Pith. sign in

REVIEW 4 major objections 5 minor 151 references

A 32-D latent makes a standard flow model beat VQ and autoregressive baselines for text-to-motion generation.

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

2026-08-03 12:10 UTC pith:3R5VEF4A

load-bearing objection The two-bottleneck diagnosis is the real contribution, and it's well-probed; the SOTA claim is only shown on MARDM's 67-D evaluator, which should be front and center as a limitation. the 4 major comments →

arxiv 2607.29180 v1 pith:3R5VEF4A submitted 2026-07-31 cs.CV cs.AI

MoRAE: Flow-Friendly Self-Supervised Latents for Text-to-Motion Generation

Yifei Zhu , Mingyi Shi , Yangyang Cai , Miao Cheng , Yoshifumi Kitamura , Taku Komura This is my paper
classification cs.CV cs.AI
keywords text-to-motion generationrepresentation autoencoderflow matchingself-supervised learningspectral conditioningdecoder alignmentJEPAlatent space
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.

MoRAE claims that the failure of directly transferring representation autoencoders (RAEs) to human motion is caused by two separable geometric bottlenecks: a spectrally ill-conditioned frozen self-supervised feature space, and flow residuals that align with decoder-sensitive directions. The paper diagnoses both with controlled probes, then shows that a compact 32-D variational latent distilled from frozen Motion-JEPA features and trained with motion-coupled decoding fixes both problems. With this 'flow-friendly' latent, a plain non-autoregressive Flow-Matching DiT achieves state-of-the-art results on HumanML3D and KIT-ML, improving both semantic metrics (FID, R-Precision) and physical plausibility (foot skating, jerk, bone variation) without vector quantization or autoregressive factorization. If correct, this means autoregression and discrete tokens are a workaround for bad latent geometry, not a fundamental requirement for high-quality motion generation.

Core claim

The paper establishes that a frozen Motion-JEPA encoder, when used directly as a generative latent space, fails because of two independent causes: (1) the feature covariance spectrum is extremely ill-conditioned (κ_flow ≈ 1.3×10^4), making Gaussian-to-data transport unstable; and (2) even after correcting the spectrum, flow residuals concentrate along directions where the motion decoder has high Jacobian gain, so small latent errors become large motion artifacts. Probe 1 shows that artificially imposing condition numbers above ~10^3 collapses generation FID despite unchanged reconstruction. Probe 2 shows that a coupled tokenizer and a decoupled control with nearly identical condition numbers

What carries the argument

The diagnostic core is a factorization of the expected decoded error into three scalar quantities: M_dec = M_lat · g_iso · ρ_align, where M_lat is residual magnitude, g_iso is mean decoder gain, and ρ_align is directional alignment computed from the pullback metric G(z) = J_D(z)^T J_D(z). This decomposition separates generator error from decoder sensitivity. The method itself is a coupled tokenizer: a single-attention variational encoder compresses frozen Motion-JEPA features to 32 dimensions, and a feature decoder plus a motion decoder are trained jointly so that the motion reconstruction loss backpropagates through both decoders and the encoder, shaping the latent geometry by its motion-sp

Load-bearing premise

The headline numbers are all computed with the 67-dimensional 'essential' representation and the corresponding MARDM evaluator, with baseline results taken at face value under that protocol; if the 263-dimensional HumanML3D evaluator ranks methods differently, the 'state-of-the-art' claim holds only within this specific benchmarking convention.

What would settle it

Evaluate MoRAE and the same baselines on the standard 263-D HumanML3D evaluator (as used by MDM, MLD, and T2M-GPT) and compare rankings; if MoRAE no longer leads on FID or R-Precision, the central claim is protocol-dependent. Alternatively, train a tokenizer with identical architecture but with ρ_align intentionally maximized to check whether generation FID degrades as predicted by the diagnostic.

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

If this is right

  • Autoregressive or masked-autoregressive generation is a workaround for poor latent geometry, not an intrinsic requirement: masked-AR rescues an ill-conditioned 512-D autoencoder (gFID 0.592→0.120) but hurts MoRAE's well-conditioned latent (0.089→0.221).
  • A spectrum within a transport-stable regime (κ_flow up to ~10^3) is sufficient for stable flow training; explicit whitening is not required, and over-squeezing beyond that regime causes catastrophic generation failure.
  • Coupled training improves generation quality by reducing residual alignment with decoder-sensitive directions (ρ_align), not by reducing residual magnitude or mean decoder gain, and this improvement appears in FID while R-Precision stays nearly unchanged.
  • Compact continuous latents can match or beat discrete VQ and autoregressive methods on both semantic and physical metrics, without quantization error or sequential sampling, while running as fast as masked generators (AIT 0.06 s).
  • The width of the bottleneck is the main control over implicit motion validity: 32-D is optimal, while over-compression (16-D) hurts condition detail and under-compression (768-D) yields the worst physical plausibility.

Where Pith is reading between the lines

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

  • The κ_flow threshold (~10^3) and the ρ_align diagnostic could be used as cheap pre-training checks for any continuous latent space paired with flow or diffusion, potentially predicting generation failure before training the generator at full scale.
  • If the causal account is right, representation autoencoders for other structured outputs (e.g., audio, articulated objects, or multi-agent trajectories) should exhibit the same two failure modes, and the same two probes would transfer to those domains.
  • Tokenizers could be regularized directly by penalizing ρ_align or by whitening the latent spectrum, which might make the decoder less critical and further improve generation quality.
  • The 'state-of-the-art' claim is tied to the MARDM evaluation protocol; under the older 263-D evaluator, the ranking of methods could shift, so the practical claim is that MoRAE is SOTA within that protocol.

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

4 major / 5 minor

Summary. The paper addresses text-to-motion generation in a continuous latent space. It first diagnoses why a frozen Motion-JEPA feature space is a poor target for flow matching, introducing a spectral condition number κ_flow (Eq. 1) and a factorization of decoded error M_dec = M_lat·g_iso·ρ_align (Eqs. 7–8). Two probes isolate the causal roles: an invertible spectral reparameterization (Eq. 9) varies condition number at fixed information content and decoder, and a coupled-versus-decoupled tokenizer comparison attributes the remaining generation gap to residual–decoder alignment. The proposed MoRAE distills the frozen Motion-JEPA features into a 32-D variational latent with a motion-coupled reconstruction objective (Eq. 14), then trains a standard flow-matching DiT in this latent. On HumanML3D and KIT-ML, under the MARDM 67-D essential evaluator, MoRAE reports improved FID, R-Precision, matching distance, and physical plausibility over VQ/AR and continuous-latent baselines.

Significance. The paper has real strengths. The diagnostic decomposition is transparent and is used as a measurement instrument rather than as a derivation of the improvement; the causal claims rest on interventions. Probe 1 is a well-controlled spectral manipulation with rFID held fixed, and Probe 2 includes a sensible decoupled control and validates the local Jacobian approximation (Pearson r ≥ 0.92). If the claims hold, the paper offers a useful geometric design principle for latent-space motion generation and demonstrates that a non-autoregressive flow on a compact, decoder-aligned SSL latent can match or beat discrete-token methods. The release of a project page and code is also positive. The main caveat is that the headline comparison is reported only under the non-standard 67-D essential evaluator, and several diagnostic quantities are reported without uncertainty estimates; these issues currently prevent full confidence in the SOTA claim.

major comments (4)
  1. [§5.1, Table 5] The central SOTA claim is established only under the MARDM 67-D 'essential' evaluator. No 263-D standard HumanML3D FID/R-Precision is reported for MoRAE, and the baseline numbers under the 67-D protocol are taken from MARDM rather than recomputed with a shared pipeline. Because the 67-D representation omits velocities and rotations, the two evaluators can rank methods differently; a model that wins on root state and relative joint positions may not win on the canonical kinematic features. Please report MoRAE under the standard 263-D evaluator, or provide a systematic comparison (e.g., recompute all baselines under both protocols) showing that the two metrics rank methods consistently. Without this, 'state-of-the-art' in §6 is supported only within MARDM's yardstick.
  2. [Table 3, §3.4] The central diagnostic attribution in Probe 2 rests on the difference ρ_align = 6.08 (decoupled) versus 4.85 (coupled). Table 3 reports only point estimates, with no confidence intervals or number of seeds, so the ~25% difference in directional alignment could be within run-to-run variability. The validation of the local Jacobian approximation (Pearson r ≥ 0.92, actual-to-predicted ratio 0.8–1.1) is mentioned in the text but no scatter plot, coverage analysis, or seed detail is given. Since this measurement is load-bearing for the claim that residual–decoder alignment, rather than residual magnitude or mean gain, explains the gap, the paper should provide uncertainty estimates or repeated-seed results.
  3. [§5.2, Table 5] The physical plausibility metrics (foot skating, jerk, bone variation) are computed after decoding with MoRAE's own decoder and skeleton conventions. The table lists comparable values for baselines, but the manuscript does not state how baseline motions were produced for these metrics, whether official checkpoints were used, or whether the same skeleton/retargeting conventions apply. Without this information, the physical-metric improvements could reflect decoder or normalization differences rather than generation quality. Please specify the protocol, and where possible report numbers under a shared evaluation skeleton.
  4. [§5.3, Tables 6 and 7] The ablations are described as using 'the diagnostic protocol' (Table 6) and 'the final protocol' (Table 7), but these protocols are never defined. This makes it difficult to interpret why gFID values for MoRAE differ between tables (0.089 in Table 6 and 0.089 in Table 7 is consistent, but AE-32 differs: 0.127 in Table 7 versus implied values in Table 6; the relationship is unclear) and to reproduce the ablation setup. Please define both protocols explicitly, including guidance scale, number of sampling steps, and any differences in evaluation features.
minor comments (5)
  1. [§4.1] The paper refers to 'Motion-JEPA' but cites Assran et al. 2023, which is I-JEPA for images. Please clarify whether this is a new motion-specific variant or a typo, and describe any modifications made for motion data.
  2. [Eq. (13)] The dimension of ε' is not specified. Since Vhat{H} is a feature-space quantity and x is a motion-space quantity, make explicit that ε' has the same shape as the feature representation.
  3. [Table 1] The notation 'SSL (rawH)' is confusing because H is defined later in §4.1. Consider naming it 'Motion-JEPA features' or 'frozen features'.
  4. [Table 5] The CLIP-score row for GT is listed as '—' in HumanML3D and '0.699' in KIT-ML; please clarify why the HumanML3D GT CLIP-score is missing.
  5. [Figure 2] The qualitative comparison would be stronger if the figure caption noted which baselines are shown and how representative the prompt is across multiple seeds.

Circularity Check

0 steps flagged

No significant circularity: central claims rest on controlled probes and external baselines, not on self-citation or definitional identities.

full rationale

The paper's central claim is that a compact, coupled JEPA latent makes a standard flow-matching DiT reach state-of-the-art results. That claim is supported by external comparisons (MDM, MLD, T2M-GPT, MMM, MoMask, MARDM) and by ablations, not by a self-referential derivation. The geometric decomposition in Eq. (7) is an exact identity by construction (M_dec = M_lat * g_iso * rho_align), but the paper uses it as a measurement instrument to factor measured residuals, and the causal attribution in Probes 1 and 2 is anchored by interventions (spectral reparameterization with fixed decoder; a decoupled training control), so the identity is not used to derive the improvement. There is no load-bearing self-citation: Motion-JEPA, RAE, DiT, flow-matching, and MARDM protocol are all cited from external groups, and the authors' own training procedure is described in full. The decision to evaluate on the 67-D MARDM 'essential' representation is a protocol/validity concern about which yardstick ranks methods, not a circularity: it is not fitted to produce the reported numbers, and the baseline numbers are taken from prior work rather than defined from MoRAE's outputs. Consistent with the rules, that concern belongs under correctness/benchmark risk, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced; the 'flow-friendly latent' is a design property, not a new object. The load-bearing free choices are the bottleneck width (ablated), the σ_p floor entering every reported κ_flow, the unspecified coupling weights λ/β/η, and the guidance scales. The axioms are mostly standard generative-model machinery plus two domain assumptions: the MARDM evaluation yardstick and the informativeness of frozen JEPA features.

free parameters (7)
  • bottleneck width d = 32
    Hand-chosen; Table 6(b) shows gFID minimized at z=32 (0.089) vs 0.228 (z=16) and 0.320 (z=64). The design principle ('compact bottleneck for transport stability') depends on this width.
  • perturbation floor σ_p = 0.02
    Eq. (1) defines κ_flow with an added floor σ_p=0.02; every reported condition number, including the headline κ_flow=27, depends on this chosen constant.
  • motion loss weight λ (Eq. 14) = not stated
    Governs how strongly the latent geometry is coupled to the decoder — the paper's core mechanism — yet no value or sensitivity analysis is given.
  • KL weight β (Eq. 14) = not stated
    Controls the Gaussian-prior match of the variational encoder; no value or ablation reported.
  • perturbation scale η (Eq. 13) = not stated
    Gaussian perturbation added to reconstructed features before motion decoding during tokenizer training; no value given.
  • CFG guidance scales = 8.5 (HumanML3D), 7.5 (KIT-ML)
    Inference-time classifier-free guidance scales (§5.1); chosen per dataset and directly affect all reported metrics.
  • Motion-JEPA pretraining hyperparameters (EMA momentum µ, masking ratios) = not stated
    Eq. (11) pretraining details are unspecified; the frozen features are the foundation of the entire pipeline.
axioms (5)
  • domain assumption The 67-D essential representation and its evaluator are the correct protocol for FID/R-Precision (following MARDM)
    §5.1 adopts MARDM's protocol; the headline SOTA numbers depend on this yardstick rather than the historical 263-D evaluator.
  • standard math Flow matching with linear interpolation is a valid generative model (Lipman et al. 2023; Ma et al. 2024)
    Used as the generator (Eqs. 15-16); a well-established external result.
  • standard math Local Taylor validity of the decoded-error decomposition (Eq. 4)
    O(||δ||³) truncation of the decoder Taylor expansion; the paper validates empirically (r ≥ 0.92 at observed residual scale) but the decomposition is used globally.
  • domain assumption Frozen Motion-JEPA features carry decodable semantic structure (Eq. 10)
    The whole pipeline assumes f(x) is a faithful semantic representation; supported indirectly by the two-stage w/o-SSL ablation (Table 7: 0.152 vs 0.089 gFID) but not independently verified.
  • domain assumption Coordinate standardization makes decomposition factors comparable across tokenizers
    §3.2 admits the factors are coordinate-dependent and compares only after standardization; comparability across tokenizers with different decoder architectures is assumed.

pith-pipeline@v1.3.0-daily-deepseek · 11987 in / 15431 out tokens · 140863 ms · 2026-08-03T12:10:33.736441+00:00 · methodology

0 comments
read the original abstract

Text-to-motion generation must produce motions that are semantically correct, temporally coherent, and physically plausible. A natural approach is to first project motion data into a structured semantic space and then train a generative model within that space. Such a paradigm has been highly successful in image generation through Representation Autoencoders (RAEs), where a frozen self-supervised encoder provides semantic features for diffusion or flow models to learn from. However, direct transfer of such a paradigm to motion space using Motion-JEPA as the frozen encoder fails dramatically. We diagnose this failure geometrically and identify two motion-specific bottlenecks: (1) the JEPA feature space is spectrally ill-conditioned, making the Gaussian-to-data transport unstable; and (2) even with a well-conditioned spectrum, flow residuals tend to align with decoder-sensitive directions, where small latent errors are amplified into large motion artifacts after decoding. Based on these insights, we propose MoRAE. MoRAE addresses the two bottlenecks separately. A compact bottleneck distills the structured JEPA representation while removing weak and redundant directions, bringing the latent spectrum into a transport-stable regime. Motion-coupled training then aligns the retained latent geometry with the decoder, making characteristic flow errors less costly after decoding. With this flow-friendly latent, a standard non-autoregressive Flow-Matching DiT achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2607.29180 by Miao Cheng, Mingyi Shi, Taku Komura, Yangyang Cai, Yifei Zhu, Yoshifumi Kitamura.

Figure 1
Figure 1. Figure 1: Overview of MoRAE. (a) Motion-JEPA predicts EMA-teacher features under temporal-span and joint-group masks. (b) Frozen features are compressed into a 32-D latent and jointly decoded back to features and motion. (c) After freezing the tokenizer, a non-autoregressive DiT learns flow matching in the compact latent space, and the frozen decoder maps generated latents back to motion. text encoder (Radford et al… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison on a compositional text prompt. [PITH_FULL_IMAGE:figures/full_fig_p007_2.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

151 extracted references · 43 linked inside Pith

  1. [1]

    Generating Diverse and Natural

    Guo, Chuan and Zou, Shihao and Zuo, Xinxin and Wang, Sen and Ji, Wei and Li, Xingyu and Cheng, Li , booktitle =. Generating Diverse and Natural

  2. [2]

    arXiv preprint arXiv:2411.16575 , year =

    Rethinking Diffusion for Text-Driven Human Motion Generation: Redundant Representations, Evaluation, and Masked Autoregression , author =. arXiv preprint arXiv:2411.16575 , year =

    Pith/arXiv arXiv

  3. [3]

    International Conference on Learning Representations (ICLR) , year =

    Human Motion Diffusion Model , author =. International Conference on Learning Representations (ICLR) , year =

  4. [4]

    Zhang, Mingyuan and Cai, Zhongang and Pan, Liang and Hong, Fangzhou and Guo, Xinying and Yang, Lei and Liu, Ziwei , journal =

  5. [5]

    Zhang, Mingyuan and Guo, Xinying and Pan, Liang and Cai, Zhongang and Hong, Fangzhou and Li, Huirong and Yang, Lei and Liu, Ziwei , booktitle =

  6. [6]

    Zhang, Jianrong and Zhang, Yangsong and Cun, Xiaodong and Huang, Shaoli and Zhang, Yong and Zhao, Hongwei and Lu, Hongtao and Shen, Xi , booktitle =

  7. [7]

    Guo, Chuan and Mu, Yuxuan and Javed, Muhammad Gohar and Wang, Sen and Cheng, Li , booktitle =

  8. [8]

    Pinyoanuntapong, Ekkasit and Wang, Pu and Lee, Minwoo and Chen, Chen , booktitle =

  9. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Executing Your Commands via Motion Diffusion in Latent Space , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  10. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    High-Resolution Image Synthesis with Latent Diffusion Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  12. [12]

    International Conference on Machine Learning (ICML) , year =

    Learning Transferable Visual Models from Natural Language Supervision , author =. International Conference on Machine Learning (ICML) , year =

  13. [13]

    International Conference on Learning Representations (ICLR) , year =

    Flow Matching for Generative Modeling , author =. International Conference on Learning Representations (ICLR) , year =

  14. [14]

    International Conference on Learning Representations (ICLR) , year =

    Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think , author =. International Conference on Learning Representations (ICLR) , year =

  15. [15]

    and Boffi, Nicholas M

    Ma, Nanye and Goldstein, Mark and Albergo, Michael S. and Boffi, Nicholas M. and Vanden-Eijnden, Eric and Xie, Saining , booktitle =

  16. [16]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Scalable Diffusion Models with Transformers , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  17. [17]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Autoregressive Image Generation without Vector Quantization , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  18. [18]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Neural Discrete Representation Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  19. [19]

    Transactions on Machine Learning Research (TMLR) , year =

    Convergence of Denoising Diffusion Models under the Manifold Hypothesis , author =. Transactions on Machine Learning Research (TMLR) , year =

  20. [20]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Score-Based Generative Models Detect Manifolds , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  21. [21]

    Big data , volume=

    The KIT motion-language dataset , author=. Big data , volume=. 2016 , publisher=

    2016

  22. [22]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Generating diverse and natural 3d human motions from text , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  23. [23]

    Biometrika , volume=

    Learning internal representations by error propagation , author=. Biometrika , volume=

  24. [24]

    arXiv preprint arXiv:1312.6114 , year=

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    Pith/arXiv arXiv

  25. [25]

    International Journal of Computer Vision , pages=

    Intergen: Diffusion-based multi-human motion generation under complex interactions , author=. International Journal of Computer Vision , pages=. 2024 , publisher=

    2024

  26. [26]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Flame: Free-form language-based motion synthesis & editing , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  27. [27]

    European Conference on Computer Vision , pages=

    Motionclip: Exposing human motion generation to clip space , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  28. [28]

    G lo V e: Global Vectors for Word Representation

    Pennington, Jeffrey and Socher, Richard and Manning, Christopher. G lo V e: Global Vectors for Word Representation. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing ( EMNLP ). 2014. doi:10.3115/v1/D14-1162

    doi:10.3115/v1/d14-1162 2014

  29. [29]

    arXiv preprint arXiv:1412.3555 , year=

    Empirical evaluation of gated recurrent neural networks on sequence modeling , author=. arXiv preprint arXiv:1412.3555 , year=

    Pith/arXiv arXiv

  30. [30]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  31. [31]

    Advances in Neural Information Processing Systems , year=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , year=

  32. [32]

    Advances in neural information processing systems , volume=

    Neural discrete representation learning , author=. Advances in neural information processing systems , volume=

  33. [33]

    Carnegie Mellon University - CMU Graphics Lab - motion capture library , year=

    Carnegie Mellon University - CMU Graphics Lab - motion capture library , title=. Carnegie Mellon University - CMU Graphics Lab - motion capture library , year=

  34. [34]

    Advances in neural information processing systems , year=

    Generating animated videos of human activities from natural language descriptions , author=. Advances in neural information processing systems , year=

  35. [35]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    AMASS: Archive of motion capture as surface shapes , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  36. [36]

    North American Chapter of the Association for Computational Linguistics , year=

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author=. North American Chapter of the Association for Computational Linguistics , year=

  37. [37]

    2019 International Conference on 3D Vision (3DV) , pages=

    Language2pose: Natural language grounded pose forecasting , author=. 2019 International Conference on 3D Vision (3DV) , pages=. 2019 , organization=

    2019

  38. [38]

    Proceedings of the 28th ACM International Conference on Multimedia , pages=

    Action2motion: Conditioned generation of 3d human motions , author=. Proceedings of the 28th ACM International Conference on Multimedia , pages=

  39. [39]

    arXiv preprint arXiv:2005.14165 , year=

    Language models are few-shot learners , author=. arXiv preprint arXiv:2005.14165 , year=

    Pith/arXiv arXiv 2005

  40. [40]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  41. [41]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  42. [42]

    arXiv preprint arXiv:2104.08718 , year=

    Clipscore: A reference-free evaluation metric for image captioning , author=. arXiv preprint arXiv:2104.08718 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2107.06252 , year=

    Dance2music: Automatic dance-driven music generation , author=. arXiv preprint arXiv:2107.06252 , year=

    Pith/arXiv arXiv

  44. [44]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  45. [45]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  46. [46]

    Advances in neural information processing systems , volume=

    Photorealistic text-to-image diffusion models with deep language understanding , author=. Advances in neural information processing systems , volume=

  47. [47]

    The Eleventh International Conference on Learning Representations , year=

    Human Motion Diffusion Model , author=. The Eleventh International Conference on Learning Representations , year=

  48. [48]

    arXiv preprint arXiv:2208.15001 , year=

    MotionDiffuse: Text-Driven Human Motion Generation with Diffusion Model , author=. arXiv preprint arXiv:2208.15001 , year=

    Pith/arXiv arXiv

  49. [49]

    Journal of Computing in Civil Engineering , volume=

    Adaln: a vision transformer for multidomain learning and predisaster building information extraction from images , author=. Journal of Computing in Civil Engineering , volume=. 2022 , publisher=

    2022

  50. [50]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  51. [51]

    arXiv preprint arXiv:2301.06052 , year=

    T2m-gpt: Generating human motion from textual descriptions with discrete representations , author=. arXiv preprint arXiv:2301.06052 , year=

    Pith/arXiv arXiv

  52. [52]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Executing your commands via motion diffusion in latent space , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  53. [53]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Remodiffuse: Retrieval-augmented motion diffusion model , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  54. [54]

    ACM SIGGRAPH 2024 Conference Papers , pages=

    Tedi: Temporally-entangled diffusion for long-term motion synthesis , author=. ACM SIGGRAPH 2024 Conference Papers , pages=

    2024

  55. [55]

    ACM SIGGRAPH 2024 Conference Papers , pages=

    Taming diffusion probabilistic models for character control , author=. ACM SIGGRAPH 2024 Conference Papers , pages=

    2024

  56. [56]

    ACM Transactions on Graphics (TOG) , volume=

    Interactive character control with auto-regressive motion diffusion models , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  57. [57]

    arXiv preprint arXiv:2410.08159 , year=

    Dart: Denoising autoregressive transformer for scalable text-to-image generation , author=. arXiv preprint arXiv:2410.08159 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2410.03441 , year=

    CLoSD: Closing the Loop between Simulation and Diffusion for multi-task character control , author=. arXiv preprint arXiv:2410.03441 , year=

    Pith/arXiv arXiv

  59. [59]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Momask: Generative masked modeling of 3d human motions , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  60. [60]

    arXiv preprint arXiv:2401.08740 , year=

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. arXiv preprint arXiv:2401.08740 , year=

    Pith/arXiv arXiv

  61. [61]

    European Conference on Computer Vision , pages=

    Motion Mamba: Efficient and Long Sequence Motion Generation , author=. European Conference on Computer Vision , pages=. 2025 , organization=

    2025

  62. [62]

    arXiv preprint arXiv:2403.19435 , year=

    BAMM: Bidirectional Autoregressive Motion Model , author=. arXiv preprint arXiv:2403.19435 , year=

    Pith/arXiv arXiv

  63. [63]

    arXiv preprint arXiv:2406.11838 , year=

    Autoregressive Image Generation without Vector Quantization , author=. arXiv preprint arXiv:2406.11838 , year=

    Pith/arXiv arXiv

  64. [64]

    European Conference on Computer Vision , pages=

    Givt: Generative infinite-vocabulary transformers , author=. European Conference on Computer Vision , pages=. 2025 , organization=

    2025

  65. [65]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mmm: Generative masked motion model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  66. [66]

    European Conference on Computer Vision , pages=

    Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  67. [67]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Motiongpt: Finetuned llms are general-purpose motion generators , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  68. [68]

    Advances in Neural Information Processing Systems , volume=

    Motiongpt: Human motion as a foreign language , author=. Advances in Neural Information Processing Systems , volume=

  69. [69]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Attt2m: Text-driven human motion generation with multi-perspective attention mechanism , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  70. [70]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Maskgit: Masked generative image transformer , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  71. [71]

    arXiv preprint arXiv:2010.02502 , year=

    Denoising diffusion implicit models , author=. arXiv preprint arXiv:2010.02502 , year=

    Pith/arXiv arXiv 2010

  72. [72]

    arXiv preprint arXiv:2404.19759 , year=

    Motionlcm: Real-time controllable motion generation via latent consistency model , author=. arXiv preprint arXiv:2404.19759 , year=

    Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2312.00752 , year=

    Mamba: Linear-time sequence modeling with selective state spaces , author=. arXiv preprint arXiv:2312.00752 , year=

    Pith/arXiv arXiv

  74. [74]

    The Twelfth International Conference on Learning Representations , year=

    OmniControl: Control Any Joint at Any Time for Human Motion Generation , author=. The Twelfth International Conference on Learning Representations , year=

  75. [75]

    arXiv preprint arXiv:2312.06553 , year=

    Hoi-diff: Text-driven synthesis of 3d human-object interactions using diffusion models , author=. arXiv preprint arXiv:2312.06553 , year=

    Pith/arXiv arXiv

  76. [76]

    European Conference on Computer Vision , pages=

    Smoodi: Stylized motion diffusion model , author=. European Conference on Computer Vision , pages=. 2025 , organization=

    2025

  77. [77]

    3DV , year=

    Language2pose: Natural language grounded pose forecasting , author=. 3DV , year=

  78. [78]

    ECCV , year=

    TEMOS: Generating diverse human motions from textual descriptions , author=. ECCV , year=

  79. [79]

    and Varol, G

    Petrovich, Mathis and Black, Michael J. and Varol, G. ICCV , year =

  80. [80]

    Proceedings of the 5th ACM International Conference on Multimedia in Asia , pages=

    Cross-Modal Retrieval for Motion and Text via DropTriple Loss , author=. Proceedings of the 5th ACM International Conference on Multimedia in Asia , pages=

Showing first 80 references.