Pith. sign in

REVIEW 2 major objections 42 references

Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation

T0 review · 2 major / 0 minor · reviewed 2026-07-10 · grok-4.5

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

Pith's one-line read Certain initial noises already encode motion meaning; retrieving and refining them improves text-to-motion alignment without retraining the diffusion model.

desk verdict Solid, training-free lever for text–motion alignment: null-prompt noise tickets work in practice; novelty is incremental, empirics are the real strength. read the letter →

arxiv 2607.06843 v1 pith:22RWFDEG submitted 2026-07-07 cs.CV

Sakuya Ota , Qing Yu , Kent Fujiwara , Satoshi Ikehata , Ikuro Sato This is my paper
classification cs.CV
keywords humanmotiongenerationdiffusionmodelstext-to-motionnoiseoptimizationwinningticketssemanticalignmenttraining-freecontrol
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

Diffusion text-to-motion models often drift from the prompt, especially on long or multi-action sequences. This paper argues that the starting Gaussian noise is not neutral: some seeds—called winning noise tickets—already bias the denoiser toward particular actions and styles, even when no text is given. WINRO builds a dictionary of such seeds by generating null-prompt motions, retrieves the seed whose induced motion best matches the query text in a shared embedding space, then gently refines that seed with a similarity objective kept near the Gaussian prior. An optional LoRA adapter can replace the iterative refine step with one forward pass. Across two base generators, the method raises text–motion fidelity on HumanML3D, helps multi-track long timelines, and transfers to stylization and spatial constraints—all without changing the frozen model weights.

What carries the argument

Winning Noise Retrieval and Optimization (WINRO): a null-prompt noise dictionary mapped into a text–motion embedding space, followed by frame-adjusted retrieval and KL-regularized seed refinement (optionally amortized by a LoRA residual predictor).

What would settle it

Build the same null-prompt dictionary, then measure how often the top-retrieved seed fails to rank highly under the true text-conditioned similarity score across a held-out prompt set; if that transfer rate is near chance, the retrieval stage collapses.

Watch

Extended reading notes

Core claim

Within the Gaussian noise space of a frozen diffusion motion generator there exist winning noise tickets: seeds that carry latent structure biasing the entire denoising trajectory toward coherent motion semantics even under a null prompt. Selecting the ticket best aligned with a text query and refining it under a KL-regularized text–motion similarity objective measurably improves alignment and long-horizon consistency without retraining the base model.

Load-bearing premise

A seed that looks good under a blank prompt will still look good once text is turned on, because the text only nudges the motion features a little in embedding space.

Editorial extensions

If this is right

  • Text–motion fidelity on existing diffusion backbones can be raised at inference time by seed selection alone, with no weight updates.
  • Long multi-action timelines become more segment-faithful when the same seed is refined against interval-level text constraints.
  • Style-specific or constraint-aware dictionaries can be swapped in without redesigning the generator.
  • A trained LoRA noise refiner can keep most of the gain while cutting per-query cost to roughly one generation pass.

Reading between the lines

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

  • If noise tickets truly form smooth semantic regions, denser dictionaries or learned density models over noise could replace best-of-N retrieval for rare prompts.
  • The same null-prompt dictionary idea may transfer to other scarce-data diffusion domains (e.g., gesture, facial motion) where models are equally seed-sensitive.
  • Temporal localization in the retrieval encoder could let tickets encode segment-level rather than whole-sequence semantics, tightening multi-track control.
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

2 major / 0 minor

Summary. The paper argues that in diffusion-based text-to-motion models, certain initial Gaussian noises ("winning noise tickets") already bias the denoising trajectory toward particular motion semantics, even under null prompts. It introduces WINRO: a training-free, model-agnostic pipeline that (i) builds a null-prompt noise dictionary whose entries are motion features from a frozen retrieval model, (ii) retrieves the best-aligned seed for a query text (with frame-adjusted retrieval for variable length), and (iii) optionally refines that seed by maximizing text–motion cosine similarity under a KL regularizer that keeps the noise near N(0,I). An optional LoRA adapter amortizes the iterative refinement. Empirically, WINRO improves R-Precision, MMDist, and FID on HumanML3D for MDM-50step and MotionLCM, helps long-horizon multi-track generation on MTT, and transfers to stylization and spatial control without retraining the base generator.

Significance. If the result holds, it reframes initial noise in motion diffusion as a controllable semantic prior rather than unstructured randomness, and supplies a practical, training-free lever that improves alignment on scarce motion data without touching base weights. Strengths include multi-backbone and multi-task evidence (MDM, MotionLCM; HumanML3D, KIT-ML, MTT, stylization, spatial constraints), 95% CIs over 20 runs, null-prompt isolation of the ticket effect (Table 1), ablations on dictionary size, λ, retrieval backbone, and LoRA rank, plus a clear modular design (retrieval vs. optimization vs. amortized LoRA). The formal smoothness and transfer propositions (Sec. 3.2, Props. 1–3) give a coherent justification for why dictionary search and local KL-refined steps are well-posed.

major comments (2)
  1. The load-bearing transfer claim (Prop. 3 / Sec. 3.2 and 4.1) asserts that conditioning induces only a bounded feature shift ||Φ_c(x_T)−Φ_∅(x_T)||≤δ(c), so high null alignment implies high conditional alignment. Table 1 strongly supports intrinsic semantics under pure null generation, and Table 2 shows the same seeds help under text conditioning, but the manuscript never measures or bounds δ(c) (e.g., mean/quantile of ||Φ_c−Φ_∅|| over the dictionary, or correlation of null vs. conditional scores). Without that diagnostic, it remains unclear when retrieval can select the wrong seed for hard or compositional prompts. A short measurement or failure-case analysis would make the hinge assumption falsifiable rather than only empirically plausible.
  2. Both dictionary construction and the refinement loss L_sim (Eqs. 9–11) optimize cosine similarity in the same TMR embedding used for retrieval. HumanML3D R-Precision/FID/MMDist use a separate evaluation protocol, and the PartTMR ablation (Table 9) plus rare-prompt split (Table 12) partially address metric-chasing, but the paper does not report cross-metric correlation (TMR vs. evaluation encoder) or any independent human preference study. For claims of “semantic fidelity” beyond chasing one retrieval space—especially on rare or multi-track prompts—this is a residual load-bearing gap that should be closed or more carefully scoped.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: winning-ticket claims are empirically isolated (null-prompt Table 1) and evaluated with encoders independent of the TMR objective used for retrieval/refinement.

full rationale

WINRO is a method paper, not a first-principles derivation that claims to predict a quantity from fitted inputs. The definition of a winning noise ticket (˜S_c(x_T) ≥ γ under null conditioning) is operational, not a self-definitional loop that forces the reported HumanML3D/MTT gains. Retrieval builds a dictionary from null-prompt motions; refinement maximizes TMR cosine similarity plus a KL prior term; evaluation R-Precision/FID/MMDist use a separate evaluation protocol that the paper states is independent of TMR, and an alternate retrieval backbone (PartTMR) still helps. Table 1 further isolates intrinsic noise semantics by generating under null prompts after text-based retrieval, so alignment cannot be an artifact of text-conditioned optimization. Propositions 1–5 are standard Lipschitz/Grönwall, sampling-probability, bounded-shift, and Lagrangian arguments under stated assumptions (especially Prop. 3’s δ(c) bound); none reduce a claimed prediction to a fitted parameter or to a load-bearing self-citation uniqueness theorem. Related self-citations (e.g., PINO) appear only as prior noise-optimization context, not as the hinge that forces the central claim. Residual metric-chasing risk (optimizing a retrieval similarity related in spirit to R-Precision) is a correctness/evaluation concern, not circularity by construction under the required reduction test.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim rests on standard diffusion/DDIM assumptions, the existence of a shared text-motion embedding, and several paper-specific modeling choices (null-prompt dictionary, KL on empirical mean/variance, hand-chosen λ and dictionary size). No new physical entities are postulated; free parameters are ordinary hyperparameters.

free parameters (4)
  • KL weight λ = 5000 / 250 / 50000
    Hand-chosen per backbone and sequence length (5000 MDM HumanML3D, 250 MotionLCM, 50000 MTT); ablation shows λ=0 degrades FID, so the claim depends on this balance.
  • Noise dictionary size N = 10000
    Default 10k; ablation shows diminishing returns beyond ~10k; retrieval quality depends on N.
  • Perturbation σ and top-k = σ=0.5, k=1 or 150
    σ=0.5 and optional top-k=150 used to restore diversity after discrete retrieval; affect reported diversity/MModality.
  • LoRA rank / α / residual λ = r=64, α=64, λ=0.5
    Rank 64, α=64, residual weight 0.5 chosen for the amortized refiner; ablated but still free design choices.
assumptions (4)
  • domain assumption Deterministic DDIM/ODE sampling yields a continuous, locally Lipschitz map from initial noise to motion, so high-alignment seeds form regions rather than isolated points.
    Sec. 3.2 Lemma 1 / Prop. 1; underpins both dictionary retrieval and local refinement.
  • ad hoc to paper Conditioning induces only a bounded shift in the retrieval feature space, so null-prompt alignment transfers to text-conditioned alignment.
    Prop. 3; load-bearing for why a null-prompt dictionary works at inference.
  • domain assumption A pretrained contrastive text-motion encoder (TMR or PartTMR) provides a reliable shared embedding for both retrieval and the refinement objective.
    Used throughout Sec. 4 and experiments; method fails if the embedding is poorly aligned.
  • ad hoc to paper Empirical mean/variance KL to N(0,1) is a sufficient trust-region regularizer to keep optimized noise inside the diffusion prior.
    Eq. 10–11 and Prop. 4–5; alternative regularizers are not compared.
invented entities (2)
  • Winning noise ticket (motion) independent evidence
    purpose: Name for initial Gaussian samples that bias denoising toward particular motion semantics even under null prompts.
    Definitional label for observed semantic regions in noise space; not a new physical object, but the paper's central postulated structure.
  • Null-prompt noise dictionary with frame-adjusted retrieval (FAR) independent evidence
    purpose: Offline map from noise to null-prompt motion features at multiple lengths for text-based seed selection.
    Constructed artifact enabling training-free retrieval; evidence is the reported null-prompt and text-conditioned gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation." pith.science (2026). https://pith.science/paper/22RWFDEG

@misc{pith2026260706843,
  author       = {Pith},
  title        = {Pith review of: Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22RWFDEG}},
  note         = {Machine review of arXiv:2607.06843}
}
read the original abstract

Diffusion-based text-to-motion models synthesize realistic human motions but often exhibit semantic drift from the input text. Motion is inherently temporal, especially in compositional and long-duration sequences that require semantic consistency across multiple action segments and smooth kinematic transitions throughout the trajectory. We posit that the initial noise is central to this consistency: within the Gaussian noise space, certain instances, i.e. winning noise tickets, carry latent structure that biases denoising toward particular motion semantics, even under null prompts. We propose WInning Noise Retrieval and Optimization (WINRO), a training-free, model-agnostic framework that improves text-motion alignment by selecting and refining such tickets before diffusion sampling. WINRO maps random noises to motion features generated under null prompts, retrieves the best-aligned noise for a given text, and refines it via a KL-regularized objective that reduces the residual semantic gap while preserving the Gaussian prior. An optional LoRA-based adapter amortizes this refinement into a single forward pass. WINRO consistently improves text-motion fidelity across different base models, MDM and MotionLCM, on HumanML3D without retraining, improves temporal robustness on the MTT benchmark, and generalizes to applications such as motion stylization and spatial constraint satisfaction.

Figures

Figures reproduced from arXiv: 2607.06843 by the authors.

Figure 1
Figure 1. Winning Noise Ticket Hypothesis for motion. Certain initial noises inherently encode motion semantics, producing meaningful motion even under a null prompt (right), while others yield static poses regardless of conditioning (left). WINRO re￾trieves and refines such tickets to improve text–motion alignment. the generated trajectory must realize them with consistent semantics and smooth kinematic transitions. This lim… view at source ↗
Figure 2
Figure 2. z-score heatmaps of action-specific tendencies under four fixed noise seeds. For each seed and prompt group, we generate ten motions with varied prompts, compute the mean TMR similarity per category, and apply per-seed z-score normalization. 0.5 0.6 0.7 0.8 Text similarity score 0 10 20 30 40 Count (a) “walk” 0.650 0.675 0.700 0.725 0.750 0.775 0.800 0.825 Text similarity score 0 10 20 30 Count (b) “run” 0.6 0.7 0.8… view at source ↗
Figure 3
Figure 3. Similarity-score histograms for the fixed prompts “the person walks/runs/jumps.” For each prompt, we compute TMR similarity scores over 1,000 initial noise seeds. Smoothness implies semantic regions. If the deterministic sampler and FMotion are locally Lipschitz, then S˜ c is locally Lipschitz in xT . Therefore, whenever S˜ c(x ⋆ T ) ≥ γ, there exists a radius r > 0 such that all xT with ∥xT −x ⋆ T ∥ ≤ r also satisf… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: The pipeline of building noise dictionary D. We randomly sample noise x i T and use a diffusion model Gθ and encoder F of a retrieval model to generate a feature z i l . For the frame-adjusted dictionary, we truncate the output motion x i 0 at frame l and generate the …
Figure 5
Figure 5. Figure 5: Overview of Winning Noise Ticket Retrieval. The input text c and, optionally, the desired motion length l are used to retrieve the closest feature z † l , whose corre￾sponding noise x † T is then used to generate the desired motion. Note that we freeze the feature enco…
Figure 6
Figure 6. Figure 6: Overview of Winning Noise Ticket Refinement. The retrieved noise x † T can further be refined to improve the quality of the output using input text c, and optionally the desired motion length l. The loss L is minimized to obtain the refined noise x ∗ T . As in retrieva…
Figure 7
Figure 7. Figure 7: Comparison of motions generated with random, retrieved, and refined initial noise for each text prompt, using MDM-50step with and without our method. The texts shown above each row are used for generation. We recommend viewing the sup￾plementary video for better visual…
Figure 8
Figure 8. Figure 8: Comparison of motions generated by STMC [27] with and without initial noise refinement on multi-track long-duration generation. STMC produces motions that are not faithful to the text prompt of each segment (e.g., missing throw or hop actions), while refining the initi…
Figure 9
Figure 9. Figure 9: Overview of WINRO for motion stylization. The generic noise dictionary is replaced with a style-specific one built from 100STYLE motions [23] via DDIM inver￾sion. The optimization jointly minimizes semantic loss Lsim, style loss Lstyle, and KL regularization term λLKL.…
Figure 10
Figure 10. Figure 10: Qualitative results from the stylization experiment. Reference style motion at the left and the text prompt at the top of each example are used for generation. xs is the reference style motion. We compare WIN-O (optimization only from random noise) and WINRO (retrieva…
Figure 11
Figure 11. Figure 11: Comparison of motions generated using the initial noise optimized by Prog￾MoGen [20] and the refined noise produced by the proposed WINRO method. WINRO produces motions that better satisfy both textual and spatial constraints. 20]. Concretely, we augment Eq. 11 with a…
Figure 12
Figure 12. Figure 12: Additional comparison of motions generated with random, retrieved, and re￾fined initial noise for each text prompt, using MDM-50step [PITH_FULL_IMAGE:figures/full_fig_p029_12.png]
Figure 13
Figure 13. Figure 13: Additional compositional generation results on the multi-track timeline bench￾mark. STMC [27] often fails to realize individual action segments (e.g., missing jump, punch, or hop), while applying our noise refinement (w/ Ours) improves faithfulness to each segment’s t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    In: 3DV (2022)

    Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 3DV (2022)

    work page 2022

  2. [2]

    In: ACMMM (2024)

    Chen, C., Yang, L., Yang, X., Chen, L., He, G., Wang, C., Li, Y.: Find: Fine- tuning initial noise distribution with policy optimization for diffusion models. In: ACMMM (2024)

    work page 2024

  3. [3]

    In: CVPR (2023)

    Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR (2023)

    work page 2023

  4. [4]

    In: ECCV (2024)

    Dai, W., Chen, L.H., Wang, J., Liu, J., Dai, B., Tang, Y.: Motionlcm: Real-time controllable motion generation via latent consistency model. In: ECCV (2024)

    work page 2024

  5. [5]

    In: NeurIPS (2025)

    Eyring, L., Karthik, S., Dosovitskiy, A., Ruiz, N., Akata, Z.: Noise hypernetworks: Amortizing test-time compute in diffusion models. In: NeurIPS (2025)

    work page 2025

  6. [6]

    In: NeurIPS (2024)

    Eyring, L., Karthik, S., Roth, K., Dosovitskiy, A., Akata, Z.: Reno: Enhancing one- step text-to-image models through reward-based noise optimization. In: NeurIPS (2024)

    work page 2024

  7. [7]

    In: ICLR (2019)

    Frankle, J., Carbin, M.: The lottery ticket hypothesis: Finding sparse, trainable neural networks. In: ICLR (2019)

    work page 2019

  8. [8]

    In: ECCV (2024)

    Fujiwara, K., Tanaka, M., Yu, Q.: Chronologically accurate retrieval for temporal grounding of motion-language models. In: ECCV (2024)

    work page 2024

Show all 42 references
  1. [9]

    In: ICCV (2021)

    Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compo- sitional animations from textual descriptions. In: ICCV (2021)

    2021

  2. [10]

    In: CVPR (2022)

    Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)

    2022

  3. [11]

    In: CVPR (2024)

    Guo, X., Liu, J., Cui, M., Li, J., Yang, H., Huang, D.: Initno: Boosting text-to- image diffusion models via initial noise optimization. In: CVPR (2024)

    2024

  4. [12]

    Holden,D.,Kanoun,O.,Perepichka,M.,Popa,T.:Learnedmotionmatching.TOG 39(4), 53–1 (2020)

    2020

  5. [13]

    In: CVPR (2025)

    Hong, S., Kim, C., Yoon, S., Nam, J., Cha, S., Noh, J.y.: Salad: Skeleton-aware latent diffusion for text-driven motion generation and editing. In: CVPR (2025)

    2025

  6. [14]

    In: ICLR (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)

    2022

  7. [15]

    In: NeurIPS (2023)

    Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. In: NeurIPS (2023)

    2023

  8. [16]

    arXiv preprint arXiv:2211.15603 (2022)

    Kalakonda, S.S., Maheshwari, S., Sarvadevabhatla, R.K.: Action-gpt: Leveraging large-scale language models for improved and generalized zero shot action genera- tion. arXiv preprint arXiv:2211.15603 (2022)

    2022 arXiv

  9. [17]

    In: CVPR (2024)

    Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. In: CVPR (2024)

    2024

  10. [18]

    In: CVPR (2020)

    Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: CVPR (2020)

    2020

  11. [19]

    arXiv preprint arXiv:2508.02605 (2025)

    Li, Z., Wang, S., Zhang, Z., Tang, H.: Remomask: Retrieval-augmented masked motion generation. arXiv preprint arXiv:2508.02605 (2025)

    2025

  12. [20]

    In: CVPR (2024)

    Liu, H., Zhan, X., Huang, S., Mu, T.J., Shan, Y.: Programmable motion generation for open-set motion control tasks. In: CVPR (2024)

    2024

  13. [21]

    In: ACMMM (2023)

    Mao, J., Wang, X., Aizawa, K.: Guided image synthesis via initial image editing in diffusion model. In: ACMMM (2023)

    2023

  14. [22]

    In: ECCV (2024) Winning Noise Tickets for Motion Generation 17

    Mao, J., Wang, X., Aizawa, K.: The lottery ticket hypothesis in denoising: Towards semantic-driven initialization. In: ECCV (2024) Winning Noise Tickets for Motion Generation 17

    2024

  15. [23]

    Proceedings of the ACM on Computer Graphics and Interactive Techniques (2022)

    Mason, I., Starke, S., Komura, T.: Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques (2022)

    2022

  16. [24]

    In: ICCV (2025)

    Ota, S., Yu, Q., Fujiwara, K., Ikehata, S., Sato, I.: Pino: Person-interaction noise optimization for long-duration and customizable motion generation of arbitrary- sized groups. In: ICCV (2025)

    2025

  17. [25]

    In: ICCV (2021)

    Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion syn- thesis with transformer vae. In: ICCV (2021)

    2021

  18. [26]

    In: ICCV (2023)

    Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using con- trastive 3d human motion synthesis. In: ICCV (2023)

    2023

  19. [27]

    In: CVPRW (2024)

    Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Peng, X.B., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPRW (2024)

    2024

  20. [28]

    Big Data 4(4), 236–252 (2016)

    Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big Data 4(4), 236–252 (2016)

    2016

  21. [29]

    In: CVPR (2024)

    Shin, S., Kim, J., Halilaj, E., Black, M.J.: WHAM: Reconstructing world-grounded humans with accurate 3D motion. In: CVPR (2024)

    2024

  22. [30]

    In: ICLR (2021)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)

    2021

  23. [31]

    In: ICLR (2023)

    Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)

    2023

  24. [32]

    arXiv preprint arXiv:2412.05101 (2024)

    Wang, R., Huang, H., Zhu, Y., Russakovsky, O., Wu, Y.: The silent prompt: Ini- tial noise as implicit guidance for goal-driven image generation. arXiv preprint arXiv:2412.05101 (2024)

    2024 arXiv

  25. [33]

    arXiv preprint arXiv:2405.14828 (2024)

    Xu, K., Zhang, L., Shi, J.: Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv preprint arXiv:2405.14828 (2024)

    2024 arXiv

  26. [34]

    In: ICCV (2023)

    Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: ICCV (2023)

    2023

  27. [35]

    In: CVPR (2023)

    Ye, V., Pavlakos, G., Malik, J., Kanazawa, A.: Decoupling human and camera motion from videos in the wild. In: CVPR (2023)

    2023

  28. [36]

    In: CVPR (2024)

    Yu, Q., Tanaka, M., Fujiwara, K.: Exploring vision transformers for 3d human motion-language models with motion patches. In: CVPR (2024)

    2024

  29. [37]

    In: AAAI (2025)

    Yu,Q.,Tanaka,M.,Fujiwara,K.:Remogpt:Part-levelretrieval-augmentedmotion- language models. In: AAAI (2025)

    2025

  30. [38]

    In: ECCV (2020)

    Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: ECCV (2020)

    2020

  31. [39]

    In: CVPR (2023)

    Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: CVPR (2023)

    2023

  32. [40]

    IEEE TPAMI (2024)

    Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI (2024)

    2024

  33. [41]

    In: ICCV (2023)

    Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remod- iffuse: Retrieval-augmented motion diffusion model. In: ICCV (2023)

    2023

  34. [42]

    In: ECCV (2024) 18 S

    Zhong, L., Xie, Y., Jampani, V., Sun, D., Jiang, H.: Smoodi: Stylized motion diffusion model. In: ECCV (2024) 18 S. Ota et al. Supplementary Material Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation A Theoretical Perspective on Winning Noise ...

    2024

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.