–9504

    2024

  39. [2022]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20333–20342

  40. [2024]

    React 2024: the second multiple appropriate facial reaction generation challenge.arXiv preprint arXiv:2401.05166(2024)

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.