Pith. sign in

REVIEW 3 major objections 5 minor 48 references

A synthetic dataset of 3D-rendered pain faces with precisely controlled muscle activations can train automated pain assessment models to clinical-level accuracy, reaching 0.91 AUROC on real patient data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 16:04 UTC pith:RHFV3EQC

load-bearing objection A serious synthetic-data engineering effort whose 'clinically grounded' claim needs external AU validation and control augmentations before it convinces me. the 3 major comments →

arxiv 2509.16727 v5 pith:RHFV3EQC submitted 2025-09-20 cs.CV cs.LG

Pain in 3D: Generating Controllable Synthetic Faces for Automated Pain Assessment

Xin Lei Lin , Soroush Mehraban , Abhishek Moturu , Babak Taati This is my paper
classification cs.CV cs.LG
keywords synthetic dataautomated pain assessmentfacial action unitsPSPI3D face generationneural face riggingcross-modal knowledge distillationVision Transformer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Automated pain assessment relies on facial expressions, but real pain datasets are small, demographically narrow, and dominated by low-pain frames because high-intensity pain is ethically impossible to collect on demand. The authors set out to show that a fully synthetic dataset can break this bottleneck. Their 3DPain pipeline rigs diverse 3D face meshes with randomly sampled intensities of six pain-related action units, generating 82,500 frames with exact AU labels, PSPI clinical pain scores, and pain-region heatmaps. They then train a Vision Transformer (ViTPain) with this data as augmentation and with heatmap-based cross-modal distillation, reaching 0.91 AUROC on the UNBC-McMaster real pain dataset under subject-independent cross-validation—matching or exceeding prior state-of-the-art. If true, this means pain training data can be manufactured with controlled demographics and intensities, sidestepping ethical and practical limits on clinical data collection.

Core claim

3DPain is a synthetic dataset of 82,500 frames from 2,500 identities, generated by a three-stage pipeline: FLAME parametric meshes provide diverse facial structures; diffusion models (depth-conditioned Kandinsky 2.2 and Hunyuan3D 2.1) produce photorealistic neutral images and textures; and Neural Face Rigging (NFR) maps a vector of six pain-relevant action-unit (AU) intensities to mesh vertex displacements, yielding pain expressions whose AU activations and PSPI scores are known exactly and whose vertex displacements generate pain-region heatmaps. On the UNBC-McMaster real-pain benchmark, a Vision Transformer (ViTPain) with AU-query cross-attention and cross-modal distillation from a heatmap

What carries the argument

The key mechanism is AU-driven neural face rigging on FLAME meshes: a network (DiffusionNet + Neural Jacobian Fields) translates action-unit intensity values into 3D vertex displacements. Because the PSPI pain index is a fixed linear combination of six AUs (AU4, AU6, AU7, AU9, AU10, AU43), random sampling of these intensities yields expressions with known PSPI scores and known AU labels. Heatmaps are computed by subtracting neutral vs. rigged mesh vertex positions, providing a spatial, geometric teaching signal that is distilled into a 2D image model via cross-modal knowledge distillation.

Load-bearing premise

The load-bearing premise is that the AU intensities sampled for neural rigging produce facial expressions that are clinically valid pain expressions that transfer to real patient faces; the paper never tests the rendered images against human pain ratings, FACS annotation, or an AU detector.

What would settle it

Run a validated AU detector on the rendered 3DPain faces and compare its outputs to the rigged AU intensities; if the detector's intensities deviate beyond the tolerance used in clinical PSPI scoring for a non-negligible fraction of frames, the claim of clinically grounded, precisely controlled pain expressions is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Augmenting training with 3DPain raises subject-independent AUROC on the real UNBC-McMaster benchmark from 0.83 to 0.90; heatmap distillation adds another point to 0.91.
  • Rare high-intensity pain expressions (high PSPI) can be generated at any desired frequency, directly mitigating the severe class imbalance in real datasets.
  • Because demographic attributes are controlled by the mesh-generation prompts, the dataset provides balanced coverage of age, gender, and ethnicity that most clinical datasets lack.
  • The pipeline yields paired neutral/pain images, AU intensities, PSPI scores, and pain heatmaps per identity, enabling multi-task supervision—something that would be prohibitively expensive to annotate manually on real video.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • I infer that the same AU-rigging machinery could be retargeted to other clinically defined facial-expression domains (e.g., depression, fatigue, or stroke-related paresis) by swapping the target AU set, since the rigging and distillation components are not pain-specific.
  • The paper does not validate that the rendered faces actually match FACS-coded AUs or human pain perception; an independent AU-detector or rater study on 3DPain frames would be a decisive and publishable extension.
  • A possible confound: the UNBC improvement might stem from regularization or domain randomization rather than realistic pain content; ablating with random non-pain AU rigging would isolate whether the gain is pain-specific.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces 3DPain, a large synthetic dataset of 82,500 frames from 2,500 FLAME-based identities with demographic balance, AU-level rigging, PSPI scores, and pain-region heatmaps, generated through a three-stage pipeline (depth-conditioned diffusion for neutral faces, Hunyuan3D texturing, and Neural Face Rigging with inpainting). It also presents ViTPain, a Vision Transformer with AU-query cross-attention and heatmap-teacher/RGB-student distillation. On the UNBC-McMaster benchmark, ViTPain with 3DPain augmentation and heatmap supervision reaches 0.91 AUROC for binary pain classification, compared to a 0.83 AUROC baseline. The paper claims that 3DPain is a 'controllable, diverse, and clinically grounded foundation' for generalizable automated pain assessment, with ViTPain demonstrating the utility of synthetic data and heatmap distillation.

Significance. If the claims hold, the contribution is significant: a publicly released, large-scale, demographically balanced synthetic dataset with precise AU and PSPI annotations and heatmaps could address several long-standing limitations in automated pain assessment. The ViTPain architecture, with AU-query cross-attention and cross-modal distillation, is a reasonable design and the UNBC result, if robust, would be a useful step. The dataset itself, with its deliberate demographic balance, is a potentially valuable community resource. However, the current evidence does not yet establish the 'clinically grounded' claim: the synthetic expressions are not verified to match real pain morphology, the internal 3DPain ablations are partly self-consistent by construction, and the single external benchmark lacks control conditions and statistical grounding. The paper would be substantially strengthened by targeted validation and control experiments.

major comments (3)
  1. [§3, Neural Face Rigging for Controllable Expressions] The central claim that rigging produces clinically valid pain expressions is not verified on rendered images. The text states that 'By controlling the intensity of individual AUs, we can generate expressions corresponding to specific PSPI scores,' but no FACS-annotator agreement, no AU-detector evaluation on 3DPain images, and no human pain ratings are reported. Because the depth-to-image and inpainting stages (§3) can alter subtle expression cues, mesh-level AU displacement does not guarantee that the final 2D images exhibit the intended AUs. Please add a direct validation of AU fidelity (e.g., AU detector predictions on rendered images, or FACS coder agreement) and, ideally, a small human rating study to support the 'clinically grounded' language.
  2. [Table 2 and §3, Heatmap Generation] The 0.96 macro AUROC of the heatmap-trained teacher is partly a self-consistency check, not an independent measure of clinical realism. The heatmaps are computed from vertex displacements between neutral and rigged meshes, and the same mesh displacements define the AU and PSPI labels. Thus the teacher's high performance reflects that heatmaps and labels are derived from the same construction. This does not validate the visual fidelity of the expressions or the transferability of the heatmaps to real faces. Please report AU-detector or human-annotation agreement on 3DPain images, and evaluate the teacher on an independent real-image heatmap benchmark if available.
  3. [Table 3 and §5, Performance Comparison] The UNBC-McMaster result lacks a control augmentation condition and statistical significance reporting. The gains from 0.83 (baseline) to 0.90 (with 3DPain) and 0.91 (with heatmap supervision) are reported without per-fold standard deviations, confidence intervals, or significance tests, on a 25-subject 5-fold split. The improvement could plausibly reflect regularization or domain randomization rather than clinically valid pain content. Please include a control augmentation condition (e.g., adding non-pain synthetic faces with matched demographics, or label-shuffled 3DPain frames), and report per-fold results and significance testing. This is necessary to support the claim that 3DPain's pain-specific content, rather than generic synthetic-image augmentation, drives the improvement.
minor comments (5)
  1. [Abstract and Introduction] The dataset is inconsistently named '3DPain' and 'Pain3D' across the abstract, introduction, and Section 3. Please standardize.
  2. [§4.1, PSPI Classification Branch] The number of PSPI classes is inconsistent: the text says '17-class classification (scores 0-16)', while Table 2's caption says '16-class ordinal pain classification (PSPI 0-15)' and reports tolerance metrics accordingly. Please reconcile the class count and the PSPI range.
  3. [§3, Data Generation] The sampling ranges for AU intensities, FLAME shape/expression parameters, and the demographic prompt set are not specified, which limits reproducibility of the dataset-generation pipeline. Consider providing these details in an appendix or supplement.
  4. [Table 1] The Black racial/ethnic group contains only 82 identities out of 2,500. This is a relatively small subgroup and should be discussed as a limitation when claiming balanced demographic representation.
  5. [Throughout] There are several typographical errors, e.g., 'for for' in §3, 'non-communicative patient' in the abstract, and 'anatomically plausible' assertions without citations. A careful proofread is needed.

Circularity Check

2 steps flagged

Internal 3DPain heatmap-teacher results are self-consistency by construction; UNBC transfer is independently evaluated but the synthetic-label 'predictions' partially reduce to the defining equations.

specific steps
  1. self definitional [Section 3, 'Heatmap Generation'; Eq. (1); Table 2]
    "We generate AU heatmaps by comparing the meshes before and after neural face rigging. Specifically, we analyze changes in vertex positions between neutral FLAME meshes and their corresponding pain FLAME meshes to identify which action units have been activated and quantify their intensity. ... The teacher model trained on heatmaps alone reaches 0.96 macro AUROC and 0.67 exact accuracy, highlighting the value of the auxiliary modality."

    By Eq. (1), PSPI is a deterministic function of the six AU intensities (AU4, max(AU6,AU7), max(AU9,AU10), AU43). The heatmaps are computed directly from the vertex displacements produced by the neural rigging that is configured from those same sampled AU intensities. Thus the heatmap input contains a spatial encoding of exactly the variables that define the PSPI target. The teacher's 0.96 AUROC on 3DPain therefore measures label-construction self-consistency, not independent clinical validity or visual fidelity of the rendered expressions. Using this number to 'highlight the value of the auxiliary modality' partially reduces to the fact that the auxiliary modality is generated from the label-defining parameters.

  2. self definitional [Section 3, 'Neural Face Rigging for Controllable Expressions']
    "By controlling the intensity of individual AUs, we can generate expressions corresponding to specific PSPI scores, enabling the creation of training data with known pain levels."

    The claimed correspondence is definitional: the PSPI score is computed by Eq. (1) from the same AU intensities that drive the rigging. The paper presents this as yielding 'clinically validated pain levels,' but no evidence (FACS coding, AU detection, or human ratings) links the rigged parameters to observable expression in the rendered images. Consequently, the 'expressions corresponding to PSPI' is an identity between the label algebra and the rigging input, not an empirically verified prediction about real pain morphology. This makes the internal synthetic labels self-consistent rather than clinically validated.

full rationale

The paper's central external claim—3DPain augmentation plus heatmap distillation lifts ViTPain to 0.91 AUROC on the real UNBC-McMaster benchmark (Tab. 3)—is not circular: the benchmark labels are independent of the synthetic rigging parameters, and the evaluation is subject-independent 5-fold CV. The circularity is confined to the internal 3DPain evaluation. The PSPI target is defined by Eq. (1) from six AU intensities, and the 'pain-region heatmaps' are computed from the vertex displacements created by rigging with exactly those same AU intensities. Hence the heatmap teacher's 0.96 macro AUROC (Tab. 2) is a self-consistency measure: the input modality is a spatial decoding of the very parameters that define the target score. Similarly, the statement that controlling AU intensities yields expressions corresponding to PSPI scores is a definitional identity, not an empirical validation of visual fidelity. The paper cites its own SynPain dataset to motivate the need for 3D control, but that self-citation is motivational rather than load-bearing for the UNBC result. No uniqueness theorem or ansatz is smuggled via self-citation. Because the central transfer claim has independent external support, the paper is only partially circular; the synthetic-data-internal 'predictions' reduce by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim relies on assumptions that the synthetic faces are clinically plausible, that PSPI-like labels can be generated from sampled AUs, that FLAME plus diffusion preserves identity and demographics, and that heatmaps derived from rigging are meaningful teaching signals. The dataset and code are not released, so none of these assumptions can be independently checked from artifacts.

free parameters (6)
  • AU intensity sampling ranges = not reported
    Figure 2 says AUs are randomly sampled, but per-AU ranges and distributions are omitted; they drive PSPI distribution and realism.
  • FLAME shape and expression sampling = not reported
    Diverse identities come from varying FLAME parameters; no ranges or validation are given.
  • Loss weights lambda = 1.0, 1.0, 0.1, 0.3, 0.5
    Chosen by hyperparameter optimization on validation data in Section 4.2; they affect both training and distillation balance.
  • Distillation temperature T = 4.0
    Set in implementation details; softness of teacher targets depends on it.
  • Heatmap computation details = not reported
    Heatmaps use vertex distances and displacements, but exact normalization and aggregation are unspecified; teacher signal depends on them.
  • Demographic prompt set for Kandinsky = 2,500 prompts
    The pipeline uses 2,500 hand-written prompts to control ethnicity, age, gender, and attire; prompts are not released, so demographic composition is not reproducible.
axioms (5)
  • domain assumption PSPI formula (Eq. 1) is a valid clinical mapping of AUs to pain intensity.
    All synthetic labels are assigned through this formula; validity is taken from the cited clinical literature rather than demonstrated on the generated faces.
  • ad hoc to paper NFR-displaced meshes at sampled AU intensities look like real human pain expressions.
    No AU detector, human rater, or clinical review checks the rendered images; the synthetic-to-real transfer claim depends on this.
  • domain assumption FLAME shape and expression sampling spans the intended demographic groups.
    Demographic balance in Table 1 is achieved by varying FLAME parameters and prompts, with no independent check that output faces match each group.
  • domain assumption Diffusion-generated textures preserve the identity and geometry of the FLAME mesh.
    Kandinsky and Hunyuan3D outputs are assumed faithful enough for pain recognition training; there is no identity or geometry consistency evaluation.
  • ad hoc to paper Vertex-displacement heatmaps localize pain-relevant regions.
    Heatmaps are derived from mesh differences, but their anatomical and clinical relevance is not validated against any external standard.

pith-pipeline@v1.3.0-alltime-deepseek · 10779 in / 16508 out tokens · 139993 ms · 2026-08-04T16:04:09.132622+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Pain in 3D: Generating Controllable Synthetic Faces for Automated Pain Assessment." pith.science (2026). https://pith.science/paper/RHFV3EQC

@misc{pith2026250916727,
  author       = {Pith},
  title        = {Pith review of: Pain in 3D: Generating Controllable Synthetic Faces for Automated Pain Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RHFV3EQC}},
  note         = {Machine review of arXiv:2509.16727}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automated pain assessment from facial expressions is crucial for non-communicative patient. Progress has been limited by two challenges: (i) existing datasets exhibit severe demographic and label imbalance due to ethical constraints, and (ii) current generative models cannot precisely control facial action units (AUs), facial structure, or clinically validated pain levels. We introduce 3DPain, a large-scale synthetic dataset designed to overcome data scarcity in automated pain assessment. Comprising 82,500 frames across 2,500 unique identities, 3DPain offers extensive heterogeneity in facial pain responses across demographic groups balanced by age, gender, and ethnicity. Our three-stage framework samples diverse 3D meshes, textures them with diffusion models, and applies AU-driven face rigging to synthesize multi-view faces with paired neutral/pain images, facial action units, PSPI scores, and pain-region heatmaps. We further introduce ViTPain, a Vision Transformer based framework leveraging cross-attention with a neutral reference face to achieve identityaware pain estimation. Together, 3DPain and ViTPain establish a controllable, diverse, and clinically grounded foundation for generalizable automated pain assessment.

Figures

Figures reproduced from arXiv: 2509.16727 by Abhishek Moturu, Babak Taati, Soroush Mehraban, Xin Lei Lin.

Figure 1
Figure 1. Figure 1: 3DPain synthetic dataset. Our dataset provides multiview images and corresponding heatmaps of facial expressions, enabling robust training and evaluation for automatic pain recognition. ing equity and generalizability in automated pain detection [2, 7, 22]. To bridge the gap between current synthesis methods and clinical needs, we introduce Pain3D, a dataset containing 82,500 frames containing photorealist… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline for generating facial expressions of pain. The method first rigs the facial mesh using randomly sampled action units (AUs), then passes the frontal view depth of the mesh to a Depth ControlNet to synthesize a frontal face image. A texture generation model maps textures onto both the neutral and rigged meshes, and finally an inpainting model refines the hair and background using a background segmen… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed ViTPain framework. The model uses patch tokens with cross-attention to AU query tokens, supervised [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Quantitative comparison of the number of frames and heatmaps in the synthetic 3DPain dataset as well as the UNBC-McMaster [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 8 linked inside Pith

  1. [1]

    Kandinsky 2.2 controlnet depth model. 2, 4, 6

  2. [2]

    Dissecting racial bias in an algorithm used to manage the health of populations.Science, 366(6464):447–453, 2019. 1, 2, 5

    2019

  3. [3]

    arXiv preprint, 2024

    A synthetic dataset of pain and non-pain facial expressions. arXiv preprint, 2024. 1, 2, 3

    2024

  4. [4]

    Neural jaco- bian fields: Learning intrinsic mappings of arbitrary meshes

    Noam Aigerman, Kunal Gupta, Vladimir G Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. Neural jaco- bian fields: Learning intrinsic mappings of arbitrary meshes. arXiv preprint arXiv:2205.02904, 2022. 5

    Pith/arXiv arXiv 2022

  5. [5]

    Bedlam: A synthetic dataset and benchmark for fitting 3d human models to video

    Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset and benchmark for fitting 3d human models to video. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8784– 8794, 2023. 3

    2023

  6. [6]

    Idiff-face: Synthetic- based face recognition through fizzy identity-conditioned diffusion models.arXiv preprint arXiv:2308.04995, 2023

    Fadi Boutros, Marco Huber, Maximilian Klemt, Naser Damer, and Florian Kirchbuchner. Idiff-face: Synthetic- based face recognition through fizzy identity-conditioned diffusion models.arXiv preprint arXiv:2308.04995, 2023. 3

    Pith/arXiv arXiv 2023

  7. [7]

    Gender shades: Inter- sectional accuracy disparities in commercial gender classifi- cation

    Joy Buolamwini and Timnit Gebru. Gender shades: Inter- sectional accuracy disparities in commercial gender classifi- cation. InConference on Fairness, Accountability and Trans- parency, pages 77–91. PMLR, 2018. 1, 2, 5

    2018

  8. [8]

    Ilvr: Conditioning method for denoising diffusion probabilistic models, 2021

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models, 2021. 1, 2

    2021

  9. [9]

    Xinmeng Deng, Xu Xu, Yifan Zhang, Zhenhua Cheng, and Yanmin Qian. Audio-visual speech codecs: Rethinking audio-visual speech enhancement by re-synthesis.Proceed- ings of the ACM on Interactive, Mobile, Wearable and Ubiq- uitous Technologies, 7(1):1–24, 2023. 3

    2023

  10. [10]

    Diffusion models beat gans on image synthesis, 2021

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 3

    2021

  11. [11]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 6

    Pith/arXiv arXiv 2010

  12. [12]

    Paul Ekman and Wallace V . Friesen. Facial action coding system.Environmental Psychology & Nonverbal Behavior,

  13. [13]

    Jeni, Wanqiao Ding, and Jef- frey F

    Itir Onal Ertugrul, L ´aszl´o A. Jeni, Wanqiao Ding, and Jef- frey F. Cohn. Afar: A deep learning based tool for automated facial affect recognition. In2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019), pages 1–1, 2019. 8

    2019

  14. [14]

    Humans in 4d: Re- constructing and tracking humans with transformers, 2023

    Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Re- constructing and tracking humans with transformers, 2023. 6

    2023

  15. [15]

    Maisi: Medical ai for synthetic imaging — high-resolution 3d ct image generation with diffusion models.arXiv preprint arXiv:2409.11169, 2024

    Pengfei Guo, Can Zhao, Dong Yang, Ziyue Xu, Vish- wesh Nath, Yucheng Tang, Benjamin Simon, et al. Maisi: Medical ai for synthetic imaging — high-resolution 3d ct image generation with diffusion models.arXiv preprint arXiv:2409.11169, 2024. 1

    arXiv 2024

  16. [16]

    Thomas Hadjistavropoulos.Assessing pain in older persons with severe limitations in ability to communicate, pages 135–

  17. [17]

    Prkachin, Kenneth D

    Thomas Hadjistavropoulos, Keela Herr, Kenneth M. Prkachin, Kenneth D. Craig, Stephen J. Gibson, Andreas Lukas, and Jennifer H. Smith. Pain assessment in elderly adults with dementia.The Lancet Neurology, 13(12):1216– 1227, 2014. 1, 5

    2014

  18. [18]

    Zakia Hammal and Jeffrey F. Cohn. Automatic detection of pain intensity.Proceedings of the ACM International Con- ference on Multimodal Interaction, pages 47–52, 2012. 2, 6

    2012

  19. [19]

    Coyne, Marilyn McCaffery, Renee Man- worren, and Sylvia Merkel

    Keela Herr, Paula J. Coyne, Marilyn McCaffery, Renee Man- worren, and Sylvia Merkel. Pain assessment in the patient unable to self-report: position statement with clinical prac- tice recommendations.Pain Management Nursing, 12(4): 230–250, 2011. 1

    2011

  20. [20]

    Denoising diffu- sion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models, 2020. 1, 3, 4

    2020

  21. [21]

    Synthlight: Por- trait shadow removal with synthetic data.arXiv preprint arXiv:2501.09756, 2025

    Yingqing Hu, Zhihao Chen, Youjia Wang, Lingbo Yang, Wenhan Wang, and Qifeng Yang. Synthlight: Por- trait shadow removal with synthetic data.arXiv preprint arXiv:2501.09756, 2025. 3

    Pith/arXiv arXiv 2025

  22. [22]

    Biases in artificial intelligence application in pain medicine

    Oranicha Jumreornvong, Aliza M Perez, Brian Malave, Fa- timah Mozawalla, Arash Kia, and Chinwe A Nwaneshiudu. Biases in artificial intelligence application in pain medicine. Journal of Pain Research, 18:1021–1033, 2025. 2

    2025

  23. [23]

    Continuous pain intensity estimation from facial expres- sions

    Sebastian Kaltwang, Ognjen Rudovic, and Maja Pantic. Continuous pain intensity estimation from facial expres- sions. InAdvances in Visual Computing, pages 368–377, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg. 8

    2012

  24. [24]

    M. Kunz, D. Seuss, T. Hassan, J.U. Garbas, M. Siebers, U. Schmid, M. Schoberl, and S. Lautenbacher. Problems of video-based pain detection in patients with dementia: a road map to an interdisciplinary solution.BMC Geriatrics, 17: 1–8, 2017. 6

    2017

  25. [25]

    Semantic-guided inversion for versatile and high- fidelity face editing.arXiv preprint arXiv:2404.05063, 2024

    Kang Liu, Ting Yao, Luoqi Zhou, Yingwei Zhang, and Tao Mei. Semantic-guided inversion for versatile and high- fidelity face editing.arXiv preprint arXiv:2404.05063, 2024. 3

    Pith/arXiv arXiv 2024

  26. [26]

    Cohn, Kenneth M

    Patrick Lucey, Jeffrey F. Cohn, Kenneth M. Prkachin, Pa- tricia E. Solomon, and Iain Matthews. Painful data: The unbc-mcmaster shoulder pain expression archive database. In2011 IEEE International Conference on Automatic Face & Gesture Recognition (FG), pages 57–64. IEEE, 2011. 3, 7, 8

    2011

  27. [27]

    Parkhi, Andrea Vedaldi, and Andrew Zisserman

    Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. InProceedings of the British Machine Vision Conference (BMVC), pages 41.1–41.12. BMV A Press,

  28. [28]

    Prkachin and Patricia E

    Kenneth M. Prkachin and Patricia E. Solomon. The struc- ture, reliability and validity of pain expression: Evidence from patients with shoulder pain.Pain, 139(2):267–274,

  29. [29]

    Neural face rigging for animating and retargeting facial meshes in the wild

    Dafei Qin, Jun Saito, Noam Aigerman, Thibault Groueix, and Taku Komura. Neural face rigging for animating and retargeting facial meshes in the wild. InACM SIGGRAPH 2023 Conference Proceedings, New York, NY , USA, 2023. Association for Computing Machinery. 2, 5

    2023

  30. [30]

    Prkachin, Thomas Hadjistavropoulos, and Babak Taati

    Siavash Rezaei, Abhishek Moturu, Shun Zhao, Kenneth M. Prkachin, Thomas Hadjistavropoulos, and Babak Taati. Un- obtrusive pain monitoring in older adults with dementia us- ing pairwise and contrastive training.IEEE Journal of Biomedical and Health Informatics, 25(5):1450–1462, 2021. 3, 7, 8

    2021

  31. [31]

    Klados, Raman Chopra, Ab- bas Sohrabpour, and Bin He

    Shahed Rezaei, Michalis A. Klados, Raman Chopra, Ab- bas Sohrabpour, and Bin He. Conditional adversarial net- works for multi-subject facial expression generation with ac- tion unit intensity control.Frontiers in Signal Processing, 2,

  32. [32]

    Gonfaus, Kamal Nasrollahi, Thomas B

    Pau Rodriguez, Guillem Cucurull, Jordi Gonz `alez, Josep M. Gonfaus, Kamal Nasrollahi, Thomas B. Moeslund, and F. Xavier Roca. Deep pain: Exploiting long short-term memory networks for facial expression classification.IEEE Transactions on Cybernetics, 47(10):2898–2910, 2017. 8

    2017

  33. [33]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 1, 4, 6

    2022

  34. [34]

    Flame: Articulated expressive 3d head model (pytorch implementation), 2020

    Soubhik Sanyal. Flame: Articulated expressive 3d head model (pytorch implementation), 2020. 2, 3, 4

    2020

  35. [35]

    DiffusionNet: Discretization agnostic learning on surfaces.ACM Transactions on Graphics (TOG), 41(3): 1–16, 2022

    Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. DiffusionNet: Discretization agnostic learning on surfaces.ACM Transactions on Graphics (TOG), 41(3): 1–16, 2022. 5

    2022

  36. [36]

    Weiss, Niru Mah- eswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 1, 3

    2015

  37. [37]

    Synpain: A synthetic dataset of pain and non-pain facial expressions

    Babak Taati, Muhammad Muzammil, Yasamin Zarghami, Abhishek Moturu, Amirhossein Kazerouni, Hailey Reimer, Alex Mihailidis, and Thomas Hadjistavropoulos. Synpain: A synthetic dataset of pain and non-pain facial expressions

  38. [38]

    Hunyuan3d 1.0: A unified framework for text-to-3d and image-to-3d generation, 2024

    Tencent Hunyuan3D Team. Hunyuan3d 1.0: A unified framework for text-to-3d and image-to-3d generation, 2024. 5

    2024

  39. [39]

    Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material,

    Tencent Hunyuan3D Team. Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material,

  40. [40]

    Hunyuan3d 2.0: Scaling diffu- sion models for high resolution textured 3d assets generation,

    Tencent Hunyuan3D Team. Hunyuan3d 2.0: Scaling diffu- sion models for high resolution textured 3d assets generation,

  41. [41]

    Traue, Philipp Werner, Ayoub Al-Hamadi, Stephen Crawcour, Adriano O

    Steffen Walter, Sascha Gruss, Hagen Ehleiter, Jiaxi Tan, Har- ald C. Traue, Philipp Werner, Ayoub Al-Hamadi, Stephen Crawcour, Adriano O. Andrade, and Gustavo Moreira da Silva. The biovid heat pain database data for the advance- ment and systematic validation of an automated pain recog- nition system. InProceedings of the IEEE International Conference on ...

    2013

  42. [42]

    Fineface: High-quality face generation with fine-grained control.arXiv preprint arXiv:2407.20175,

    Haoqian Wang, Zhongcong Li, Jing Zhang, Ziwei Liu, and Chen Change Loy. Fineface: High-quality face generation with fine-grained control.arXiv preprint arXiv:2407.20175,

    Pith/arXiv arXiv

  43. [43]

    Magicface: High-fidelity facial expression editing with action-unit control, 2025

    Mengting Wei, Tuomas Varanka, Xingxun Jiang, Huai-Qian Khor, and Guoying Zhao. Magicface: High-fidelity facial expression editing with action-unit control, 2025. 1, 2

    2025

  44. [44]

    Fake it till you make it: Face analysis in the wild using synthetic data alone

    Erroll Woody, Luan Trinh, Esther Ahn, Georgios Pavlakos, and Stella X Yu. Fake it till you make it: Face analysis in the wild using synthetic data alone. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3681–3691, 2021. 3

    2021

  45. [45]

    Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond.arXiv preprint arXiv:2404.01243, 2024

    Tianxin Zhang, Aston Wu, Wayne Xin Zhao, and Ji-Rong Wen. Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond.arXiv preprint arXiv:2404.01243, 2024. 3

    Pith/arXiv arXiv 2024

  46. [46]

    Girard, Yue Wu, Xing Zhang, Peng Liu, Umur Ciftci, Shaun Canavan, Michael Reale, Andy Horowitz, Huiyuan Yang, Jeffrey F

    Zheng Zhang, Jeff M. Girard, Yue Wu, Xing Zhang, Peng Liu, Umur Ciftci, Shaun Canavan, Michael Reale, Andy Horowitz, Huiyuan Yang, Jeffrey F. Cohn, Qiang Ji, and Li- jun Yin. Multimodal spontaneous emotion corpus for human behavior analysis. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3438–3446, 2016. 3

    2016

  47. [151]

    IASP Press, Seattle, 2005. 1

    2005

  48. [2025]

    arXiv:2507.19673v2 [cs.CV], 2025. 3

    Pith/arXiv arXiv 2025