Pith. sign in

REVIEW 4 major objections 5 minor 84 references

ClusterStyle claims that each motion style contains multiple distinct sub-style patterns, and modeling them as non-learnable cluster prototypes yields more diverse, content-faithful stylized motion generation than a single style embedding.

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 ↗

ClusterStyle clusters each motion style into global and local prototypes and conditions a latent diffusion model on them, improving stylized motion generation fidelity and enabling controllable within-style diversity on 100STYLE/HumanML3D.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection ClusterStyle is a well-executed system with credible SOTA numbers, but its headline claim about interpretable intra-style diversity is not supported by the evidence. the 4 major comments →

arxiv 2512.02453 v2 pith:D5KMLODZ submitted 2025-12-02 cs.CV

ClusterStyle: Modeling Intra-Style Diversity with Prototypical Clustering for Stylized Motion Generation

Kerui Chen , Jianrong Zhang , Ming Li , Zhonglong Zheng , Hehe Fan This is my paper
classification cs.CV
keywords stylized motion generationintra-style diversityprototype clusteringmotion style transferlatent diffusion modelcontrastive learningoptimal transport assignmentstyle modulation adapter
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.

The reading

The paper sets out to establish that intra-style diversity — the fact that one style, such as an 'aeroplane' walk, can be expressed in many different ways — is a structure worth modeling explicitly rather than a nuisance. It replaces the single unstructured style embedding used by prior diffusion-based stylization methods with a set of cluster prototypes per style, learned at both the whole-sequence and temporal-segment level. The authors argue this yields two gains: generated motions vary within a style when different prototypes are chosen, and the style representation is disentangled from the content of the style reference clip, so text prompts like 'a person is walking forward' are not overridden. They support this with state-of-the-art scores on content alignment (R-Precision), style recognition accuracy, and motion fidelity (FID) on the standard stylized-motion benchmark.

Core claim

On its own terms, the paper's central claim is that a motion style is better represented by a set of K non-learnable prototypes (cluster centroids) than by a single learned embedding. The prototypes are computed by clustering style-motion features through an optimal-transport assignment with a balancing constraint that forces roughly equal use of every prototype, and are updated by exponential moving average during training. Two contrastive losses shape the space: one pulls each motion feature toward its own category's prototypes and away from other categories, while the other treats the assigned prototype as positive and all other prototypes, even within the same style, as negatives. The sa

What carries the argument

The central object is the prototype set P_s = {p_i_s}: a small collection of non-learnable cluster centroids assigned to each style category. These centroids carry the argument by (1) giving each style a multi-modal, interpretable structure; (2) being formed from style-motion features via optimal-transport assignment with a balanced-usage constraint, exponential-moving-average updates, and two prototype-based contrastive losses; (3) being organized hierarchically into global (whole-sequence) and local (temporal-segment) prototypes; and (4) serving at inference as guidance anchors that, combined through a Style Modulation Adapter (which blends text cross-attention and style cross-attention),

Load-bearing premise

The method assumes that every style category really contains K_g (here 3) comparably sized and distinct sub-style modes, because the optimal-transport assignment is constrained to use each prototype roughly equally; for a style that is actually homogeneous, that constraint may fabricate pseudo-diverse prototypes rather than discover real variation.

What would settle it

Measure, per style, the actual cluster structure of style-motion features (for instance, silhouette score or the mutual information between prototype assignment and hand-labeled variation in amplitude/tempo). If a style known to be unimodal still yields K_g well-separated prototypes that produce visibly different generated motions, the claimed diversity is likely an artifact of the balanced-assignment constraint rather than discovered sub-styles. A second check: retrain without the balanced-usage constraint and see whether style recognition accuracy and diversity drop; if they do not, the cons

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

If this is right

  • If the central claim is correct, a single style can be rendered as multiple visibly different motions at inference time simply by choosing a different prototype, making style control interpretable rather than an opaque embedding.
  • Content preservation improves because the style branch is anchored to style-only centroids, so the semantic content of the text (e.g., walking forward) is not corrupted by whatever content happens to be in the style reference clip.
  • The hierarchical global-plus-local prototype scheme carries over to motion style transfer, where it yields lower foot-skating ratios and better motion fidelity than previous approaches.
  • The architecture attaches to a pretrained text-to-motion diffusion model with only a small number of additional trainable parameters, so the prototype encoder and the style modulation adapter can be reused across different text-to-motion backbones.
  • Ablation results in the paper show that both the clustering procedure and the combination of inter- and intra-style contrastive losses are necessary; replacing them with a classification head or a single loss degrades all evaluation metrics.

Where Pith is reading between the lines

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

  • We infer that the balanced-usage constraint is the most fragile component: for a style that is genuinely unimodal, forcing K prototypes to be used equally may manufacture artificial sub-styles, so an important test is to compare against an unconstrained assignment and to measure per-style prototype separation.
  • Because the prototypes are non-learnable centroids of style features, the same machinery could transfer to other generative domains that need intra-class variation, such as facial expression or gesture generation; this is an extension the paper does not explore.
  • The paper's qualitative evidence of sub-style semantics rests on a single aeroplane example; a stronger test of interpretability would be a user study or automatic retrieval showing that different prototypes correspond to perceptually distinct variants across many styles.
Share X Bluesky LinkedIn Reddit HN

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. ClusterStyle proposes a clustering-based framework for stylized motion generation and motion style transfer. Instead of encoding each style motion into a single unstructured embedding, the method represents each style category by multiple non-learnable prototypes at two levels: global prototypes for whole-sequence style and local prototypes for temporal segments. Features are assigned to prototypes via a Sinkhorn-regularized optimal transport problem, prototypes are updated by EMA, and two contrastive losses (inter-style and intra-style) shape the style embedding space. A Style Modulation Adapter (SMA) injects the resulting style features into a pretrained text-to-motion latent diffusion model. The paper reports state-of-the-art results on 100STYLE and HumanML3D for both stylized generation (e.g., R-Precision 0.708, FID 1.137 in Table 1) and style transfer (FID 0.768, SRA 73.849 in Table 2), with ablations over the prototype counts, loss components, guidance strength, and SMA hyperparameters.

Significance. If the claims hold, the paper offers a concrete way to move from unstructured style embeddings to a structured, partially interpretable representation of intra-style variation, and it shows measurable improvements in content preservation and style fidelity over SMooDi, BiFlow, and StyleMotif. The work is well positioned within an active line of research and ships a reproducible-looking pipeline with component-level ablations (Tables 3–8) and a modest training cost. The main advertised novelty, however, is the claim that the learned prototypes correspond to meaningful sub-styles that can be selected at inference to produce diverse but style-consistent motions. That interpretability claim is currently supported only by one qualitative aeroplane example (Fig. 5), and the training objective contains a suspicious sign in Eq. (5) that, if not fixed, would invert the intended inter-style contrastive learning. The aggregate SOTA numbers are encouraging but are reported as single-seed point estimates.

major comments (4)
  1. [§3.2, Eq. (5)] The inter-style loss is written as -log exp(-sim[f_g, s]) / Σ_{s'} exp(-sim[f_g, s']), where sim[f_g, s] is defined as min_k cos(f_g, p_k^s), a cosine similarity. With this sign, a feature close to its own style prototype gives a small numerator, so minimizing the loss pushes the feature away from the correct category. This is the opposite of the stated goal of pulling features toward the prototypes of their own style. Please correct the sign or clarify if sim is intended to be a distance, and re-run the affected experiments (Tables 1, 3, 4, and the SRA numbers). Since L_inter is one of two core style losses, this issue is load-bearing.
  2. [§3.2, Eq. (4) and §4.4/Fig. 5] The central claim that prototypes model intra-style diversity requires evidence that the prototypes correspond to stable, perceptually distinct sub-styles rather than arbitrary partitions induced by the Sinkhorn balancing constraint. Eq. (4) enforces L_s 1_Ns = (N_s/K_g) 1_Kg, i.e., every prototype must be selected approximately N_s/K_g times. For a style category that is unimodal or unevenly clustered, this constraint artificially splits the feature distribution into equal-mass pseudo-clusters, and the EMA centroids in Eq. (8) inherit that arbitrary partition. The paper currently provides no cluster-validity analysis (e.g., silhouette scores, within/between-prototype variance), no seed-stability or permutation tests, and only one qualitative aeroplane example. Please add quantitative evidence that the discovered prototypes are stable and correspond to distinct sub-style modes, or temper
  3. [Tables 1–3 and §4.2] The headline improvements, e.g., the 15.5% relative R-Precision gain over BiFlow in Table 1, are based on single-run point estimates with no confidence intervals or multiple-seed evaluation. This is especially important because the claimed differences in the ablations are modest (e.g., FID 1.137 vs 1.170 for K_g=3 vs 5, R-Precision 0.708 vs 0.701 in Table 3). Without variance estimates, the reader cannot judge whether the reported margins are significant. Please report means and standard deviations over at least three seeds, or a paired significance test for the main comparisons.
  4. [§3.4, Eq. (11) and Appendix F] The paper states in the main text that 'λ is a trainable parameter' but the final model (Appendix F, Table 6) uses the learnable variant. The ablation in Table 6 compares fixed λ and learnable λ, which is appropriate, but the main text should be explicit that the default is the learnable version. The differences between λ=1.0 and learnable are small (SRA 77.213 vs 78.101); please specify whether the reported Table 1/2 results use the learnable λ and, if so, provide the learned value or its range for reproducibility.
minor comments (5)
  1. [§3.2, Eq. (4)] The constraint 'L_s ∈ R^{K_g×N_s}_+ ensuring that each feature is assigned once' is not enforced by the written row-sum condition alone; a column-sum condition 1_{K_g}^T L_s = 1_{N_s}^T is needed. If this condition is implied by the Sinkhorn procedure, please state it explicitly.
  2. [§3.2, Eq. (9)] The assignment 'ϵθ(zt, t, c, s) = ϵθ(zt, t, c, s) + γ_g ∇zt G_g(...)' uses the same symbol on both sides; the left-hand side should be a new, guided noise prediction (e.g., ε̃_θ).
  3. [§3.2, Eq. (5)] The use of min{cos(...)} for sim[f_g, s] means the inter-style loss is driven by the closest prototype in each category. This is reasonable, but it is not discussed why the min over prototypes is preferred over, e.g., the max or the average; a brief justification would help.
  4. [Figure 1, Figure 3, Table 1 caption] 'Areoplane' is a typo for 'Aeroplane' in multiple places. In Table 1, the header 'R-Precision → SRA ↑ (Top-3) ↑' is confusingly formatted; the arrows and the parenthetical should be cleaned up.
  5. [§3.5 and §4.1] The paper alternates between 'classifier-based guidance' and 'classifier guidance'; please use consistent terminology. Also, the classifier-based guidance details appear only in the appendix, but the main-text Section 3.5 says 'Please find more information corresponding to the inference in the Appendix'; this is acceptable but should be cross-referenced more explicitly, e.g., 'Appendix A'.

Circularity Check

1 steps flagged

Fig. 5 'intra-style diversity' is generated by optimizing toward the same fitted prototypes it is offered as evidence for.

specific steps
  1. fitted input called prediction [Sec. 3.2 (Eqs. 4, 6, 8), Sec. 3.2/Appendix B (Eqs. 9, 16), Sec. 4.4]
    "To prevent all style motion features from being assigned to a single prototype, we introduce a balancing constraint ... L_s 1_Ns = Ns/Kg 1_Kg, encouraging balanced usage of prototypes. ... our prototypes are non-parametric and non-learnable statistics computed as the centroids of their assigned style-feature sets. ... Gp(zt, t, pks) = 1−cos(f_x0_g, pks) ... We find that both different global prototypes and permutations of local prototypes yield stylistic results that vary in the extent of spread arms."

    Eq. 8 defines each prototype as the EMA mean of the very features assigned to it by Eq. 4's balanced Sinkhorn assignment, and Eq. 6 contrastively pulls features toward their assigned prototype. Eq. 16 then 'guides' generation by gradient ascent on 1−cos(f_x0_g, pks), so the generated motion is optimized to match whichever prototype is supplied. Different prototypes therefore produce different outputs by construction; the Fig. 5 'diversity' is the objective itself, not evidence that prototypes correspond to real sub-styles. The equal-usage constraint L_s 1_Ns = Ns/Kg 1_Kg can split a unimodal style into K_g equal-mass pseudo-clusters, so the claimed discovery of intra-style modes reduces to the construction rather than an independent finding.

full rationale

The quantitative claims in Tables 1 and 2 are evaluated with external metrics (R-Precision, MM-Dist, FID, SRA, FSR) against held-out content/style data, so the core SOTA generation result is not circular. There is no load-bearing self-citation: the co-authored [75] is used only for a pretrained VAE implementation detail. The circularity is confined to the paper's interpretability/diversity claim: the demonstration that different prototypes produce different stylized motions is a direct consequence of the prototype-guidance objective (Eq. 16) applied to prototypes that are themselves EMA centroids of the same feature space (Eq. 8). No stability, permutation, or within-cluster-coherence analysis is provided to show the prototypes are semantically distinct modes; the only qualitative evidence is the aeroplane example in Fig. 5. Thus the 'diversity' evidence reduces by construction, while the external benchmark performance remains independent.

Axiom & Free-Parameter Ledger

13 free parameters · 6 axioms · 2 invented entities

The framework rests on a moderate number of hand-set hyperparameters (K_g, K_l, gamma, lambda_style, w_c/w_s, tau, beta, lambda_p) and on domain assumptions about the meaningfulness of balanced prototype clusters, content-style disentanglement, and the validity of an unstated style feature extractor used both for guidance and evaluation. No new physical entities are introduced; the prototypes and SMA are representational/architectural constructs whose evidence is internal to the paper. Several parameters that the central claims depend on (mu, w, gamma_g, prototype initialization) are not reported at all.

free parameters (13)
  • K_g (global prototypes per style) = 3
    Ablated in Table 3 (1, 3, 5, 10); 3 chosen as the best trade-off, with the paper noting content alignment degrades as K_g grows while SRA improves.
  • K_l (local prototypes) = 30
    Ablated in Table 3 (1, 5, 10, 30, 50); 30 chosen for comparable performance to 50 at lower cost.
  • gamma (classifier-based guidance weight) = 3.0 (applied only when t<300)
    Ablated in Appendix Table 8 (2, 2.5, 3.0, 3.5, 4.0); FID and R-Precision degrade monotonically as gamma grows while SRA improves; 3.0 chosen as a trade-off.
  • lambda_style (style loss weight) = 1
    Ablated in Appendix Table 5 (0.5, 1, 2); 1 balances FID/R-Precision against SRA.
  • w_c, w_s (dual classifier-free guidance weights) = 7.5, 1
    Hand-set following SMooDi practice (Appendix A.1); not ablated in the paper.
  • tau (intra-style contrastive temperature) = 0.05
    Taken from SimCLR [8] (Section 3.2); not ablated.
  • beta (intra-category negative prototype weight) = 5
    Hand-set in Eq. 6 to up-weight negatives within the same style category; no ablation shown.
  • lambda_p (prototype EMA momentum) = 0.95
    Hand-set in Section 3.5; not ablated.
  • mu (Sinkhorn entropy regularizer) = not reported
    Controls assignment smoothness in Eq. 4; value never specified in the paper.
  • lambda (SMA balance) = learnable, initial value unstated
    Ablated in Appendix Table 6: fixed 0.5 and 1.0 vs learnable; learnable preferred.
  • local segment window w = not reported
    Defines local temporal segments in Section 3.3 (L_s = L_w x w); value not given.
  • gamma_g (prototype guidance weight) = not reported
    Appears in Eq. 9; never given a value or an ablation.
  • style-transfer inference settings = w_s=2.25, T'=500, gamma=2.5
    Hand-set in Appendix C for the SDEdit-based transfer task.
axioms (6)
  • domain assumption The style encoder's feature distribution, balanced by Sinkhorn's uniform-usage constraint (Eq. 4), contains K_g genuinely distinct, comparably sized sub-style modes per category.
    If a style is monomodal or uneven, the balancing constraint fabricates pseudo-diverse centroids and the 'diversity' results are artifacts of the constraint. This is the load-bearing premise for the intra-style diversity claim.
  • domain assumption Style motions in 100STYLE carry separable content and style components, and the prototype contrastive losses suffice to extract style without content leakage.
    The disentanglement is never measured directly; it is inferred from downstream R-Precision and SRA.
  • domain assumption The pretrained latent diffusion prior (MLD/EnergyMoGen, trained on HumanML3D) transfers to 100STYLE stylization when augmented with SMA.
    Standard in the SMooDi/BiFlow/StyleMotif line (refs 25, 36, 81); reused here without re-validation.
  • domain assumption The style feature extractor f used for classifier-based guidance (Appendix A.2) and the SRA evaluator are valid, consistent style measures.
    Their architecture, training data, and mutual relationship are not reported; both are trained on the same 100STYLE labels, creating a possible guidance-evaluation loop.
  • ad hoc to paper EMA prototype updates with momentum 0.95, frozen after 1800 iterations, remain aligned with the still-training style encoder.
    The freeze schedule (1800 of ~3500 iterations) is a paper-specific choice with no ablation; if the encoder drifts after freezing, the prototype anchors become stale.
  • standard math Gaussian diffusion noise model and DDIM sampling with 50 steps adequately approximate the reverse process for stylized generation.
    Standard diffusion assumptions following MLD [9].
invented entities (2)
  • Global and local style prototypes (cluster centroids) no independent evidence
    purpose: Represent sub-style modes per style category; condition generation through SMA and serve as guidance targets at inference (Eqs. 8, 9, 16, 17).
    A learned representational construct, not a physical entity. The only evidence that prototypes correspond to meaningful sub-styles is the paper's own qualitative Fig 5 and internal ablations; no externally testable prediction (e.g., a kinematic signature on held-out data) is provided.
  • Stylistic Modulation Adapter (SMA) no independent evidence
    purpose: Fuses style features into the pretrained latent diffusion model via dual cross-attention with a trainable balance lambda (Eq. 11).
    An architectural component. Its effect is measurable only through the paper's own ablations (Appendix Table 6); no external benchmark or prediction independent of the paper's pipeline is offered.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ClusterStyle: Modeling Intra-Style Diversity with Prototypical Clustering for Stylized Motion Generation." pith.science (2026). https://pith.science/paper/D5KMLODZ

@misc{pith2026251202453,
  author       = {Pith},
  title        = {Pith review of: ClusterStyle: Modeling Intra-Style Diversity with Prototypical Clustering for Stylized Motion Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D5KMLODZ}},
  note         = {Machine review of arXiv:2512.02453}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing stylized motion generation models have shown their remarkable ability to understand specific style information from the style motion, and insert it into the content motion. However, capturing intra-style diversity, where a single style should correspond to diverse motion variations, remains a significant challenge. In this paper, we propose a clustering-based framework, ClusterStyle, to address this limitation. Instead of learning an unstructured embedding from each style motion, we leverage a set of prototypes to effectively model diverse style patterns across motions belonging to the same style category. We consider two types of style diversity: global-level diversity among style motions of the same category, and local-level diversity within the temporal dynamics of motion sequences. These components jointly shape two structured style embedding spaces, i.e., global and local, optimized via alignment with non-learnable prototype anchors. Furthermore, we augment the pretrained text-to-motion generation model with the Stylistic Modulation Adapter (SMA) to integrate the style features. Extensive experiments demonstrate that our approach outperforms existing state-of-the-art models in stylized motion generation and motion style transfer.

Figures

Figures reproduced from arXiv: 2512.02453 by Hehe Fan, Jianrong Zhang, Kerui Chen, Ming Li, Zhonglong Zheng.

Figure 1
Figure 1. Figure 1: Comparison between SMooDi [81] and ClusterStyle. Given two style motions with the same style but differing in ex￾pression, along with a content text, (a) SMooDi generates similar stylized motions. Meanwhile, the generated motions still reflect elements from the style motions’ content, leading to inconsistency with the intended content text. (b) In contrast, using different prototypes corresponding to the a… view at source ↗
Figure 2
Figure 2. Figure 2: The overview of the ClusterStyle. Our method consists of a style encoder and a motion latent diffusion model. In the style encoder, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results of stylized motion generation. We compare our method with SMooDi under various text prompts and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of motion style transfer. Our approach [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of prototype guiding. We visualize how [PITH_FULL_IMAGE:figures/full_fig_p008_5.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

84 extracted references · 11 linked inside Pith

  1. [1]

    Unpaired motion style transfer from video to animation.ACM Transactions on Graphics (TOG), 39(4):64–1, 2020

    Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen- Or, and Baoquan Chen. Unpaired motion style transfer from video to animation.ACM Transactions on Graphics (TOG), 39(4):64–1, 2020. 1, 2, 6, 7

    2020

  2. [2]

    Lan- guage2pose: Natural language grounded pose forecasting

    Chaitanya Ahuja and Louis-Philippe Morency. Lan- guage2pose: Natural language grounded pose forecasting. In2019 International conference on 3D vision (3DV), pages 719–728. IEEE, 2019. 2

    2019

  3. [3]

    Coolmoves: User motion ac- centuation in virtual reality.Proceedings of the ACM on In- teractive, Mobile, Wearable and Ubiquitous Technologies, 5 (2):1–23, 2021

    Karan Ahuja, Eyal Ofek, Mar Gonzalez-Franco, Christian Holz, and Andrew D Wilson. Coolmoves: User motion ac- centuation in virtual reality.Proceedings of the ACM on In- teractive, Mobile, Wearable and Ubiquitous Technologies, 5 (2):1–23, 2021. 1

    2021

  4. [4]

    Deep learning-based motion style transfer tools, techniques and fu- ture challenges.Sensors, 23(5):2597, 2023

    Syed Muhammad Abrar Akber, Sadia Nishat Kazmi, Syed Muhammad Mohsin, and Agnieszka Szcz˛ esna. Deep learning-based motion style transfer tools, techniques and fu- ture challenges.Sensors, 23(5):2597, 2023. 1

    2023

  5. [5]

    Deep clustering for unsupervised learning of visual features

    Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. InProceedings of the European confer- ence on computer vision (ECCV), pages 132–149, 2018. 3

    2018

  6. [6]

    Unsupervised learning of visual features by contrasting cluster assignments.Ad- vances in neural information processing systems, 33:9912– 9924, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Pi- otr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.Ad- vances in neural information processing systems, 33:9912– 9924, 2020. 3

    2020

  7. [7]

    Neural clustering based visual representation learning

    Guikun Chen, Xia Li, Yi Yang, and Wenguan Wang. Neural clustering based visual representation learning. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5714–5725, 2024. 3

    2024

  8. [8]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on ma- chine learning, pages 1597–1607. PmLR, 2020. 5

    2020

  9. [9]

    Executing your commands via motion diffusion in latent space

    Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18000–18010, 2023. 1, 2, 4

    2023

  10. [10]

    Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information pro- cessing systems, 26, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information pro- cessing systems, 26, 2013. 4

    2013

  11. [11]

    Mofusion: A framework for denoising-diffusion-based motion synthesis

    Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9760–9770, 2023. 2

    2023

  12. [12]

    Decou- pling with entropy-based equalization for semi-supervised semantic segmentation

    Chuanghao Ding, Jianrong Zhang, Henghui Ding, Hongwei Zhao, Zhihui Wang, Tengfei Xing, and Runbo Hu. Decou- pling with entropy-based equalization for semi-supervised semantic segmentation. InInternational Joint Conference on Artificial Intelligence (IJCAI), 2023. 3

    2023

  13. [13]

    Clustering propagation for universal medical image segmen- tation

    Yuhang Ding, Liulei Li, Wenguan Wang, and Yi Yang. Clustering propagation for universal medical image segmen- tation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3357–3369,

  14. [14]

    Clustering based point cloud representation learning for 3d analysis

    Tuo Feng, Wenguan Wang, Xiaohan Wang, Yi Yang, and Qinghua Zheng. Clustering based point cloud representation learning for 3d analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8283– 8294, 2023. 3

    2023

  15. [15]

    Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds

    Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. InProceedings of the AAAI Conference on Artificial Intelligence, pages 1761–1769, 2024. 3

    2024

  16. [16]

    A robust com- petitive clustering algorithm with applications in computer vision.Ieee transactions on pattern analysis and machine intelligence, 21(5):450–465, 2002

    Hichem Frigui and Raghu Krishnapuram. A robust com- petitive clustering algorithm with applications in computer vision.Ieee transactions on pattern analysis and machine intelligence, 21(5):450–465, 2002. 3

    2002

  17. [17]

    Guess: Gradually enriching synthesis for text-driven human motion generation.IEEE Transactions on Visualization and Computer Graphics, 30 (12):7518–7530, 2024

    Xuehao Gao, Yang Yang, Zhenyu Xie, Shaoyi Du, Zhongqian Sun, and Yang Wu. Guess: Gradually enriching synthesis for text-driven human motion generation.IEEE Transactions on Visualization and Computer Graphics, 30 (12):7518–7530, 2024. 2

    2024

  18. [18]

    Im- age style transfer using convolutional neural networks

    Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Im- age style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414–2423, 2016. 1

    2016

  19. [19]

    Synthesis of compositional animations from textual descriptions

    Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. InProceedings of the IEEE/CVF international conference on computer vision, pages 1396–1406, 2021. 2

    2021

  20. [20]

    Iterative motion editing with natural language

    Purvi Goel, Kuan-Chieh Wang, C Karen Liu, and Kayvon Fatahalian. Iterative motion editing with natural language. InACM SIGGRAPH 2024 Conference Papers, pages 1–9,

    2024

  21. [21]

    Generating diverse and natural 3d human motions from text

    Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152–5161, 2022. 2, 4, 6, 15

    2022

  22. [22]

    Tm2t: Stochastic and tokenized modeling for the reciprocal gener- ation of 3d human motions and texts

    Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal gener- ation of 3d human motions and texts. InECCV, 2022. 2

    2022

  23. [23]

    Momask: Generative masked model- ing of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked model- ing of 3d human motions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910, 2024. 2

    1900

  24. [24]

    Generative human motion stylization in latent space.arXiv preprint arXiv:2401.13505,

    Chuan Guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, and Li Cheng. Generative human motion stylization in latent space.arXiv preprint arXiv:2401.13505,

    Pith/arXiv arXiv

  25. [25]

    Stylemotif: Multi- modal motion stylization using style-content cross fusion

    Ziyu Guo, Young Yoon Lee, Joseph Liu, Yizhak Ben-Shabat, Victor Zordan, and Mubbasir Kapadia. Stylemotif: Multi- modal motion stylization using style-content cross fusion. arXiv preprint arXiv:2503.21775, 2025. 1, 3, 4, 6, 7, 15

    Pith/arXiv arXiv 2025

  26. [26]

    Amd: Autoregressive motion diffusion

    Bo Han, Hao Peng, Minjing Dong, Yi Ren, Yixuan Shen, and Chang Xu. Amd: Autoregressive motion diffusion. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 2022–2030, 2024. 2

    2022

  27. [27]

    Diffusion-based human motion style transfer with semantic guidance

    Lei Hu, Zihao Zhang, Yongjing Ye, Yiwen Xu, and Shi- hong Xia. Diffusion-based human motion style transfer with semantic guidance. InComputer Graphics Forum, page e15169. Wiley Online Library, 2024. 1, 2, 3

    2024

  28. [28]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. InProceed- ings of the IEEE international conference on computer vi- sion, pages 1501–1510, 2017. 2

    2017

  29. [29]

    Mo- tion puzzle: Arbitrary motion style transfer by body part

    Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Mo- tion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1–16, 2022. 2, 6, 7

    2022

  30. [30]

    Motiongpt: Human motion as a foreign lan- guage.Advances in Neural Information Processing Systems, 36:20067–20079, 2023

    Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign lan- guage.Advances in Neural Information Processing Systems, 36:20067–20079, 2023. 15

    2023

  31. [31]

    Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs

    Peng Jin, Yang Wu, Yanbo Fan, Zhongqian Sun, Wei Yang, and Li Yuan. Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. Advances in Neural Information Processing Systems, 36: 15497–15518, 2023. 2

    2023

  32. [32]

    Ro- bust clustering with applications in computer vision.IEEE transactions on pattern analysis and machine intelligence, 13(8):791–802, 1991

    Jean-Michel Jolion, Peter Meer, and Samira Bataouche. Ro- bust clustering with applications in computer vision.IEEE transactions on pattern analysis and machine intelligence, 13(8):791–802, 1991. 3

    1991

  33. [33]

    Guided motion diffusion for controllable human motion synthesis

    Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151–2162, 2023. 2

    2023

  34. [34]

    Most: Motion style transformer between diverse ac- tion contents

    Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse ac- tion contents. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705– 1714, 2024. 2

    2024

  35. [35]

    Personabooth: Per- sonalized text-to-motion generation

    Boeun Kim, Hea In Jeong, JungHoon Sung, Yihua Cheng, Jeongmin Lee, Ju Yong Chang, Sang-Il Choi, Younggeun Choi, Saim Shin, Jungho Kim, et al. Personabooth: Per- sonalized text-to-motion generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22756–22765, 2025. 2

    2025

  36. [36]

    Mulsmo: Multimodal stylized motion generation by bidirectional control flow.arXiv preprint arXiv:2412.09901, 2024

    Zhe Li, Yisheng He, Lei Zhong, Weichao Shen, Qi Zuo, Lingteng Qiu, Zilong Dong, Laurence Tianruo Yang, and Weihao Yuan. Mulsmo: Multimodal stylized motion generation by bidirectional control flow.arXiv preprint arXiv:2412.09901, 2024. 1, 2, 3, 4, 6, 7, 15

    Pith/arXiv arXiv 2024

  37. [37]

    Clusterfomer: clustering as a universal visual learner.Advances in neural information processing systems, 36:64029–64042, 2023

    James Liang, Yiming Cui, Qifan Wang, Tong Geng, Wen- guan Wang, and Dongfang Liu. Clusterfomer: clustering as a universal visual learner.Advances in neural information processing systems, 36:64029–64042, 2023. 3

    2023

  38. [38]

    Clustseg: Clustering for universal segmentation

    James Liang, Tianfei Zhou, Dongfang Liu, and Wenguan Wang. Clustseg: Clustering for universal segmentation. arXiv preprint arXiv:2305.02187, 2023. 2, 3

    Pith/arXiv arXiv 2023

  39. [39]

    Pointcluster: Deep clustering of 3-d point clouds with se- mantic pseudo-labeling.IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024

    Xiu Liu, Xinxin Han, Huan Xia, Kang Li, Haochen Zhao, Jia Jia, Gang Zhen, Linzhi Su, Fengjun Zhao, and Xin Cao. Pointcluster: Deep clustering of 3-d point clouds with se- mantic pseudo-labeling.IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024. 3

    2024

  40. [40]

    Point cloud classification using content-based trans- former via clustering in feature space.IEEE/CAA Journal of Automatica Sinica, 11(1):231–239, 2023

    Yahui Liu, Bin Tian, Yisheng Lv, Lingxi Li, and Fei-Yue Wang. Point cloud classification using content-based trans- former via clustering in feature space.IEEE/CAA Journal of Automatica Sinica, 11(1):231–239, 2023. 3

    2023

  41. [41]

    Zero- shot video grounding with pseudo query lookup and veri- fication.IEEE Transactions on Image Processing, 33:1643– 1654, 2024

    Yu Lu, Ruijie Quan, Linchao Zhu, and Yi Yang. Zero- shot video grounding with pseudo query lookup and veri- fication.IEEE Transactions on Image Processing, 33:1643– 1654, 2024. 3

    2024

  42. [42]

    Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise trans- formations and local motion phases.Proceedings of the ACM on Computer Graphics and Interactive Techniques, 5(1):1– 18, 2022. 1, 2

    2022

  43. [43]

    Sdedit: Guided image synthesis and editing with stochastic differential equa- tions.arXiv preprint arXiv:2108.01073, 2021

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equa- tions.arXiv preprint arXiv:2108.01073, 2021. 7, 14

    Pith/arXiv arXiv 2021

  44. [44]

    Spice: Seman- tic pseudo-labeling for image clustering.IEEE Transactions on Image Processing, 31:7264–7278, 2022

    Chuang Niu, Hongming Shan, and Ge Wang. Spice: Seman- tic pseudo-labeling for image clustering.IEEE Transactions on Image Processing, 31:7264–7278, 2022. 3

    2022

  45. [45]

    Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. Di- verse motion stylization for multiple style domains via spatial-temporal graph-based generative model.Proceedings of the ACM on Computer Graphics and Interactive Tech- niques, 4(3):1–17, 2021. 2

    2021

  46. [46]

    Black, and Gül Varol

    Mathis Petrovich, Michael J. Black, and Gül Varol. Action- conditioned 3D human motion synthesis with transformer V AE. InInternational Conference on Computer Vision (ICCV), 2021. 2

    2021

  47. [47]

    Temos: Generating diverse human motions from textual descriptions

    Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. InEuropean Conference on Computer Vision, pages 480–

  48. [48]

    Black, and Gül Varol

    Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. InInternational Conference on Computer Vision (ICCV), 2023. 2

    2023

  49. [49]

    Bamm: Bidirectional autoregressive motion model

    Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. InComputer Vi- sion – ECCV 2024, 2024. 2

    2024

  50. [50]

    Mmm: Generative masked motion model

    Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  51. [51]

    Mafd: Fine-grained motion style trans- fer with adaptive signal fusion

    Ziyun Qian, Dingkang Yang, Mingcheng Li, Dongliang Kou, and Lihua Zhang. Mafd: Fine-grained motion style trans- fer with adaptive signal fusion. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025. 1, 2 10

    2025

  52. [52]

    Learning clustering-based prototypes for composi- tional zero-shot learning.arXiv preprint arXiv:2502.06501,

    Hongyu Qu, Jianan Wei, Xiangbo Shu, and Wenguan Wang. Learning clustering-based prototypes for composi- tional zero-shot learning.arXiv preprint arXiv:2502.06501,

    Pith/arXiv arXiv

  53. [53]

    Monkey see, monkey do: Harnessing self- attention in motion diffusion for zero-shot motion transfer

    Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit Haim Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self- attention in motion diffusion for zero-shot motion transfer. InSIGGRAPH Asia 2024 Conference Papers, pages 1–13,

    2024

  54. [54]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 3

    2021

  55. [55]

    Gaussian mixture models

    Douglas Reynolds. Gaussian mixture models. InEncyclope- dia of biometrics, pages 827–832. Springer, 2015. 3

    2015

  56. [56]

    Dance like a chicken: Low- rank stylization for human motion diffusion.arXiv preprint arXiv:2503.19557, 2025

    Haim Sawdayee, Chuan Guo, Guy Tevet, Bing Zhou, Jian Wang, and Amit H Bermano. Dance like a chicken: Low- rank stylization for human motion diffusion.arXiv preprint arXiv:2503.19557, 2025. 1, 2

    Pith/arXiv arXiv 2025

  57. [57]

    Arbitrary mo- tion style transfer with multi-condition motion latent diffu- sion model

    Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary mo- tion style transfer with multi-condition motion latent diffu- sion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821–830,

  58. [58]

    Rsmt: Real-time stylized motion transition for characters

    Xiangjun Tang, Linjun Wu, He Wang, Bo Hu, Xu Gong, Yuchen Liao, Songnan Li, Qilong Kou, and Xiaogang Jin. Rsmt: Real-time stylized motion transition for characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1– 10, 2023. 1, 2

    2023

  59. [59]

    Style-erd: Responsive and coherent online motion style transfer

    Tianxin Tao, Xiaohang Zhan, Zhongquan Chen, and Michiel van de Panne. Style-erd: Responsive and coherent online motion style transfer. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 6593–6603, 2022. 1, 2

    2022

  60. [60]

    Motionclip: Exposing human motion generation to clip space

    Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. InEuropean Conference on Com- puter Vision, pages 358–374. Springer, 2022. 2

    2022

  61. [61]

    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. 2

    2023

  62. [62]

    Tlcontrol: Trajec- tory and language control for human motion synthesis

    Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajec- tory and language control for human motion synthesis. In European Conference on Computer Vision, pages 37–54. Springer, 2024. 2

    2024

  63. [63]

    Visual recognition with deep nearest centroids.arXiv preprint arXiv:2209.07383, 2022

    Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids.arXiv preprint arXiv:2209.07383, 2022. 3

    Pith/arXiv arXiv 2022

  64. [64]

    Fg-t2m: Fine-grained text-driven human motion generation via diffusion model

    Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 22035–22044, 2023. 2

    2023

  65. [65]

    Acmo: Attribute controllable motion generation.arXiv preprint arXiv:2503.11038, 2025

    Mingjie Wei, Xuemei Xie, and Guangming Shi. Acmo: Attribute controllable motion generation.arXiv preprint arXiv:2503.11038, 2025. 2

    Pith/arXiv arXiv 2025

  66. [66]

    Autoregressive stylized motion syn- thesis with generative flow

    Yu-Hui Wen, Zhipeng Yang, Hongbo Fu, Lin Gao, Yanan Sun, and Yong-Jin Liu. Autoregressive stylized motion syn- thesis with generative flow. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13612–13621, 2021. 2

    2021

  67. [67]

    Omnicontrol: Control any joint at any time for human motion generation.arXiv preprint arXiv:2310.08580,

    Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation.arXiv preprint arXiv:2310.08580,

    Pith/arXiv arXiv

  68. [68]

    Towards detailed text-to- motion synthesis via basic-to-advanced hierarchical diffu- sion model

    Zhenyu Xie, Yang Wu, Xuehao Gao, Zhongqian Sun, Wei Yang, and Xiaodan Liang. Towards detailed text-to- motion synthesis via basic-to-advanced hierarchical diffu- sion model. InProceedings of the AAAI Conference on Arti- ficial Intelligence, pages 6252–6260, 2024. 2

    2024

  69. [69]

    Hierarchical style-based networks for motion synthesis

    Jingwei Xu, Huazhe Xu, Bingbing Ni, Xiaokang Yang, Xi- aolong Wang, and Trevor Darrell. Hierarchical style-based networks for motion synthesis. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part XI 16, pages 178–194. Springer,

    2020

  70. [70]

    Clusterfit: Improving gen- eralization of visual representations

    Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadi- yaram, and Dhruv Mahajan. Clusterfit: Improving gen- eralization of visual representations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6509–6518, 2020. 3

    2020

  71. [71]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models.arXiv preprint arXiv:2308.06721,

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models.arXiv preprint arXiv:2308.06721,

    Pith/arXiv arXiv

  72. [72]

    Physdiff: Physics-guided human motion diffusion model

    Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 16010–16021, 2023. 2

    2023

  73. [73]

    Region-level contrastive and con- sistency learning for semi-supervised semantic segmenta- tion

    Jianrong Zhang, Tianyi Wu, Chuanghao Ding, Hongwei Zhao, and Guodong Guo. Region-level contrastive and con- sistency learning for semi-supervised semantic segmenta- tion. InInternational Joint Conference on Artificial Intel- ligence (IJCAI), 2022. 3

    2022

  74. [74]

    Generating human motion from textual descrip- tions with discrete representations

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descrip- tions with discrete representations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730–14740, 2023. 2

    2023

  75. [75]

    Energymogen: Compositional human motion generation with energy-based diffusion model in latent space

    Jianrong Zhang, Hehe Fan, and Yi Yang. Energymogen: Compositional human motion generation with energy-based diffusion model in latent space. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 17592–17602, 2025. 2, 4, 6

    2025

  76. [76]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 4 11

    2023

  77. [77]

    Re- modiffuse: Retrieval-augmented motion diffusion model

    Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Re- modiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364–373, 2023. 2

    2023

  78. [78]

    Motiondif- fuse: Text-driven human motion generation with diffusion model.IEEE transactions on pattern analysis and machine intelligence, 46(6):4115–4128, 2024

    Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondif- fuse: Text-driven human motion generation with diffusion model.IEEE transactions on pattern analysis and machine intelligence, 46(6):4115–4128, 2024. 2

    2024

  79. [79]

    Motion mamba: Efficient and long sequence motion generation

    Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. InEuropean Conference on Computer Vision, pages 265–282. Springer, 2024. 2

    2024

  80. [80]

    Attt2m: Text-driven human motion generation with multi- perspective attention mechanism

    Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multi- perspective attention mechanism. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 509–519, 2023. 2

    2023

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.