Pith. sign in

REVIEW 5 major objections 7 minor 38 references

3D Hand Mesh-Guided AI-Generated Malformed Hand Refinement with Hand Pose Transformation via Diffusion Model

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read 3D hand meshes refine AI-generated hands more reliably than depth maps, the paper claims, and also enable reference-pose hand transformation without extra training.

desk verdict Useful applied paper; the mesh-vs-depth claim is confounded by estimator choice, but the dataset and double-check are real contributions. read the letter →

arxiv 2506.12680 v2 pith:3DURMLBF submitted 2025-06-15 cs.CV

classification cs.CV
keywords 3DhandmeshmalformedrefinementdiffusioninpaintingposetransformationStableControlNetInterWilddoublecheck
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 tries to establish that conditioning a diffusion inpainting model on 3D hand meshes, rather than depth maps, produces more realistic refinements of malformed hands in AI-generated images. The mesh carries surface orientation details, such as whether a pixel belongs to the palm or the back of the hand, which depth guidance lacks. The authors also claim a training-free affine pose transformation that lets the refined hand imitate the pose in a reference image. They build a 24,411-sample mesh-image-mask training set, add a double check filter for erroneous mesh estimates, and report lower FID and KID scores plus higher detection confidence than Stable Diffusion, ControlNet, and HandRefiner. If true, this positions mesh guidance as a stronger structural prior for hand restoration than the depth-based prior used by previous refinement systems.

What carries the argument

The load-bearing object is the rendered 2D grayscale map of the 3D hand mesh produced by InterWild; it carries surface geometry and orientation information that a depth map cannot represent, such as palm-versus-back direction. This mesh map is concatenated with the masked image and fed through a ControlNet branch, initialized from depth-conditioning weights and fine-tuned on the authors' dataset, into the Stable Diffusion U-Net. The inpainting loop follows Repaint, re-noising the unmasked background at each DDIM step and denoising only the hand region before one final mask-free DDIM step to harmonize the boundary. The double check algorithm uses YOLO boxes and Mediapipe keypoints to decide whether a left or right hand exists, steering InterWild to output a single robust mesh. Pose transformation is a separate affine mechanism that resets scale, translation, and rotation of the reference mesh to match the input's wrist-to-pinky geometry, then reuses the same refinement model.

What would settle it

Compute per-image hand accuracy on a test set where the double-checked InterWild meshes are known to be wrong; if refined outputs on those images are no better than unguided inpainting, the mesh-guidance claim collapses. Alternatively, rerun the Hagrid comparison with error bars or a hand-specific score such as percentage of correct keypoints; if the margin over HandRefiner vanishes, the central claim is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that replacing depth guidance with 3D mesh guidance in a diffusion inpainting pipeline improves malformed-hand refinement, and that the same pipeline can transform the hand to a new pose without any additional training. The paper shows that rendered grayscale maps of the estimated 3D hand mesh, injected through a ControlNet-style adapter into a Stable Diffusion inpainting model with Repaint-style masking and DDIM sampling, achieve FID 80.11, KID 0.074, and detection confidence 0.92 on the Hagrid reference, versus FID 80.67, KID 0.075, and 0.91 for HandRefiner; on the FreiHAND reference it reports FID 125.75 versus 126.83. It further claims that the double check algorithm, which combines YOLO bounding boxes with Mediapipe left/right keypoint checks, removes spurious second-hand mesh predictions and contributes to the gain. The pose transformation aligns the input and reference meshes through scale, translation, and rotation about the wrist keypoint, then runs the same refinement within the union bounding box to render a new hand pose.

Load-bearing premise

The method assumes that the 3D mesh estimator delivers sufficiently accurate guidance on the very malformed hands it is meant to fix, and that the FID, KID, and detection-confidence protocol is sensitive enough to reflect true hand quality; if either assumption fails, the reported improvement may not be real.

Editorial extensions

If this is right

  • If mesh guidance is genuinely superior, future hand-refinement systems should condition on surface meshes rather than depth, and existing depth-initialized ControlNet weights can be repurposed as a starting point.
  • The training recipe combining real lab data, relighted in-the-wild data, and reannotated fashion images offers a reusable template for expanding hand refinement to more diverse skin tones, backgrounds, and hand-object contexts.
  • The double check algorithm provides a cheap, inference-time filter for any hand mesh estimator that occasionally emits spurious second-hand predictions on single-hand images.
  • Because pose transformation requires no additional training, one mesh-guided model can serve both restoration and re-pose use cases, and the affine alignment procedure may generalize to other articulated structures with a distinctive wrist-like joint.
  • The reported margins over HandRefiner are small in absolute terms, so the practical significance of the FID and KID gains depends on whether they are consistent across individual images rather than averaged over the whole test set.

Reading between the lines

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

  • A testable extension is to replace InterWild with another mesh estimator and measure whether the FID gain persists; if the gain disappears, the improvement may come from the training data or masking rather than from mesh guidance per se.
  • The palm-versus-back motivation suggests the same mesh prior could help other surface-ambiguity problems in generated humans, such as foot orientation or ear shape, where depth maps are similarly ambiguous.
  • Because the reported FID/KID differences are small and given without error bars, a hand-specific metric, such as keypoint accuracy or mesh chamfer distance on known-pose test images, would settle whether mesh guidance is the true cause of the perceived improvement.
  • The double check only filters one failure mode, so extending it to also score mesh confidence or reject implausible joint angles could make the pipeline robust against other InterWild errors on severely malformed hands.
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

5 major / 7 minor

Summary. The paper proposes a diffusion-based inpainting framework for refining malformed hands in AI-generated images. The key idea is to replace the depth-map guidance used by HandRefiner/ControlNet with 3D hand meshes estimated by InterWild, arguing that meshes preserve hand-structure details such as palm/back orientation. The authors construct a 24,411-sample training set from InterHand, ReIH, and a reannotated FashionIQ subset; introduce a double-check heuristic that uses YOLO and MediaPipe to suppress spurious second-hand mesh predictions; and propose a training-free affine pose-transformation procedure that aligns a reference hand pose before refinement. Quantitative evaluation on Hagrid and FreiHAND reports small FID/KID/detection-confidence gains over Stable Diffusion, ControlNet, and HandRefiner, and the pose-transformation contribution is shown only with qualitative examples.

Significance. If the central claim were established, the mesh-guidance idea would be a plausible and useful step beyond depth-guided hand refinement, and the curated 24,411-triplet dataset plus the double-check heuristic are practical assets. The paper also states a clear limitation (occluded hands and hand-object interaction) and proposes a training-free pose-transfer mechanism, which is attractive for deployment. However, the empirical core does not currently isolate the causal factor: the comparison changes both the conditioning modality and the pose estimator, all reported margins are very small, and no uncertainty quantification is provided. The pose-transformation contribution is supported only by visual examples. These issues require substantial additional experimentation before the superiority claims can be accepted.

major comments (5)
  1. [§4.4, Tables 1 and 2] The comparison is confounded: the proposed method uses 3D meshes from InterWild, while the HandRefiner and ControlNet baselines use depth maps from Graphormer. The reported improvements (Hagrid FID 80.67 to 80.11 and KID 0.075 to 0.074; FreiHAND FID 126.83 to 125.75) therefore cannot be attributed to mesh guidance per se; they may reflect better pose estimation by InterWild. Please add ablations that hold the estimator fixed, e.g., a depth map rendered from the same InterWild mesh fed into the same pipeline, or a mesh-conditioned variant using Graphormer meshes, so that modality and estimator effects are separated.
  2. [§4.4, Tables 1-4] No error bars, repeated seeds, or significance tests are reported for any of the FID/KID/detection-confidence results. Differences as small as 0.001 KID and 0.01 detection confidence are plausibly within run-to-run noise. Please report bootstrap confidence intervals for FID/KID over the generated sets, run at least three seeds, and, where possible, a significance test; otherwise the claim of superior performance is not quantitatively established.
  3. [§4.5.1, Table 3] The double-check ablation is internally inconsistent. The text says the table evaluates the method without the double-check algorithm, but the first row is labeled w/ Double Check and the second row is labeled Ours. Please clarify which configuration produced each set of numbers and relabel the rows; in its current form the effect of the double-check cannot be verified.
  4. [§3.5, Figure 12] The second contribution, hand pose transformation, is evaluated only qualitatively. There is no quantitative measure of whether the output actually mimics the reference pose, nor any image-quality metric for the transformed outputs. Please add a metric such as keypoint alignment error against the target pose, FID/KID on the transformed outputs, or a user study, so that the pose-transformation claim can be assessed.
  5. [§3.3-3.4] The method's guidance quality depends on the accuracy of InterWild meshes on malformed in-the-wild hands, but this premise is not directly tested. The double-check heuristic filters one specific failure mode (spurious second-hand predictions) and does not address general mesh errors. Please include a failure analysis or report mesh-estimation accuracy on the malformed-hand evaluation set, and discuss how estimation errors propagate to the refined outputs.
minor comments (7)
  1. [§1 and Abstract] There are copy-editing issues, e.g., 'Stable diffusion [1] have recently attracted a lot of attention' and 'The depth map-based hand pose estimator fails'; please proofread the manuscript.
  2. [§2.2] Mesh Graphomer should be Mesh Graphormer, and the spelling Interhand is used inconsistently with InterHand.
  3. [§4.4] In the FreiHAND evaluation, the text says hand region mask by Interhand [30] but the method uses InterWild; please correct the name and reference.
  4. [Table 4] The row InterHand + ReIH contains a missing space and reads 0.0740.92; it should be 0.074 0.92.
  5. [References] References [23] and [35] cite the same FreiHAND paper; please merge them or disambiguate the citations.
  6. [§3.3, Eqs. (2)-(3)] The notation DDIM(·) is used without explicitly defining how xmask and M esh enter the DDIM step; please clarify.
  7. [§4.3] The implementation details report 2,307 training steps but do not state the number of epochs or the train/validation split of the 24,411-triplet dataset; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical refinement pipeline evaluated against external benchmarks, with mesh guidance from an external estimator and a deterministic affine pose transform.

full rationale

The paper's central claim is that a 3D mesh-guided diffusion inpainting model refines malformed hands better than depth-guided baselines. This is supported by training on a collected dataset (InterHand, ReIH, and FashionIQ annotated with InterWild meshes) and evaluating FID, KID, and detection confidence against Stable Diffusion, ControlNet, and HandRefiner on external reference datasets Hagrid and FreiHAND. The mesh guidance is supplied by InterWild, an external estimator, and the hand pose transformation is a deterministic affine alignment computed from Mediapipe keypoints; neither reduces by construction to the reported improvements. No fitted parameter is renamed as a prediction, and no load-bearing claim depends on a self-citation chain. The comparison between mesh guidance and depth guidance does change the pose estimator (InterWild vs. Graphormer) simultaneously with the conditioning modality, which is a valid experimental-control concern but not a circularity. The apparent mislabeling in Table 3 ('w/ Double Check' in the ablation row) and the absence of error bars are reporting and evaluation-robustness issues, not circular reasoning. Overall, the derivation and evaluation are self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the quality of external hand mesh estimation, the validity of the diffusion inpainting formulation, the sufficiency of an affine pose alignment, and the sensitivity of the chosen evaluation metrics. No new physical entities are introduced.

free parameters (3)
  • Training dataset composition = InterHand 5,472 + ReIH 13,918 + Fashion 5,021 = 24,411 triplets
    The mixture of sub-datasets is chosen by hand; Table 4 shows that different compositions change FID, so the central result depends on this choice.
  • Double-check criterion = Mediapipe keypoints 1-8 must fall inside the YOLO bounding box
    Algorithm 1 uses an ad hoc rule to decide left/right hand existence; no sensitivity analysis or learned thresholds are provided.
  • Fine-tuning hyperparameters = learning rate 2e-5, batch size 4, 2307 steps
    These are chosen without reported search or convergence analysis; they affect the trained model but are standard values.
assumptions (4)
  • standard math Diffusion inpainting with DDIM and RePaint equations (Eqs. 1-3) is a valid conditional generation method.
    The paper builds on [1,5,32] without proving these steps, relying on established diffusion results.
  • domain assumption Interwild [30] provides accurate 3D hand meshes for malformed in-the-wild hands.
    Section 3.3 uses Interwild at inference; double check only filters some two-hand errors, not all mesh errors.
  • domain assumption An affine transform with isotropic scale, translation, and wrist-centered rotation suffices to align hand poses.
    Section 3.5 uses this without validation for pose transfer; it ignores out-of-plane rotation and finger articulation changes.
  • domain assumption FID/KID and hand detection confidence score are valid proxies for hand refinement quality.
    Sections 4.2 and 4.4 evaluate hand quality using these global metrics following [2].

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D Hand Mesh-Guided AI-Generated Malformed Hand Refinement with Hand Pose Transformation via Diffusion Model." pith.science (2026). https://pith.science/paper/3DURMLBF

@misc{pith2026250612680,
  author       = {Pith},
  title        = {Pith review of: 3D Hand Mesh-Guided AI-Generated Malformed Hand Refinement with Hand Pose Transformation via Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3DURMLBF}},
  note         = {Machine review of arXiv:2506.12680}
}
read the original abstract

The malformed hands in the AI-generated images seriously affect the authenticity of the images. To refine malformed hands, existing depth-based approaches use a hand depth estimator to guide the refinement of malformed hands. Due to the performance limitations of the hand depth estimator, many hand details cannot be represented, resulting in errors in the generated hands, such as confusing the palm and the back of the hand. To solve this problem, we propose a 3D mesh-guided refinement framework using a diffusion pipeline. We use a state-of-the-art 3D hand mesh estimator, which provides more details of the hands. For training, we collect and reannotate a dataset consisting of RGB images and 3D hand mesh. Then we design a diffusion inpainting model to generate refined outputs guided by 3D hand meshes. For inference, we propose a double check algorithm to facilitate the 3D hand mesh estimator to obtain robust hand mesh guidance to obtain our refined results. Beyond malformed hand refinement, we propose a novel hand pose transformation method. It increases the flexibility and diversity of the malformed hand refinement task. We made the restored images mimic the hand poses of the reference images. The pose transformation requires no additional training. Extensive experimental results demonstrate the superior performance of our proposed method.

Figures

Figures reproduced from arXiv: 2506.12680 by the authors.

Figure 1
Figure 1. The visual comparison of input malformed hand image by Stable Diffusion [1], refined result by ControlNet [6], refined result by HandRefiner [2], and refined result by our method. The former two methods use depth pose, which lacks details and is thus hard to guide the refining model to generate the palm or the back of the hand. Our method uses 3D hand mesh that provides rich details for generating realistic hands. •… view at source ↗
Figure 2
Figure 2. The visual result of our proposed hand pose transformation. The outputs change pose according to reference images. Faces of the real person are obscured due to privacy protection. [10], uses two kinds of human pose templates, the open palm and the fist, to fix the badly shaped hand. Hand1000 [11] uses 1000 hand images to fine-tune the stable diffusion model, resulting in better hand image generation capability. It c… view at source ↗
Figure 3
Figure 3. Training dataset visualization. We have three sub-datasets from different scenes. 3.2. Training Dataset Preparation Previous works [2, 16] utilize hand depth as guidance for malformed hand refinement. They are limited by the coarseness of depth maps and cannot provide sufficiently clear details. Therefore, we train our model using 3D hand mesh as hand pose guidance. Specifically, we sample images and ground-truth ha… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The illustration of our 3D hand mesh guided hand refinement pipelines combined with double check algorithm. The 3D mesh estimator are used during inference. 3.3. 3D Hand Mesh Guided Diffusion Inpainting 3D Hand Mesh Guidance. To ensure the refined hand is both precise …
Figure 5
Figure 5. Figure 5: The illustration of the double check. It provides more credible results for the hand image generation model. keypoints or the right hand keypoint predicted by Mediapipe are within the range of the YOLO bounding box, we consider that the left or right hand exists. We pr…
Figure 6
Figure 6. Figure 6: The illustration of the hand pose transformation process. We use a 3D hand mesh estimator to extract hand pose, then use an affine transformation to align pose location and scale, and then refine hands with our refinement method. movement. We compute the coordinate shi…
Figure 7
Figure 7. Figure 7: Visualization of the reference dataset: Hagrid [34] and FreiHAND [35]. The faces of the real persons in the pictures are obscured due to privacy protection. 4.2. Evaluation metrics Following the baseline model HandRefiner [2], we adopt the Frechet Inception Distance (F…
Figure 8
Figure 8. Figure 8: Visual results on the Hagrid dataset. The results have the whole human body with hands. see that our generated refinement results are better than previous methods. Pose Transformation. We provide a pose transformation experiment using human pose images from the Hagrid …
Figure 9
Figure 9. Figure 9: Visual results on the Hagrid dataset. The results have the whole human body with hands. 4.5. Ablation Studies 4.5.1. Double Check We evaluate the performance of our method without using the double￾check algorithm for hand mesh pre-processing during inference. The resul…
Figure 10
Figure 10. Figure 10: Visual results on the FreiHAND dataset. The results are cropped hand regions [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Visual results from the ablation study of the double-check mechanism. It clearly demonstrates the mechanismâĂŹs ability to eliminate wrong hand predictions. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Visual results of pose transformation. The faces are masked for privacy protection. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 30 canonical work pages

  1. [1]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High- resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10684–10695

  2. [2]

    W. Lu, Y. Xu, J. Zhang, C. Wang, D. Tao, Handrefiner: Refining malformed hands in generated images by diffusion-based conditional inpainting, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 7085–7093

  3. [3]

    K. Lin, L. Wang, Z. Liu, Mesh graphormer, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12939– 12948

  4. [4]

    Zhang, V

    F. Zhang, V. Bazarevsky, A. Vakunov, A. Tkachenka, G. Sung, C. Chang, M. Grundmann, Mediapipe hands: On-device real-time hand tracking, ArXiv Preprint ArXiv:2006.10214 (2020)

  5. [5]

    J. Song, C. Meng, S. Ermon, Denoising diffusion implicit models, arXiv preprint arXiv:2010.02502 (2020). 19

  6. [6]

    Zhang, A

    L. Zhang, A. Rao, M. Agrawala, Adding conditional control to text-to- image diffusion models, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3836–3847

  7. [7]

    Y. Zhou, J. Qian, H. Zhang, X. Xu, H. Sun, F. Zeng, Y. Zhou, Adaptive multi-text union for stable text-to-image synthesis learning, Pattern Recognition 152 (2024) 110438

  8. [8]

    W. Song, Z. Ye, M. Sun, X. Hou, S. Li, A. Hao, Attridiffuser: Adversari- ally enhanced diffusion model for text-to-facial attribute image synthesis, Pattern Recognition (2025) 111447

Show all 38 references
  1. [9]

    T. Zhen, J. Cao, X. Sun, J. Pan, Z. Ji, Y. Pang, Token-aware and step- aware acceleration for stable diffusion, Pattern Recognition 164 (2025) 111479

  2. [10]

    Z. Qin, Y. Zhang, Y. Liu, D. Campbell, Handcraft: Anatomically correct restoration of malformed hands in diffusion generated images, ArXiv Preprint ArXiv:2411.04332 (2024)

  3. [11]

    Zhang, B

    H. Zhang, B. Zhu, Y. Cao, Y. Hao, Hand1000: Generating realistic hands from text with only 1,000 images, ArXiv Preprint ArXiv:2408.15461 (2024)

  4. [12]

    J. Park, K. Kong, S. Kang, Attentionhand: Text-driven controllable hand image generation for 3d hand reconstruction in the wild, in: European Conference on Computer Vision, 2024, pp. 329–345

  5. [13]

    P. Kwon, H. Joo, Graspdiffusion: Synthesizing realistic whole-body hand-object interaction, ArXiv Preprint ArXiv:2410.13911 (2024)

  6. [14]

    Y. Pang, R. Shao, J. Zhang, H. Tu, Y. Liu, B. Zhou, H. Zhang, Y. Liu, Manivideo: Generating hand-object manipulation video with dexterous and generalizable grasping, ArXiv Preprint ArXiv:2412.16212 (2024)

  7. [15]

    Liang, X

    Y. Liang, X. Hu, B. Jiang, D. Luo, K. Wu, W. Han, T. Jin, C. Wang, Vton-handfit: Virtual try-on for arbitrary hand pose guided by hand priors embedding, ArXiv Preprint ArXiv:2408.12340 (2024)

  8. [16]

    C. Wang, P. Liu, M. Zhou, M. Zeng, X. Li, T. Ge, B. Zheng, Rhands: Refining malformed hands for generated images with decoupled structure 20 and style guidance, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, pp. 7573–7581

  9. [17]

    Pelykh, O

    A. Pelykh, O. Sincan, R. Bowden, Giving a hand to diffusion models: a two-stage approach to improving conditional human image generation, in: 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), 2024, pp. 1–10

  10. [18]

    T. Eum, J. Choi, T. Kim, Mghand: Multi-modal guidance for authentic hand diffusion, ArXiv Preprint ArXiv:2503.08133 (2025)

  11. [19]

    B. Wang, J. Zhou, J. Bai, Y. Yang, W. Chen, F. Wang, Z. Lei, Real- ishuman: A two-stage approach for refining malformed human parts in generated images, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, pp. 7509–7517

  12. [20]

    Q. Fu, X. Chen, M. Asad, S. Yuan, C. Oh, G. Slabaugh, Handrawer: Leveraging spatial information to render realistic hands using a condi- tional diffusion model in single stage, ArXiv Preprint ArXiv:2503.02127 (2025)

  13. [21]

    K. Gao, X. Liu, P. Ren, H. Chen, T. Zhen, L. Xie, Z. Li, Y. Yan, H. Zhang, E. Yin, Progressively global–local fusion with explicit guidance for accurate and robust 3d hand pose reconstruction, Knowledge-Based Systems 304 (2024) 112532

  14. [22]

    Romero, D

    J. Romero, D. Tzionas, M. J. Black, Embodied hands: Modeling and capturing hands and bodies together, arXiv preprint arXiv:2201.02610 (2022)

  15. [23]

    Zimmermann, D

    C. Zimmermann, D. Ceylan, J. Yang, B. Russell, M. Argus, T. Brox, Freihand: A dataset for markerless capture of hand pose and shape from single rgb images, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 813–822

  16. [24]

    Zimmermann, T

    C. Zimmermann, T. Brox, Learning to estimate 3d hand pose from single rgb images, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 4903–4911

  17. [25]

    Kapitanov, K

    A. Kapitanov, K. Kvanchiani, A. Nagaev, R. Kraynov, A. Makhliarchuk, Hagrid–hand gesture recognition image dataset, in: Proceedings of the 21 IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 4572–4581

  18. [26]

    AI, Static gestures dataset, https://synthesis.ai/ static-gestures-dataset/, data retrieved from Synthesis AI (2023)

    S. AI, Static gestures dataset, https://synthesis.ai/ static-gestures-dataset/, data retrieved from Synthesis AI (2023)

  19. [27]

    Tzionas, L

    D. Tzionas, L. Ballan, A. Srikantha, P. Aponte, M. Pollefeys, J. Gall, Capturing hands in action using discriminative salient points and physics simulation, International Journal of Computer Vision 118 (2016) 172–193

  20. [28]

    Moon, S.-I

    G. Moon, S.-I. Yu, H. Wen, T. Shiratori, K. M. Lee, Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image, in: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16, Spring...

  21. [29]

    G. Moon, S. Saito, W. Xu, R. Joshi, J. Buffalini, H. Bellan, N. Rosen, J. Richardson, M. Mize, P. De Bree, et al., A dataset of relighted 3d interacting hands, Advances in Neural Information Processing Systems 36 (2023) 17689–17701

  22. [30]

    G. Moon, Bringing inputs to shared domains for 3d interacting hands recovery in the wild, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17028–17037

  23. [31]

    H. Wu, Y. Gao, X. Guo, Z. Al-Halah, S. Rennie, K. Grauman, R. Feris, Fashion iq: A new dataset towards retrieving images by natural language feedback, in: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2021, pp. 11307–11317

  24. [32]

    Lugmayr, M

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, L. Van Gool, Repaint: Inpainting using denoising diffusion probabilistic models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11461–11471

  25. [33]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Uni- fied, real-time object detection, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779–788. 22

  26. [34]

    Kapitanov, A

    A. Kapitanov, A. Makhlyarchuk, K. Kvanchiani, Hagrid - hand gesture recognition image dataset (2022).arXiv:2206.08219

  27. [35]

    Zimmermann, D

    C. Zimmermann, D. Ceylan, J. Yang, B. Russell, M. Argus, T. Brox, Freihand: A dataset for markerless capture of hand pose and shape from single rgb images, in: Proceedings of the IEEE International Conference on Computer Vision, 2019

  28. [36]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, S. Hochreiter, Gans trained by a two time-scale update rule converge to a local nash equilibrium, in: Advances In Neural Information Processing Systems, 2017

  29. [37]

    BiÅĎkowski, D

    M. BiÅĎkowski, D. J. Sutherland, M. Arbel, A. Gretton, Demystifying mmd gans, in: International Conference on Learning Representations, 2018

  30. [38]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, in: International Conference on Learning Representations, 2019. 23

Pith tools

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