Motion-example-controlled Co-speech Gesture Generation Leveraging Large Language Models

Bohong Chen; Kun Zhou; Yao-Xiang Ding; Youyi Zheng; Yumeng Li

REVIEW 5 major objections 5 minor 5 references

Motion-example-controlled Co-speech Gesture Generation Leveraging Large Language Models

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A fine-tuned LLM can follow a motion example when generating co-speech gestures, beating specialized pipelines on quality, diversity, and example fidelity.

desk verdict A genuinely novel way to condition an LLM on motion-example tokens for co-speech gestures, with strong in-distribution results, but the example-control claim needs out-of-distribution tests before I trust it. read the letter →

arxiv 2507.20220 v1 pith:W5ACXMZP submitted 2025-07-27 cs.CV

Bohong Chen , Yumeng Li , Youyi Zheng , Yao-Xiang Ding , Kun Zhou This is my paper

classification cs.CV

keywords co-speechgesturegenerationmotion-examplecontrollargelanguagemodelsresidualvectorquantizationmotiontokensaudiomultimodalfine-tuning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a large language model, fine-tuned to read speech audio and motion tokens, can generate co-speech gestures that follow a user's motion example while remaining in sync with the spoken audio. This matters because current example-based systems either reduce examples to coarse style labels or collapse them into pseudo-labels, losing the kinematic details that make a gesture recognizably from that example. The proposed mechanism treats the motion example as an explicit token-level system prompt and the audio as the user query, with a three-stage fine-tuning schedule that first adapts new token embeddings, then learns speech-to-gesture mapping, then learns example-conditioned generation. The paper reports the lowest Fréchet Gesture Distance, the highest motion diversity, and the highest example-gesture similarity among compared methods on two mocap datasets, while losing almost none of the base LLM's text abilities.

What carries the argument

The central object is the motion-example prompt: a token sequence produced by encoding a user-supplied motion with a residual-quantized VQ-VAE split into three functional regions (upper body, lower body, hands), then applying deduplication, shuffling, and random dropout to form a system prompt. It carries the argument because it is the only channel through which example identity enters the model, and a penalty term in the loss keeps the generative distribution near the example's token set. The second mechanism is the three-stage fine-tuning schedule: first only token embeddings and the output projection are trained to place new audio and motion tokens in the existing embedding space, then all parameters are trained for speech-to-gesture, then all parameters are trained for example-conditioned generation. That schedule is what preserves the LLM's original text capabilities while teaching it the new modalities.

What would settle it

Record a person's gesture from a monocular in-the-wild video, tokenize it, use it as the motion example with a fixed neutral audio clip, and measure whether the generated motion's distance to the example stays as low as it does for mocap-captured examples; if it rises sharply, the tokenizer-transfer premise is falsified.

Watch

Extended reading notes

Core claim

The paper's discovery, stated on its own terms, is that explicit kinematic reference beats compressed style encoding for controllable co-speech gesture generation. A single 0.5-billion-parameter instruction-tuned LLM, fine-tuned in three stages, simultaneously interprets audio tokens and motion tokens from a residual-quantized VQ-VAE, generating gesture tokens autoregressively. The motion example is not encoded into a style vector; it is deduplicated, shuffled, and partially dropped into a system-prompt context, and an auxiliary loss term penalizes output tokens that do not appear in the example. At inference a logit boost-and-decay sampling rule lets the user dial how strictly the output follows the example. The paper reports that this yields state-of-the-art FGD, diversity, and example-gesture similarity on the BEAT2 and ZeroEGGS benchmarks, preserves the LLM's text benchmarks nearly unchanged, and supports granular control of upper body, lower body, and hands individually.

Load-bearing premise

The approach assumes the motion tokenizer preserves the style of whatever example a user supplies, but the authors report that in-the-wild video-derived motions suffer significant reconstruction errors after tokenization, so example control can fail precisely when the example is far from the training motion-capture distribution.

Editorial extensions

If this is right

Motion examples can replace pseudo-label feature extractors entirely: the same prompt slot accepts a clip, a single pose, a video-derived motion, or a text-retrieved motion.
Gesture fidelity becomes a tunable dial: the logit boost-and-decay parameter lets a user trade strict example adherence against speech congruence without retraining.
Body-part control follows from tokenization: combining upper-body tokens from one example and lower-body or hand tokens from another produces mixed-style output.
Fine-tuning on audio and motion tokens does not destroy the model's language abilities; text benchmarks remain nearly at base levels, so the same checkpoint can still do text tasks.
At current data scales, a 0.5B backbone is enough; scaling to 7B gives no quality gain, implying that data, not capacity, is the limiting factor.

Reading between the lines

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

If the motion tokenizer is the bottleneck, a causal residual-quantized tokenizer could extend this example-prompt conditioning to real-time streaming generation; the paper itself notes that its current tokenizer is non-causal.
The deduplicate-and-shuffle training construction suggests the model learns style at the level of token sets rather than exact sequences, so one could test whether retrieving a clean, in-distribution example closes the gap for out-of-distribution user inputs without retraining.
Any control signal that can be tokenized, such as emotion, speaker identity, trajectory, or music, might occupy the same system-prompt slot, generalizing the framework well beyond kinematic examples.
Because the paper finds no benefit from a 7B backbone under current data scarcity, larger gains are more likely to come from larger clean co-speech mocap corpora or from video-derived motion data that passes the tokenizer's distribution test.

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.

Desk Editor's Note

read the letter

Short version: this is a real step forward for example-controlled co-speech gesture, with a genuinely new way of conditioning an LLM on motion-example tokens. The BEAT2 numbers are strong, and the paper is honest about its limits. But the example-control claim is only tested in-distribution, and the training objective makes it hard to separate style transfer from token copying. Needs OOD evaluation before I'd trust the headline.

What's new: MECo treats motion examples as explicit token-level prompt contexts rather than style vectors or pseudo-labels. Three-stage fine-tuning (embedding init, speech-to-gesture SFT, example-conditioned SFT) with logit boosting is a clean design. The result that a 0.5B Qwen retains near-full text capability after multimodal fine-tuning is worth noting. Code, models, and videos are promised, and the paper runs a user study.

Soft spots, in order of importance. First, the central conditioning premise: Eq. 4 constructs training examples by dropout/shuffle/dedup of the target token sequence itself. So during training, the example is a subset of the tokens the model must output. At inference, the example comes from a different motion; the model must generalize from 'subset of my output' to 'token set of another style'. That is a different task, and the paper gives no direct evidence it works. Table 2 measures similarity on in-distribution examples from the same datasets, so it cannot detect failure. The paper's own Section 5 admits the motion VQ-VAE produces significant reconstruction errors on in-the-wild video examples, which limits the advertised video modality exactly where the method would need to work. This is not fatal to the no-example FGD/diversity result, but it is a load-bearing issue for the paper's main selling point.

Second, the quantitative comparisons have fixable problems: Table 1 has no error bars and relies on greedy decoding; the ZeroEGGS comparison uses SynTalker's checkpoint, which was not trained on ZeroEGGS (Section 4.2.2), so that margin is inflated; and the higher BC of a baseline is dismissed as 'meaningless jitters' without quantitative support. The lambda penalty also directly optimizes token overlap with the example, so the example-similarity metric is partly a training target. None of these are fatal, but they need addressing.

Third, minor: 29 participants in the user study is small but not unusual; foot sliding is acknowledged.

The paper deserves a serious referee. The novelty is solid, the engineering is careful, and the limitations are discussed honestly. My recommendation: accept conditionally, with required experiments on out-of-distribution motion examples (held-out styles, different subjects, the video pipeline), plus error bars on Table 1 and a redo of the ZeroEGGS comparison with a properly trained baseline.

Referee Report

5 major / 5 minor

Summary. The paper presents MECo, a three-stage LLM fine-tuning framework for motion-example-controlled co-speech gesture generation. Speech audio is tokenized with HuBERT and motion with a body-part RQ-VAE; a Qwen2.5-0.5B-instruct model is fine-tuned first on the LLM's own token-prediction task for new embeddings, then on speech-to-gesture SFT, then on example-conditioned SFT where the training example is a Drop&Shuffle&Dedup subset of the target token sequence (Eq. 4) with a token-set penalty (Eq. 5). Inference uses a logit boost for example tokens and a decay factor (Eq. 6). On BEAT2 and ZeroEGGS, the paper reports state-of-the-art FGD, diversity, and example-gesture similarity, supported by ablations, a user study, preservation of LLM text capabilities, and demos of multimodal control (motion clips, poses, video, text).

Significance. If the results hold, the paper is significant: it gives a concrete recipe for injecting audio and motion tokens into a small instruction-tuned LLM, reports substantially better FGD than prior specialized architectures (3.401e-1 vs 5.512e-1 for EMAGE on BEAT2 in Table 1), verifies that LLM text capabilities are largely preserved (MMLU 46.27 vs 46.50 in Table 3), and quantifies inference speed (270 tokens/sec with vLLM, Section 4.1.2). The release of code, models, and videos is a clear strength, as is the explicit acknowledgment of limitations in Section 5 and Appendix G. The central example-control mechanism is plausible, but the evidence as currently presented leaves a load-bearing gap between training-condition construction and the advertised user-facing 'motion-example' control.

major comments (5)

[§4.2.1, Table 1] The headline 'state-of-the-art' claim on BEAT2 FGD/diversity rests on a single greedy decoding run with no error bars or multiple-seed reporting. Table 1 lists MECo at FGD 3.401e-1 versus EMAGE 5.512e-1; given FGD's sensitivity to feature distributions and test-split protocol, the paper should report at least three decoding runs (or seeds) with mean/std. This is load-bearing because the SOTA claim in the abstract depends entirely on these numbers.
[§4.2.2, Table 2] The SynTalker comparison on BEAT2 is unfair as stated: the paper says it 'directly use[s] their published code and pre-trained checkpoints' (Section 4.2.2), but SynTalker's released checkpoint was not trained on the BEAT2 protocol used here (BEAT2 single-speaker split). Comparing an untrained checkpoint on a different data distribution to a model trained/tested on that split explains much of the large FGD gap (8.21e-1 vs 4.12e-1 test FGD). The authors should either retrain SynTalker with its official recipe on this split or clearly label the row as an out-of-distribution reference.
[§3.2.3, Eq. (4)] The training-condition construction E_c = Drop&Shuffle&Dedup(c_1,...,c_Tc) means that at training time the model only sees 'examples' that are token subsets of the ground-truth target itself. At inference, a user example is a tokenized version of a different motion, so the learned conditioning must generalize from 'subset of what I should emit' to 'token set of a different style'. The example-similarity evaluation in Table 2 samples in-distribution examples from the same dataset, so it cannot detect a failure of this generalization. The paper should add an out-of-distribution test (e.g., examples from BEAT2 speakers not seen in training, or examples from a different mocap corpus) to support the claimed example-gesture similarity.
[§5, Discussions] Section 5 explicitly states that for many in-the-wild videos, monocular-mocap SMPL-X parameters have 'significant reconstruction errors after VQ-VAE processing', i.e., the tokenizer corrupts exactly the out-of-distribution inputs used in the advertised video-input modality. This is an acknowledged limitation that directly bounds the 'diverse input modalities' and 'motion-example control' claims. At minimum, the abstract and Section 1 should not describe video-input control without this caveat, and the paper should quantify how often such video examples fail and how graceful the degradation is.
[§4.4.3, Table 5] The claim that a 7B backbone gives no scaling benefit is presented as a finding, but it is based on single runs and is confounded by training budget: the paper reports the same 22-hour, 4-GPU training budget for the 7B model. To support the 'data-scarcity' interpretation (Section 4.4.3), the authors should either report a learning-rate/epoch-schedule comparison that accounts for the 14x parameter increase or soften the claim to 'no benefit under our fixed budget'.

minor comments (5)

[§4.2.1] The text says 'We have achieved SOTA performance as shown in Table 5', but Table 5 is the ablation table; the SOTA comparison is Table 1. The table reference should be corrected.
[§4.1.2 and §3.3] The hyperparameters β=5 and γ=0.9 are stated as defaults in Section 4.1.2 but their sensitivity is never studied. A one-paragraph sensitivity plot (β=0..10) would strengthen the example-control mechanism.
[§4.2.2] The sentence 'we still use FGD as the metric, as it is the most effective in determining whether two sequences are similar' conflates distribution-level Fréchet distance with pairwise example similarity. The paper should clarify that FGD1/FGD2 are computed per-pair here and that this differs from the standard FGD usage in Table 1.
[§3.1] The notation T_c vs T_C (in Eq. 5 'T_C' vs 'T_c') is inconsistent. Also, the sentence 'Since the origin LLM does not have corresponding audio and motion tokens' contains a typo ('origin' → 'original').
[§4.4.1] The ablation 'MECo (w/o freeze)' is described as making all parameters trainable in Section 4.4.1, but the table also lists 'MECo (w/o freeze&pretrain)'; the two conditions are not fully disambiguated in the text.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the speech-to-gesture core is trained on independent audio-motion pairs; example conditioning is an explicit, acknowledged training/inference objective rather than a disguised derivation.

full rationale

The paper's derivation chain is self-contained at the level where it matters. The speech-to-gesture mapping (Eq. 3) is a standard SFT objective on paired HuBERT audio tokens and RQ-VAE motion tokens; the reported FGD/diversity results in Table 1 are obtained with the motion-example branch explicitly emptied ('the motion examples in the model's input conditions are set to be empty during sampling'), so they are not fitted by construction and are falsifiable against EMAGE, SynTalker, etc. The example-conditioning branch (Eqs. 4-6) does define the training example E_c as a Drop&Shuffle&Dedup of the target token sequence and adds a penalty for tokens outside the example plus a logit boost β=5 at inference; this means high example-similarity numbers in Table 2 partly reflect the explicit objective, not an emergent discovery. That is a design choice (and an honest one: β and γ are presented as controllable adherence knobs), not an equation that by construction equals the reported test metric, because the test examples are sampled separately and the model still must order tokens audio-appropriately. The paper's own Section 5 concedes the VQ-VAE corrupts in-the-wild video examples, which is a generalization limitation rather than a circular step. The only same-author citation that could be relevant is SynTalker (Chen et al. 2024a), used as a baseline and related work; no load-bearing premise or uniqueness theorem is imported from it. Thus no specific reduction of a prediction to its inputs can be exhibited; score 2 reflects a minor in-built objective and self-citation, not substantive circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests primarily on the motion tokenizer's fidelity, the representativeness of the two mocap datasets, and the ad-hoc construction of training examples. Beta, gamma, and lambda are user-set hyperparameters that directly control the example-following behavior; their values are chosen by hand rather than derived.

free parameters (3)

beta (β) = 5 (default)
Logit adjustment for motion-example tokens during inference (Eq. 6); manually selected, directly controls example adherence.
gamma (γ) = 0.9
Decay factor applied to logits of previously sampled tokens to promote diversity (Eq. 6).
lambda (λ) = not specified
Weight of the penalty term in Eq. 5 that discourages tokens outside the motion example; no value given, likely tuned manually.

assumptions (4)

domain assumption HuBERT discrete speech units carry sufficient prosodic and phonetic information for gesture synthesis.
Section 3.2.1 tokenizes audio with HuBERT; the method assumes these units drive gestures.
domain assumption Motion RQ-VAE discrete tokens preserve the kinematic style and body-part granularity of the example motion.
The entire example-control pipeline conditions on token sets; Section 5 admits the tokenizer fails on out-of-distribution videos.
ad hoc to paper Deduplicating and shuffling the ground-truth token sequence produces motion examples whose distribution matches user-provided examples.
Eq. 4 defines training examples this way; if real user examples differ, the model may not generalize.
ad hoc to paper The penalty term in Eq. 5 that restricts output tokens to the example's token set does not suppress speech-appropriate gestures.
No ablation of λ is reported; the user study only partially addresses this.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Motion-example-controlled Co-speech Gesture Generation Leveraging Large Language Models." pith.science (2026). https://pith.science/paper/W5ACXMZP

@misc{pith2026250720220,
  author       = {Pith},
  title        = {Pith review of: Motion-example-controlled Co-speech Gesture Generation Leveraging Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W5ACXMZP}},
  note         = {Machine review of arXiv:2507.20220}
}

read the original abstract

The automatic generation of controllable co-speech gestures has recently gained growing attention. While existing systems typically achieve gesture control through predefined categorical labels or implicit pseudo-labels derived from motion examples, these approaches often compromise the rich details present in the original motion examples. We present MECo, a framework for motion-example-controlled co-speech gesture generation by leveraging large language models (LLMs). Our method capitalizes on LLMs' comprehension capabilities through fine-tuning to simultaneously interpret speech audio and motion examples, enabling the synthesis of gestures that preserve example-specific characteristics while maintaining speech congruence. Departing from conventional pseudo-labeling paradigms, we position motion examples as explicit query contexts within the prompt structure to guide gesture generation. Experimental results demonstrate state-of-the-art performance across three metrics: Fr\'echet Gesture Distance (FGD), motion diversity, and example-gesture similarity. Furthermore, our framework enables granular control of individual body parts and accommodates diverse input modalities including motion clips, static poses, human video sequences, and textual descriptions. Our code, pre-trained models, and videos are available at https://robinwitch.github.io/MECo-Page.

Figures

Figures reproduced from arXiv: 2507.20220 by the authors.

**Figure 1.** Given a motion example and a speech audio clip, our method generates vivid co-speech gestures. Motion examples can be a motion clip, a single pose, a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Our model takes motion examples and speech audio as inputs. Both [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The structure of our example-guided co-speech generation model. Both motion and audio are tokenized and fed into a large language model (LLM) to [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: We demonstrate the versatility of our method across various control modalities, including direct motion control, pose control, video control, and text [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: We can control specific body parts by tokenizing examples and combining their corresponding tokens. For instance, we tokenize two examples, use the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: A qualitative comparison between our method and ZeroEGGS. Both methods use the same input, with the motion example displayed on the left side of [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: A qualitative comparison between our method and SynTalker. Both methods use the same input, with the motion example displayed on the left side of [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: The prompt format of regular LLM and of our method. We prompt the [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: Screenshot of the user interface used for user study. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

5 extracted references · 1 canonical work pages

[2]

MoMask: Generative Masked Modeling of 3D Human Motions. (2023). arXiv:2312.00063 [cs.CV] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. 2022. A Motion Matching-based Framework for Controllable Gesture Synthesis from Speech. In ACM SIGGRAPH 2022 Conference Proceedings (Van...

arXiv 2023
[4]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Guided Motion Diffusion for Controllable Human Motion Synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2151– 2162. Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N. Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R. Thórisson, and Hannes Vilhjálmsson. 2006. Towards a Common Framework for Multimodal Genera...

arXiv 2023
[2022]

ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–19

AvatarCLIP: Zero-Shot Text-Driven Generation and Animation of 3D Avatars. ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–19. Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM tran...

work page 2022
[2023]

In arXiv preprint arXiv:2307.15818

RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In arXiv preprint arXiv:2307.15818. Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. 1994. Animated Con- versation: Rule-Based Generation of Facial Expression, Gesture & Spoken Into...

arXiv 2024
[2024]

arXiv preprint arXiv:2412.15115 (2024)

Qwen2.5 Technical Report. arXiv preprint arXiv:2412.15115 (2024). Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. 2023. DiffuseStyleGesture: Stylized Audio-Driven Co-Speech Gesture Generation with Diffusion Models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intell...

arXiv 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [2]

MoMask: Generative Masked Modeling of 3D Human Motions. (2023). arXiv:2312.00063 [cs.CV] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. 2022. A Motion Matching-based Framework for Controllable Gesture Synthesis from Speech. In ACM SIGGRAPH 2022 Conference Proceedings (Van...

arXiv 2023

[2] [4]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Guided Motion Diffusion for Controllable Human Motion Synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2151– 2162. Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N. Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R. Thórisson, and Hannes Vilhjálmsson. 2006. Towards a Common Framework for Multimodal Genera...

arXiv 2023

[3] [2022]

ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–19

AvatarCLIP: Zero-Shot Text-Driven Generation and Animation of 3D Avatars. ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–19. Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM tran...

work page 2022

[4] [2023]

In arXiv preprint arXiv:2307.15818

RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In arXiv preprint arXiv:2307.15818. Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. 1994. Animated Con- versation: Rule-Based Generation of Facial Expression, Gesture & Spoken Into...

arXiv 2024

[5] [2024]

arXiv preprint arXiv:2412.15115 (2024)

Qwen2.5 Technical Report. arXiv preprint arXiv:2412.15115 (2024). Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. 2023. DiffuseStyleGesture: Stylized Audio-Driven Co-Speech Gesture Generation with Diffusion Models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intell...

arXiv 2024