Pith. sign in

REVIEW 3 major objections 4 minor 40 references

SILK: Smooth InterpoLation frameworK for motion in-betweening A Simplified Computational Approach

T0 review · 3 major / 4 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 SILK shows that a single Transformer encoder, with zero-filled gaps and root-space pose features, matches or beats more complex motion in-betweening models on LaFAN1.

desk verdict Useful empirical paper on data-centric choices, but its central simplicity claim rests on a velocity feature leak that gives SILK ground-truth access to a frame it should predict. read the letter →

arxiv 2506.09075 v1 pith:TXBDEUI2 submitted 2025-06-09 cs.GR cs.CVcs.LG

Elly Akhoundi , Hung Yu Ling , Anup Anand Deshmukh , Judith Butepage This is my paper
classification cs.GRcs.CVcs.LG
keywords motionin-betweeningTransformerencoderLaFAN1zero-filledinputroot-spaceposefeaturesvelocitydata-centricsynthesissingle-shotprediction
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

The paper argues that high-quality motion in-betweening does not require a complex, multi-module architecture. A single Transformer encoder, fed with zero-filled gaps instead of interpolated frames, root-space pose features, velocity inputs, and a four-times-larger training sample extracted from the same data, matches or beats several heavier published models on the LaFAN1 benchmark. The result challenges the field's recent push toward skeleton-aware, multi-encoder, or multi-stage designs.

What carries the argument

The central mechanism is a single encoder-only Transformer that treats in-betweening as sequence-to-sequence with the missing frames set to all zeros. The zero pattern gives the model an implicit signal for which frames to predict, while learned relative position encodings supply temporal ordering, so no keyframe-position embeddings or masking are needed. Supporting machinery: root-space pose features (the hip trajectory projected on the ground plus joint transforms expressed relative to the root), input-only velocity features, a single L1 loss, and a denser training set produced by slicing LaFAN1 animations at offset 5.

What would settle it

Re-run every published baseline, including the omitted phase-manifold and space-time-control models and any with divergent SLERP or RMIB numbers, on the LaFAN1 subject-5 test set using exactly SILK's preprocessing: offset-5 slicing, same context length, root-space evaluation, and the same metric code. If any excluded model then reports lower L2P at 30 or 45 frames than SILK, the paper's claim that a single encoder matches or beats complex models would be contradicted.

Watch

Extended reading notes

Core claim

SILK is a minimalist, non-autoregressive Transformer encoder that produces all in-between frames in one pass. Missing frames are simply filled with zeros; the model uses learned relative position embeddings and no masking, so every frame attends to every other. On the standard LaFAN1 protocol, SILK reports L2P 0.13/0.38/0.83/1.59 and NPSS 0.0012/0.018/0.105/0.30 for gaps of 5/15/30/45 frames, matching or beating the two-stage TST model (0.10/0.39/0.89/1.68 and 0.0011/0.019/0.112/0.32) and other state-of-the-art entries. The paper attributes the gains to data and representation choices: sampling training transitions with an offset of 5 instead of 20, representing poses in root space, feeding velocities as input only, and zero-filling rather than SLERP-filling missing frames. It concludes that model complexity is not the primary determinant of animation quality.

Load-bearing premise

The central claim rests on the assumption that all compared models were evaluated under exactly the same protocol; since the paper only re-runs TST and excludes papers whose published SLERP and RMIB numbers diverge, a change in ranking under identical conditions would weaken the simplicity-wins conclusion.

Editorial extensions

If this is right

  • A single Transformer encoder suffices for competitive motion in-betweening; multi-encoder and multi-stage designs such as TST and Delta Interpolator are not required by the task.
  • Zero-filling missing frames outperforms SLERP pre-filling, contradicting a common design choice in prior work.
  • Increasing training data volume by denser slicing helps simple models and can close or reverse the gap to more complex architectures.
  • Root-space pose representation and input velocity features each improve quality enough to be worth their extra runtime costs of inverse kinematics and an additional specified frame.
  • Trained on gaps up to 30 frames, SILK extrapolates to 45-frame gaps without retraining.

Reading between the lines

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

  • Because the paper only re-runs TST under offset-5 sampling and omits methods whose published SLERP and RMIB numbers diverge, the Table 1 ranking should be treated as conditional; re-evaluating the omitted methods under identical preprocessing could shift it.
  • The offset-5 result suggests a capacity-dependent data regime: denser overlapping slices help a six-layer encoder but slightly hurt the larger TST, so the best sampling density may need to be tuned per architecture.
  • Zero-filling outperforming SLERP hints that pre-filling injects a bias toward the interpolated path; a direct test would compare zero-filling against randomized or learned placeholder fills.
  • SILK is deterministic and single-shot, so combining its data-centric recipe with a lightweight generative head is a natural next step that the authors point toward in future work.
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 / 4 minor

Summary. The paper proposes SILK, a single Transformer encoder for motion in-betweening. The missing frames are filled with zeros, the pose is represented in root space, velocity features are added to the input, and the training data are sampled with an offset of 5 rather than the offset of 20 used in prior work. On the LaFAN1 benchmark, the authors report L2P, L2Q, and NPSS numbers and claim that SILK achieves comparable or superior performance to more complex state-of-the-art models such as TST and ST-TMIB. Ablations study pose representation, velocity features, filling strategy, and key-position embeddings. The central claim is not established by the current experiments because the velocity features leak ground-truth frames that are not available in the standard in-betweening setting, and because the comparison against published baseline numbers is not fully controlled.

Significance. If the central claim held, the paper would make a useful practical point: simple architectures can match complex ones when data representation, sampling density, and input features are chosen carefully. The ablations are well designed and informative, especially the offset study in Table 2, the root-space versus local-space comparison in Table 3, and the zero-filling versus SLERP comparison in Table 5. The paper also makes a genuine effort to compare against a broad set of prior work on a standardized benchmark and to be transparent about excluded papers. However, the velocity-feature leakage in Section 4.5.2 directly undermines the headline claim, and the absence of error bars and code release limits reproducibility. The data-centric insights are still valuable, but they need to be separated from an invalid cross-method comparison.

major comments (3)
  1. [§4.5.2, Table 4] The velocity features leak information that is not available in the standard in-betweening setting. Section 4.5.2 states that velocities are computed from two consecutive frames, 'specifically, requiring one additional frame after both the start and target keyframes.' For the start keyframe, that additional frame is exactly the first in-between frame that SILK is supposed to predict; for the target keyframe, the additional frame lies outside the interpolation window. Thus the 'Input Vel' configuration receives ground-truth values of the first output frame and of a frame beyond the target, while the baselines in Table 1 do not. Table 4 quantifies the impact: removing velocity inputs ('Static Only') raises L2P from 1.59 to 2.1 at length 45 and from 0.83 to 1.0 at length 30, while TST reports 1.68 and 0.89, respectively. Under the standard MIB protocol, SILK would therefore be worse than TST at these lengths. The claim that a single Transformer encoder performs comparably to or better than complex models is not supported by the current experiments. The authors should re-run the evaluation without this leakage or explicitly redefine the task as in-betweening with additional temporal context and compare against baselines under matched conditions.
  2. [§4.3, Table 1] The comparison in Table 1 relies on published numbers from different papers without verifying that all methods were evaluated under exactly the same preprocessing, context length, and evaluation code. The paper excludes [5] and [28] because of discrepancies in SLERP/RMIB baselines, which is reasonable, but the same protocol concern applies to the included methods. Table 2 shows that training offset alone changes TST's L2P from 0.89 to 0.93 at length 30 and from 1.68 to 1.73 at length 45, which is the same scale as the differences between SILK and its closest competitors (e.g., SILK vs. ST-TMIB at length 30: 0.83 vs. 0.89). To support the central claim, the authors should run the included baselines under identical conditions or clearly limit the claim to a published-numbers comparison and release the evaluation code.
  3. [§4.2, Tables 1–5] All quantitative results are reported from single training runs without confidence intervals or significance tests. Several comparisons at short gaps differ by only 0.01–0.03 in L2P (e.g., length 5: SILK 0.13 vs. TST 0.10; length 15: SILK 0.38 vs. TST 0.39), so it is unclear whether those differences are meaningful. The paper should provide multiple seeds with mean and standard deviation, or a significance test, and should release code and model checkpoints to make the results reproducible.
minor comments (4)
  1. [§4.6] The text near Table 5 contains a typo: 'Table Table 5' should be 'Table 5'.
  2. [References] References [31] and [32] have the same title, 'Human motion diffusion model,' with overlapping content; please check whether one is a duplicate or should be distinguished by venue/year in the citation.
  3. [Table 1 caption] The caption states that the primary task of each model is added for easier comparison, but it would also help to state the training offset and context length for each method, since Section 4.4 shows that these choices affect the numbers.
  4. [§3.2.3 and §4.5.2] The architecture description says the output dimension is d_out, but the 'Full Vel' configuration in Table 4 requires predicting velocity features as well; please clarify how d_out changes for that variant.

Circularity Check

1 steps flagged · score 6.0 of 10

Velocity input features leak the first in-between frame and a frame past the target, so SILK's headline result partially reduces to its input by construction.

  1. self definitional [Section 4.5.2 (impact of velocity features); feature definition in Section 3.1.1; results in Table 4]
    "We calculate these velocity features (linear velocity for positions and angular velocity for rotations) using two consecutive frames - specifically, requiring one additional frame after both the start and target keyframes."

    With velocity v_f computed between consecutive frames, the start-keyframe velocity is v_s = (p_{s+1} - p_s)/dt, so the first in-between frame SILK is supposed to output is recoverable as p_{s+1} = p_s + dt*v_s. The model input contains p_s and v_s, so the first predicted frame is a direct linear transformation of the input; a linear output layer can copy it without learning the transition. Similarly the target velocity uses a frame after the target keyframe, outside the interpolation window, leaking boundary information unavailable to baselines. Table 4 quantifies the leakage: removing velocity inputs ('Static Only') raises L2P at length 30/45 from 0.83/1.59 to 1.0/2.1, worse than TST's 0.89/1.68, so SILK's Table 1 advantage is driven by ground-truth frames embedded in the input.

full rationale

The paper is otherwise an empirical benchmark comparison with no derivation chain that reduces to fitted values: offset-5 sampling is standard hyperparameter/data-augmentation selection, the zero-vs-slerp ablation is a direct model comparison, and there is no load-bearing self-citation or imported uniqueness theorem. However, the central simplicity-wins claim is partially produced by construction: the velocity input is computed from the first frame the model is supposed to predict (and from a frame beyond the target keyframe), making that portion of the output a deterministic function of the input. Because the ablated numbers show the leaked velocity input is what lifts SILK past TST on the longer horizons, the headline comparison is not established as the same task as the baselines. This is a genuine input/output identity, not merely an unfair protocol choice, so the circularity score is 6 rather than 1.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new mathematical objects or physical entities. Its contributions are empirical design choices and a simplified architecture. The main premises are the external benchmark's validity and the comparability of published numbers, plus the added requirement of extra frames for velocity computation.

free parameters (5)
  • training data offset = 5 frames
    Subsampling offset when segmenting training sequences; chosen after comparing offset 5 vs 20 (Section 4.4). Directly affects the 'more data helps' finding.
  • model hyperparameters = 6 layers, 8 heads, d_model=1024, d_ff=4096
    Architecture choices from Section 4.2; not tuned on test, but influence capacity and are hand-selected.
  • context frames = 10
    Number of context frames used during training (Section 4.3); follows prior work [9].
  • loss type = single L1 loss
    Choice of loss function without per-term weights (Section 3.2.4).
  • velocity computation window = one extra frame after keyframes
    Requirement to compute velocities; changes the problem setup by adding an input not present in the standard task (Section 4.5.2).
assumptions (4)
  • domain assumption LaFAN1 benchmark and evaluation protocol are representative and fair
    The paper relies on the protocol of [9] and the assumption that L2P, L2Q, NPSS capture motion quality (Section 4.3).
  • standard math Transformer self-attention is a standard building block
    Uses the architecture of [33] without modification (Section 3.2.1).
  • ad hoc to paper The extra frame after keyframes is available at inference
    Velocity inputs require one additional frame after start and target keyframes (Section 4.5.2). This is not part of the standard in-betweening setup and may not be available in all animation pipelines.
  • domain assumption Comparison numbers for prior methods in Table 1 are directly comparable
    The paper includes only methods with matching baseline reports and excludes others, but does not re-run all baselines (Section 4.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SILK: Smooth InterpoLation frameworK for motion in-betweening A Simplified Computational Approach." pith.science (2026). https://pith.science/paper/TXBDEUI2

@misc{pith2026250609075,
  author       = {Pith},
  title        = {Pith review of: SILK: Smooth InterpoLation frameworK for motion in-betweening A Simplified Computational Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TXBDEUI2}},
  note         = {Machine review of arXiv:2506.09075}
}
read the original abstract

Motion in-betweening is a crucial tool for animators, enabling intricate control over pose-level details in each keyframe. Recent machine learning solutions for motion in-betweening rely on complex models, incorporating skeleton-aware architectures or requiring multiple modules and training steps. In this work, we introduce a simple yet effective Transformer-based framework, employing a single Transformer encoder to synthesize realistic motions for motion in-betweening tasks. We find that data modeling choices play a significant role in improving in-betweening performance. Among others, we show that increasing data volume can yield equivalent or improved motion transitions, that the choice of pose representation is vital for achieving high-quality results, and that incorporating velocity input features enhances animation performance. These findings challenge the assumption that model complexity is the primary determinant of animation quality and provide insights into a more data-centric approach to motion interpolation. Additional videos and supplementary material are available at https://silk-paper.github.io.

Figures

Figures reproduced from arXiv: 2506.09075 by the authors.

Figure 1
Figure 1. A goalkeeping motion sequence generated by SILK. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. SILK’s architecture. The missing frames are filled with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Six in-between animations generated by SILK based on the LAFAN1 dataset. Length describes the number of in-between frames. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [5]

    Real-time diverse motion in- betweening with space-time control

    Yuchen Chu and Zeshi Yang. Real-time diverse motion in- betweening with space-time control. InProceedings of the 17th ACM SIGGRAPH Conference on Motion, Interaction, and Games, pages 1–8, 2024. 5, 6

    work page 2024

  2. [28]

    Motion in-betweening with phase manifolds.Pro- ceedings of the ACM on Computer Graphics and Interactive Techniques, 6(3):1–17, 2023

    Paul Starke, Sebastian Starke, Taku Komura, and Frank Steinicke. Motion in-betweening with phase manifolds.Pro- ceedings of the ACM on Computer Graphics and Interactive Techniques, 6(3):1–17, 2023. 5, 6, 7

    work page 2023

  3. [1]

    Skeleton- aware networks for deep motion retargeting.ACM Transac- tions on Graphics (TOG), 39(4):62–1, 2020

    Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine- Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton- aware networks for deep motion retargeting.ACM Transac- tions on Graphics (TOG), 39(4):62–1, 2020. 2

    work page 2020

  4. [2]

    Skel-betweener: a neural motion rig for interactive motion authoring.ACM Transactions on Graphics (TOG), 43(6):1–11, 2024

    Dhruv Agrawal, Jakob Buhmann, Dominik Borer, Robert W Sumner, and Martin Guay. Skel-betweener: a neural motion rig for interactive motion authoring.ACM Transactions on Graphics (TOG), 43(6):1–11, 2024. 1, 2

    work page 2024

  5. [3]

    Listen, denoise, action! audio-driven motion synthesis with diffusion models.ACM Transactions on Graphics (TOG), 42(4):1–20, 2023

    Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models.ACM Transactions on Graphics (TOG), 42(4):1–20, 2023. 8

    work page 2023

  6. [4]

    Dynamic motion transition: A hy- brid data-driven and model-driven method for human pose transitions.IEEE Transactions on Visualization and Com- puter Graphics, 2024

    Zhi Chai and Hong Qin. Dynamic motion transition: A hy- brid data-driven and model-driven method for human pose transitions.IEEE Transactions on Visualization and Com- puter Graphics, 2024. 3, 5

    work page 2024

  7. [6]

    Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805, 2018. 2

    arXiv 2018

  8. [7]

    A unified framework for real time motion completion

    Yinglin Duan, Yue Lin, Zhengxia Zou, Yi Yuan, Zhehui Qian, and Bohan Zhang. A unified framework for real time motion completion. InProceedings of the AAAI Conference on Artificial Intelligence, pages 4459–4467, 2022. 1, 2, 5

    work page 2022

Show all 40 references
  1. [8]

    Harvey and Christopher Pal

    F ´elix G. Harvey and Christopher Pal. Recurrent transition networks for character locomotion. InSIGGRAPH Asia 2018 Technical Briefs, New York, NY , USA, 2018. Association for Computing Machinery. 2

    2018

  2. [9]

    Robust motion in-betweening.ACM Trans- actions on Graphics (TOG), 39(4):60–1, 2020

    F ´elix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-betweening.ACM Trans- actions on Graphics (TOG), 39(4):60–1, 2020. 1, 2, 3, 4, 5, 6

    2020

  3. [10]

    Nemf: Neural motion fields for kinematic an- imation.Advances in Neural Information Processing Sys- tems, 35:4244–4256, 2022

    Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic an- imation.Advances in Neural Information Processing Sys- tems, 35:4244–4256, 2022. 2

    2022

  4. [11]

    Phase- functioned neural networks for character control.ACM Trans

    Daniel Holden, Taku Komura, and Jun Saito. Phase- functioned neural networks for character control.ACM Trans. Graph., 36(4):42:1–42:13, 2017. 3

    2017

  5. [12]

    Long-term motion in-betweening via keyframe prediction

    Seokhyeon Hong, Haemin Kim, Kyungmin Cho, and Jun- yong Noh. Long-term motion in-betweening via keyframe prediction. InComputer Graphics Forum, page e15171. Wi- ley Online Library, 2024. 6

    2024

  6. [13]

    Human3.6m: Large scale datasets and predic- tive methods for 3d human sensing in natural environments

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predic- tive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 36(7):1325–1339, 2014. 8

    2014

  7. [14]

    Convolutional autoen- coders for human motion infilling

    Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoen- coders for human motion infilling. In2020 International Conference on 3D Vision (3DV), pages 918–927, 2020. 1, 2, 7

    2020

  8. [15]

    Conditional motion in- betweening.Pattern Recognition, 132:108894, 2022

    Jihoon Kim, Taehyun Byun, Seungyoun Shin, Jung- dam Won, and Sungjoon Choi. Conditional motion in- betweening.Pattern Recognition, 132:108894, 2022. 1, 2, 5

    2022

  9. [16]

    Computer aided inbetweening

    Alexander Kort. Computer aided inbetweening. InPro- ceedings of the 2nd International Symposium on Non- Photorealistic Animation and Rendering, page 125–132, New York, NY , USA, 2002. Association for Computing Ma- chinery. 2

    2002

  10. [17]

    Foot-constrained spatial-temporal trans- former for keyframe-based complex motion synthesis.Com- puter Animation and Virtual Worlds, 35(1):e2217, 2024

    Hao Li, Ju Dai, Rui Zeng, Junxuan Bai, Zhangmeng Chen, and Junjun Pan. Foot-constrained spatial-temporal trans- former for keyframe-based complex motion synthesis.Com- puter Animation and Virtual Worlds, 35(1):e2217, 2024. 6

    2024

  11. [18]

    Task-generic hi- erarchical human motion prior using vaes

    Jiaman Li, Ruben Villegas, Duygu Ceylan, Jimei Yang, Zhengfei Kuang, Hao Li, and Yajie Zhao. Task-generic hi- erarchical human motion prior using vaes. In2021 Inter- national Conference on 3D Vision (3DV), pages 771–781. IEEE, 2021. 5

    2021

  12. [19]

    Towards robust direction invariance in character animation

    Li-Ke Ma, Zeshi Yang, Baining Guo, and KangKang Yin. Towards robust direction invariance in character animation. InComputer Graphics Forum, pages 235–242. Wiley Online Library, 2019. 3

    2019

  13. [20]

    On human motion prediction using recurrent neural networks

    Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2891–2900, 2017. 7

    2017

  14. [21]

    A unified masked autoencoder with patchified skeletons for motion synthesis

    Esteve Valls Mascar ´o, Hyemin Ahn, and Dongheui Lee. A unified masked autoencoder with patchified skeletons for motion synthesis. InProceedings of the AAAI Conference on Artificial Intelligence, pages 5261–5269, 2024. 1, 2, 3, 5

    2024

  15. [22]

    Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation

    Clinton A Mo, Kun Hu, Chengjiang Long, and Zhiyong Wang. Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894–13903, 2023. 5

    2023

  16. [23]

    Motion inbetweening via deep∆-interpolator.IEEE Trans- actions on Visualization and Computer Graphics, 2023

    Boris N Oreshkin, Antonios Valkanas, F ´elix G Harvey, Louis-Simon M´enard, Florent Bocquelet, and Mark J Coates. Motion inbetweening via deep∆-interpolator.IEEE Trans- actions on Visualization and Computer Graphics, 2023. 1, 2, 3, 5, 6

    2023

  17. [24]

    Deeploco: Dynamic locomotion skills us- ing hierarchical deep reinforcement learning.ACM Trans

    Xue Bin Peng, Glen Berseth, Kangkang Yin, and Michiel Van De Panne. Deeploco: Dynamic locomotion skills us- ing hierarchical deep reinforcement learning.ACM Trans. Graph., 36(4), 2017. 3

    2017

  18. [25]

    Motion in-betweening via two-stage transformers.ACM Trans

    Jia Qin, Youyi Zheng, and Kun Zhou. Motion in-betweening via two-stage transformers.ACM Trans. Graph., 41(6),

  19. [26]

    Di- verse motion in-betweening from sparse keyframes with dual posture stitching.IEEE Transactions on Visualization and Computer Graphics, 2024

    Tianxiang Ren, Jubo Yu, Shihui Guo, Ying Ma, Yutao Ouyang, Zijiao Zeng, Yazhan Zhang, and Yipeng Qin. Di- verse motion in-betweening from sparse keyframes with dual posture stitching.IEEE Transactions on Visualization and Computer Graphics, 2024. 3, 5

    2024

  20. [27]

    Transformer based motion in-betweening

    Pavithra Sridhar, Madhav Aggarwal, R Leela Velusamy, et al. Transformer based motion in-betweening. InPro- ceedings of the Asian Conference on Computer Vision, pages 289–302, 2022. 5

    2022

  21. [29]

    Anidance: real- time dance motion synthesize to the song

    Taoran Tang, Hanyang Mao, and Jia Jia. Anidance: real- time dance motion synthesize to the song. InProceedings of the 26th ACM international conference on Multimedia, pages 1237–1239, 2018. 8

    2018

  22. [30]

    Real-time controllable motion transition for characters.ACM Trans

    Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qi- long Kou, and Xiaogang Jin. Real-time controllable motion transition for characters.ACM Trans. Graph., 41(4), 2022. 2, 5, 7

    2022

  23. [31]

    Human motion diffu- sion model

    Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffu- sion model. InThe Eleventh International Conference on Learning Representations, 2022. 2

    2022

  24. [32]

    Human motion diffu- sion model

    Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffu- sion model. InThe Eleventh International Conference on Learning Representations, 2023. 8

    2023

  25. [33]

    Gomez, and Łukasz Kaiser

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N. Gomez, and Łukasz Kaiser. At- tention is all you need.arXiv preprint arXiv:1706.03762v5,

    arXiv

  26. [34]

    Motion in- betweening with spatial and temporal transformers.IEEE Transactions on Circuits and Systems for Video Technology,

    Zhiming Wang, Ning Ge, and Jianhua Lu. Motion in- betweening with spatial and temporal transformers.IEEE Transactions on Circuits and Systems for Video Technology,

  27. [35]

    Spatial- temporal graph u-net for skeleton-based human motion in- filling

    Leiyang Xu, Qiang Wang, and Chenguang Yang. Spatial- temporal graph u-net for skeleton-based human motion in- filling. In2024 IEEE International Conference on Industrial Technology (ICIT), pages 1–6. IEEE, 2024. 3, 5

    2024

  28. [36]

    Dancecraft: A music-reactive real-time dance im- prov system

    Ruilin Xu, Vu An Tran, Shree K Nayar, and Gurunandan Krishnan. Dancecraft: A music-reactive real-time dance im- prov system. InProceedings of the 9th International Confer- ence on Movement and Computing, pages 1–10, 2024. 1, 2, 3, 5

    2024

  29. [37]

    Mode-adaptive neural networks for quadruped motion con- trol.ACM Transactions on Graphics (TOG), 37(4):1–11,

    He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion con- trol.ACM Transactions on Graphics (TOG), 37(4):1–11,

  30. [38]

    Data-driven auto- completion for keyframe animation

    Xinyi Zhang and Michiel van de Panne. Data-driven auto- completion for keyframe animation. InProceedings of the 11th ACM SIGGRAPH Conference on Motion, Interaction and Games, New York, NY , USA, 2018. Association for Computing Machinery. 2

    2018

  31. [39]

    On the continuity of rotation representations in neu- ral networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neu- ral networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745– 5753, 2019. 3

    2019

  32. [40]

    Generative tweening: Long- term inbetweening of 3d human motions.arXiv preprint arXiv:2005.08891, 2020

    Yi Zhou, Jingwan Lu, Connelly Barnes, Jimei Yang, Sitao Xiang, and Hao Li. Generative tweening: Long- term inbetweening of 3d human motions.arXiv preprint arXiv:2005.08891, 2020. 1, 2

    2005 arXiv

Pith tools

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