Pith. sign in

REVIEW 3 major objections 5 minor 43 references

SIGNER: Temporally Grounded Sign Language Generation via Time-Resolved Conditioning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read GLOS claims that temporally aligned, gloss-level conditioning fixes incorrect lexical order and low semantic accuracy in sign language generation, outperforming prior methods on CSL-Daily and Phoenix-2014T.

desk verdict The core idea is real and the ablations support it, but the text-to-gloss component is mis-specified, the test-length row leaks information, and the evaluator is a black-box reimplementation — fix those before trusting the SOTA claims. read the letter →

arxiv 2506.07460 v2 pith:PXXTYG2V submitted 2025-06-09 cs.CV cs.CL

Taeryung Lee , Hyeongjin Nam , Gyeongsik Moon , Kyoung Mu Lee This is my paper
classification cs.CVcs.CL
keywords signlanguagegenerationtext-to-signgloss-levelconditioningtemporalalignmentVQ-diffusionproductionlexicalordermotion
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

Text-to-sign generation must produce gestures in the right order and keep each gesture faithful to its intended word-like unit, the gloss. This paper claims that prior methods fail on both because they encode the whole sentence into one vector, which destroys the temporal correspondence between glosses and motion. GLOS instead builds a temporally aligned gloss-level condition, gloss embeddings stretched along the motion timeline by estimated per-gloss durations, and injects it into a VQ-diffusion denoiser through a local fusion module called TAC. The reported results show lower joint-position error and better back-translation scores than earlier models on CSL-Daily and Phoenix-2014T, indicating correct lexical order and clearer semantics.

What carries the argument

The load-bearing mechanism is the pair formed by the temporal-gloss condition S and the TAC fusion block. S is a sequence of gloss text embeddings, each repeated for the gloss's estimated duration and stacked along the temporal axis, so that at every timestep the condition says which gloss should currently be signed. TAC fuses S into each body part's feature stream with AdaLN (a scale-and-shift modulation) followed by a 1D convolution; AdaLN places the condition at the right timestep, and the convolution lets neighboring frames share local context. The paper also adds inter-part attention so body and hand features exchange information, but the claimed improvement in lexical order is specifically attributed to the aligned gloss condition plus TAC.

What would settle it

Run the trained GLOS denoiser with the gloss order reversed while keeping durations; if back-translation WER and BLEU do not change substantially, temporal alignment is not the mechanism producing the reported lexical-order gains.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that temporal grounding, knowing which gloss is being realized at each motion timestep, is what sign language generation has been missing. GLOS constructs a temporal-gloss condition by converting text to a gloss sequence, assigning each gloss an estimated duration, repeating each gloss embedding along the timeline, and feeding this two-dimensional condition into a VQ-diffusion denoiser. A dedicated fusion module, TAC, delivers the condition through adaptive layer normalization and a temporal 1D convolution, so each motion token is modulated by the gloss it should realize and by nearby local context, rather than by all glosses globally. The paper reports that this design generates signs in correct lexical order with clearer per-gloss semantics, and that the full system outperforms previous methods on CSL-Daily and Phoenix-2014T.

Load-bearing premise

The method assumes that the off-the-shelf text-to-gloss model produces the right sequence of word-like units (glosses) and that their average training-set durations place each unit at the right time; if either assumption fails, the conditioning signal is misaligned.

Editorial extensions

If this is right

  • If GLOS's results hold, text-to-sign systems can encode linguistic order directly as a temporally aligned condition rather than hoping a single sentence vector preserves it.
  • The ablation results imply that global cross-attention over all glosses at every motion timestep is counterproductive for sign generation; condition fusion should be temporally local.
  • The length-variation experiments suggest that approximate durations from training-set averages are sufficient for alignment across length variations of roughly -25% to +50%.
  • On both CSL-Daily and Phoenix-2014T, the same framework improves over prior methods, supporting transfer across different sign languages and text embedding languages.
  • The PVQ-VAE body-part decomposition and inter-part attention contribute to better hand and body motion, which is part of the overall reported gain.

Reading between the lines

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

  • A natural extension the paper does not explore is replacing average gloss durations with a context-dependent duration predictor; this could tighten alignment on sentences with unusual pacing.
  • Because back-translation evaluation relies on a re-implemented recognition encoder, the quantitative advantages over baselines are conditional on that encoder matching the original model's behavior.
  • The temporally local conditioning idea may transfer to other discrete-latent generation tasks where output tokens correspond to text units, such as co-speech gesture or lip-sync generation.
  • A direct probe of the load-bearing mechanism would be to feed permuted gloss orders into the trained denoiser and check whether lexical-order metrics degrade; the paper does not report this 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

3 major / 5 minor

Summary. The manuscript (titled GLOS in the body and SIGNER in the arXiv metadata) proposes a sign language generation framework built on VQ-diffusion with a body-part PVQ-VAE. Instead of conditioning on a single sentence-level embedding, it constructs a temporally aligned gloss-level condition S by converting input text to a gloss sequence, sampling per-gloss durations, and repeating gloss embeddings along the time axis. The denoiser fuses S through a proposed temporal alignment conditioning (TAC) module consisting of adaptive layer normalization and 1D convolution, together with inter-part attention (IPA). Experiments on CSL-Daily and Phoenix-2014T report improved DTW-JPE and back-translation WER/BLEU/ROUGE scores over several prior methods. The central claim is that temporally aligned gloss-level conditioning and TAC correct lexical ordering and improve per-gloss semantic accuracy.

Significance. If substantiated, the idea of replacing global sentence conditioning with locally fused, temporally aligned gloss embeddings is a plausible and practically relevant step for text-to-sign generation. The paper has clear strengths: the ablations in Tables 1 and 2 isolate the contributions of gloss-level features, temporal alignment, local fusion, and inter-part attention; Table 3 provides robustness checks; and the qualitative examples illustrate the intended behavior. However, the end-to-end from-text claim currently rests on an unsupported text-to-gloss component, the headline comparison in Table 4 includes an oracle-duration row, and the linguistic metrics are computed with a reimplemented recognizer tied to the same reference model used for conditioning. These issues must be resolved before the state-of-the-art claim can be credited.

major comments (3)
  1. [§4.2, Table 3(a), Reference [7]] The paper states that glosses are extracted from input text using 'an off-the-shelf text-to-gloss model [7]', but reference [7] is a video-based sign language recognition and translation model, not a text-to-gloss model, and no adaptation, inversion, or fine-tuning is described. This is load-bearing because the gloss sequence S is the input to TAC and IPA, and the claimed lexical-order improvements are attributed to the temporal structure of S. The authors must either identify and document a genuine text-to-gloss model, describe how [7] was adapted, or state that the experiments used oracle glosses; in the latter case, the from-text SOTA claim in the abstract and Table 4 is not currently supported. The observation in Table 3(a) that 'Off-the-shelf [7]' outperforms the three fine-tuned text-to-gloss baselines makes the missing stage especially suspicious.
  2. [Table 4, §5.2] The 'Ours (Test len)' row in Table 4 uses ground-truth test-set lengths for the gloss-level condition, as acknowledged in §5.2, while all baseline methods generate without such oracle duration information. Injecting test-time duration information into the conditioning signal gives GLOS an unfair advantage and confounds the headline comparison. The 'Ours (Train len)' row is the appropriate setting for a from-text comparison, or all methods should be evaluated under an identical length protocol. As presented, the main SOTA claim is not cleanly supported.
  3. [§5.1, §4.2] Back-translation evaluation uses a reimplemented CSLR feature encoder of [7], and [7] is also cited as the source of the conditioning glosses. This creates a circularity risk: the metric may reward generated motions that match the recognition biases and gloss vocabulary of [7] rather than human-judged sign correctness, especially because the evaluator is a reimplementation rather than the original model. The authors should validate the reimplemented evaluator against the original [7] or an independent recognizer, add human evaluation, or at least show that the main ranking in Table 4 is insensitive to the choice of evaluator.
minor comments (5)
  1. [Title and metadata] The arXiv metadata calls the paper 'SIGNER: Temporally Grounded Sign Language Generation via Time-Resolved Conditioning', while the manuscript itself is titled 'GLOS: Sign Language Generation with Temporally Aligned Gloss-Level Conditioning'; these should be aligned before resubmission.
  2. [§5.3] The text says 'We report the qualitative comparison with previous state-of-the-art methods', but Table 4 reports quantitative results; this should read 'quantitative comparison'.
  3. [§3.2] The sentence 'Denoiser is trained to estimate the posterior transition distribution q(kt−1 | xt, k0)' appears to contain a typo: the conditioning variable should be kt, not xt.
  4. [§5.1 and Table 3(a)] The fine-tuned mT5, M2M, and mBART text-to-gloss baselines in Table 3(a) are not described; the paper should specify the training data, adaptation method, and hyperparameters used for these models.
  5. [§4.2] The sentence 'For the glosses which are not included in the training set, we utilize the mean length of all glosses' is grammatically awkward and should be rephrased; additionally, it is unclear how this fallback interacts with the test-length protocol used in the ablations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GLOS's TAC contribution is supported by independent ablations; the shared use of [7] in conditioning and evaluation is a validity concern, not a by-construction reduction.

full rationale

The paper's central claim is that temporally aligned gloss-level conditioning, fused through TAC, improves sign language generation. That claim is supported by the ablation in Table 1, which compares TAC against sentence-level conditioning, gloss embeddings without temporal structure, cross-attention fusion, and AdaLN+FC. These baselines are distinct condition/fusion configurations and the improvements are empirical, not forced by construction. The gloss-condition construction in Sec. 4.2 uses the text-to-gloss model [7] and training-set mean durations; neither is a fitted parameter of the generation model, and the duration sampling is explicitly varied in Table 3(b) rather than treated as a prediction. The use of test-set lengths in the main comparison ('Ours (Test len)') is an evaluation-protocol weakness that can inflate results relative to baselines, but it is not a circularity: the reference motions and glosses are independent ground-truth data. The same external model [7] appears as the stated text-to-gloss source and as the back-translation evaluator, which raises a legitimate concern that the linguistic metrics reward motions aligned with that model's gloss vocabulary. However, this is not a self-citation chain, nor a fitted-input-called-prediction reduction: [7] is a fixed external model, not a parameter fit to the evaluation metric, and the WER/BLEU/ROUGE scores are computed against ground-truth gloss/text references rather than against [7]'s own outputs. The paper contains no derivation that equates its inputs to its outputs by definition, and no load-bearing self-citation. Therefore, under the strict circularity criteria, there is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an external text-to-gloss component, training-set duration statistics, and the reimplemented back-translation recognizer. No new physical entities are introduced.

free parameters (1)
  • Per-gloss mean duration l_j = Mean length per gloss computed from training data, values not tabulated
    Used in Section 4.2 to repeat each gloss embedding for l_j steps and construct the temporal-gloss condition. Glosses absent from the training set use the global mean. This is a statistic fitted to the training split.
assumptions (4)
  • ad hoc to paper Off-the-shelf text-to-gloss model [7] converts input text into correct gloss sequences.
    Section 4.2 and Tables 3a and 4 depend on this model. The cited paper is a sign language recognition and translation system from video, so its suitability as text-to-gloss is not established.
  • domain assumption Gloss duration can be approximated by the training-set mean length per gloss.
    Section 4.2 samples lengths from the training set and assigns gloss semantics across time. This ignores context-dependent duration variability.
  • domain assumption mBART gloss embeddings carry enough semantics for conditioning.
    Section 4.2 uses pretrained mBART embeddings for glosses, with no analysis that the embedding space aligns with sign semantics.
  • ad hoc to paper Back-translation via a reimplemented CSLR feature encoder reliably evaluates sign-to-text quality.
    Section 5.1 states the official recognizer could not be reproduced, and the reimplementation is used as the metric. Evaluation validity depends on the fidelity of this reimplementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SIGNER: Temporally Grounded Sign Language Generation via Time-Resolved Conditioning." pith.science (2026). https://pith.science/paper/PXXTYG2V

@misc{pith2026250607460,
  author       = {Pith},
  title        = {Pith review of: SIGNER: Temporally Grounded Sign Language Generation via Time-Resolved Conditioning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXXTYG2V}},
  note         = {Machine review of arXiv:2506.07460}
}
read the original abstract

Sign language generation (SLG), also known as text-to-sign generation, aims to bridge the communication gap between signers and non-signers. Unlike many other generative tasks, SLG must satisfy two fundamental linguistic constraints. First, sign language expresses meaning through a sequence of gestures aligned with word-like units called glosses, and therefore requires correct lexical ordering to preserve intended meaning. Second, each gesture should faithfully reflect the intended gloss (semantic accuracy). Despite recent progress, existing SLG methods frequently produce signs with incorrect lexical order and low semantic accuracy. A common limitation of prior approaches stems from globally fused conditioning strategies, which weaken temporal grounding, the temporal correspondence between glosses and their realized sign segments. This often leads to incorrect lexical order and semantically ambiguous signs. To address this limitation, we propose SIGNER, a SIGN language generation framework with timE-Resolved conditioning to ensure temporal grounding, leveraging a temporal-gloss condition and local temporal fusion (LTF). SIGNER constructs a temporal-gloss condition by estimating a gloss sequence and its durations from input text, and assigning gloss semantics across the temporal dimension. We then introduce LTF, a temporally grounded fusion module that integrates the temporal-gloss condition within a constrained temporal window during denoising. By enforcing temporal locality in condition fusion, LTF preserves temporal grounding, leading to correct lexical ordering and clearer per-gloss semantics. Experiments on Phoenix-2014T and CSL-Daily demonstrate state-of-the-art performance, further supported by motion-smoothness analysis. The project page is available here https://taeryunglee.github.io/projects/signer.

Figures

Figures reproduced from arXiv: 2506.07460 by the authors.

Figure 1
Figure 1. Qualitative comparison between our proposed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Conceptual comparison. We compare our temporally aligned gloss-level conditioning against conventional sentence-level conditioning. Our method leverages gloss-level conditions with temporal alignment conditioning (TAC) to ensure temporal alignment between gloss conditions and associated motion segments, generating signs in correct lexical order and high semantic accuracy. 1 Introduction Sign language serves as a vit… view at source ↗
Figure 3
Figure 3. Overview. We first construct the motion latent space with PVQ-VAE (left). In generation phase (right), we first encode input text into gloss-level condition with off-the-shelf text-to-gloss model [7] and length sampling. Our proposed denoiser with TAC generates the tokens in the motion latent space and forwards them into the PVQ-VAE decoder to generate sign language. input Xpart is encoded as Z e,part = Epart(Xpart)… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Model architecture. We illustrate the detailed architecture of our proposed denoiser and inter-part attention (IPA) module. (a) In our model, each body part—body, left hand, right hand, and face—is encoded independently without inter-part interaction, except within the…
Figure 5
Figure 5. Figure 5: Comparison of condition fusion methods. We visualize two condition fusion methods: (a) Our TAC consists of AdaLN and 1D convolution to preserve the temporal alignment with local temporal context. (b) Cross-attention, which fails to maintain the temporal alignment const…
Figure 6
Figure 6. Figure 6: Qualitative result. We present the qualitative comparison of our GLOS with GT, Mo￾Mask [12], and NSA [3]. gloss is represented by their own colors, and the corresponding lexical signs are also represented in the same colors as the glosses. Human-interpreted glosses are…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [7]

    Two-stream network for sign language recognition and translation

    Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022

    work page 2022

  2. [1]

    Black, and Gül Varol

    Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal Action Compositions for 3D Humans. In 3DV, 2022

    work page 2022

  3. [2]

    Layer normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv, 2016

    work page 2016

  4. [3]

    Neural Sign Actors: A diffusion model for 3d sign language production from text

    Vasileios Baltatzis, Rolandos Alexandros Potamias, Evangelos Ververas, Guanxiong Sun, Jiankang Deng, and Stefanos Zafeiriou. Neural Sign Actors: A diffusion model for 3d sign language production from text. In CVPR, 2024

    work page 2024

  5. [4]

    Neural sign language translation

    Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, 2018

    work page 2018

  6. [5]

    The language of motion: Unifying verbal and non-verbal language of 3d human motion

    Changan Chen, Juze Zhang, Shrinidhi Kowshika Lakshmikanth, Yusu Fang, Ruizhi Shao, Gordon Wetzstein, Li Fei-Fei, and Ehsan Adeli. The language of motion: Unifying verbal and non-verbal language of 3d human motion. In arXiv, 2024

    work page 2024

  7. [6]

    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. In CVPR, 2023

    work page 2023

  8. [8]

    Acquiring a first language in adolescence: The case of basic word order in american sign language

    Qi Cheng and Rachel I Mayberry. Acquiring a first language in adolescence: The case of basic word order in american sign language. Journal of child language, 46(2):214–240, 2019

    work page 2019

Show all 43 references
  1. [9]

    Beyond english-centric multilingual machine translation

    Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, and Armand Joulin. Beyond english-...

    2020

  2. [10]

    Vector quantized diffusion model for text-to-image synthesis

    Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022

    2022

  3. [11]

    TM2T: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts

    Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. TM2T: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022

    2022

  4. [12]

    MoMask: Generative masked modeling of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. MoMask: Generative masked modeling of 3d human motions. In CVPR, 2024

    2024

  5. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  6. [14]

    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. In ICCV, 2017

    2017

  7. [15]

    Eui Jun Hwang, Jung Ho Kim, Suk Min Cho, and Jong C. Park. Non-autoregressive sign language production via knowledge distillation. In BMVC, 2021

    2021

  8. [16]

    MotionGPT: Human motion as a foreign language

    Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. MotionGPT: Human motion as a foreign language. In NeurIPS, 2024

    2024

  9. [17]

    MultiAct: Long-term 3d human motion generation from multiple action labels

    Taeryung Lee, Gyeongsik Moon, and Kyoung Mu Lee. MultiAct: Long-term 3d human motion generation from multiple action labels. In AAAI, 2023

    2023

  10. [18]

    T2LM: Long-term 3d human motion generation from multiple sentences

    Taeryung Lee, Fabien Baradel, Thomas Lucas, Kyoung Mu Lee, and Gregory Rogez. T2LM: Long-term 3d human motion generation from multiple sentences. InCVPR Workshop on Human Motion Generation, 2024

    2024

  11. [19]

    Sequential texts driven cohesive motions synthesis with natural transitions

    Shuai Li, Sisi Zhuang, Wenfeng Song, Xinyu Zhang, Hejia Chen, and Aimin Hao. Sequential texts driven cohesive motions synthesis with natural transitions. In ICCV, 2023

    2023

  12. [20]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, 2004. 10

    2004

  13. [21]

    Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics

    Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In Proceedings of the 42nd annual meeting of the association for computational linguistics (ACL-04) , pages 605–612, 2004

    2004

  14. [22]

    Multilingual denoising pre-training for neural machine translation

    Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics , 8:726–742, 2020

    2020

  15. [23]

    Humantomato: Text-aligned whole-body motion generation

    Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung- Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In arXiv, 2023

    2023

  16. [24]

    Information retrieval for music and motion , volume 2

    Meinard Müller. Information retrieval for music and motion , volume 2. Springer, 2007

    2007

  17. [25]

    Order of the major constituents in sign languages: Implications for all language

    Donna Jo Napoli and Rachel Sutton-Spence. Order of the major constituents in sign languages: Implications for all language. Frontiers in psychology, 5:376, 2014

    2014

  18. [26]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002

    2002

  19. [27]

    Black, and Gül Varol

    Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer V AE. InICCV, 2021

    2021

  20. [28]

    Black, and Gül Varol

    Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In ECCV, 2022

    2022

  21. [29]

    Adversarial training for multi- channel sign language production

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Adversarial training for multi- channel sign language production. In BMVC, 2020

    2020

  22. [30]

    Progressive transformers for end-to-end sign language production

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In ECCV, 2020

    2020

  23. [31]

    Human motion diffusion as a generative prior

    Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024

    2024

  24. [32]

    Human motion diffusion model

    Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023

    2023

  25. [33]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017

    2017

  26. [34]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017

    2017

  27. [35]

    Mg-motionllm: A unified framework for motion comprehension and generation across multiple granularities

    Bizhu Wu, Jinheng Xie, Keming Shen, Zhe Kong, Jianfeng Ren, Ruibin Bai, Rong Qu, and Linlin Shen. Mg-motionllm: A unified framework for motion comprehension and generation across multiple granularities. In CVPR, 2025

    2025

  28. [36]

    mT5: A massively multilingual pre-trained text-to-text transformer

    Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. mT5: A massively multilingual pre-trained text-to-text transformer. In NAACL, 2021

    2021

  29. [37]

    T2S-GPT: Dynamic vector quantization for autoregressive sign language production from text

    Aoxiong Yin, Haoyuan Li, Kai Shen, Siliang Tang, and Yueting Zhuang. T2S-GPT: Dynamic vector quantization for autoregressive sign language production from text. In ACL, 2024

    2024

  30. [38]

    SignAvatars: A large-scale 3d sign language holistic motion dataset and benchmark

    Zhengdi Yu, Shaoli Huang, Yongkang Cheng, and Tolga Birdal. SignAvatars: A large-scale 3d sign language holistic motion dataset and benchmark. In ECCV, 2024

    2024

  31. [39]

    T2M-GPT: Generating human motion from textual descriptions with discrete representations

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023

    2023

  32. [40]

    MotionDiffuse: Text-driven human motion generation with diffusion model

    Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. MotionDiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 11

    2024

  33. [41]

    Improving sign language translation with monolingual data by sign back-translation

    Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021

    2021

  34. [42]

    Signs as Tokens: A retrieval-enhanced multilingual sign language generator

    Ronglai Zuo, Rolandos Alexandros Potamias, Evangelos Ververas, Jiankang Deng, and Stefanos Zafeiriou. Signs as Tokens: A retrieval-enhanced multilingual sign language generator. In arXiv, 2024

    2024

  35. [43]

    A simple baseline for spoken language to sign language translation with 3d avatars

    Ronglai Zuo, Fangyun Wei, Zenggui Chen, Brian Mak, Jiaolong Yang, and Xin Tong. A simple baseline for spoken language to sign language translation with 3d avatars. In ECCV, 2024. 12 NeurIPS Paper Checklist 13

    2024

Pith tools

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