Pith. sign in

REVIEW 4 major objections 5 minor 27 references

CoT-Pose: Chain-of-Thought Reasoning for 3D Pose Generation from Abstract Prompts

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read CoT-Pose claims that generating a detailed joint-level description before decoding discrete pose tokens makes 3D pose generation from abstract prompts more plausible and semantically aligned than direct decoding.

desk verdict The CoT-for-pose idea and the synthetic triplet pipeline are genuinely new, but the paper's own Table 3 undercuts the headline claim and the evaluation is too self-referential to rescue it. read the letter →

arxiv 2508.07540 v1 pith:3XMYIHH6 submitted 2025-08-11 cs.CV

Junuk Cha , Jihyeon Kim This is my paper · ORCID
classification cs.CV
keywords text-to-posegenerationchain-of-thoughtreasoningabstractprompts3DhumanposeSMPLmultimodallargelanguagemodelsyntheticdatapipelinediscretetokens
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

CoT-Pose is trying to establish that a text-to-pose model can handle the way humans actually describe movement—abstractly, as in "doing a handstand"—if it is forced to reason out loud. The proposed mechanism is a chain-of-thought stage in which the model first writes a detailed, joint-level description of the pose and only then decodes 80 discrete pose tokens into an SMPL mesh. To make this trainable, the paper builds an automatic pipeline that converts 550 action labels into (abstract prompt, detailed prompt, 3D pose) triplets using image generation, pose estimation, captioning, and chat-model refinement. On 50 abstract prompts, CoT-Pose reports the best pose-feature distance (0.6162) and mean per-joint error (124.91 mm) against PoseScript, ChatPose, and UniPose, and the ablations tie the gain to explicit reasoning supervision rather than pose decoding alone. A sympathetic reader would care because this targets a practical gap: previous pose generators required low-level joint instructions that few users would write.

What carries the argument

The central mechanism is the two-stage token decoding scheme: an autoregressive, causal-masked text decoder writes the detailed description, then a bidirectional pose decoder maps 80 learnable pose-query tokens to discrete pose tokens in a single step, conditioned on the full generated description. The combined objective $L = L_{\text{text}} + L_{\text{pose}}$ supervises both stages jointly inside one LoRA-fine-tuned transformer, so the pose decode is guided by explicit body-part semantics rather than only by the abstract prompt.

What would settle it

Take 100 action labels never appearing in the 550-label taxonomy (for example, labels from BABEL), obtain ground-truth poses from motion capture rather than SMPLest-X estimates, and run CoT-Pose alongside ChatPose and pose-only UniPose on the resulting abstract prompts. If CoT-Pose does not beat the direct decoders on pose-feature distance and per-joint error, or if its generated detailed descriptions misstate the ground-truth joint configuration, the claimed benefit of explicit chain-of-thought reasoning is refuted; the paper's Ablation 3 already predicts that pose-only training can win on ne

Watch

Extended reading notes

Core claim

CoT-Pose claims that semantic alignment with abstract prompts comes from making the model generate a detailed, joint-level description before producing the pose. Fine-tuning UniPose with $L = L_{\text{text}} + L_{\text{pose}}$ supervises both the autoregressive detailed description and the one-step decoding of 80 pose-query tokens conditioned on it. The paper reports this reasoning bridge beats direct prompt-to-pose decoding, and that removing the reasoning loss ($L_{\text{text}}$) still looks competitive on training prompts but fails on unseen ones. The synthesis pipeline that builds the training triplets—ChatGPT taxonomy, FLUX images, SMPLest-X poses, Pose2Text captions, ChatGPT refinement

Load-bearing premise

The load-bearing premise is that the 3D poses estimated by SMPLest-X from 50 online images are correct ground truth for the abstract actions, and that prompts drawn from the same 550-label taxonomy used to train the model are a fair test of abstract-prompt understanding—if either fails, the reported PFD and MPJPE figures do not measure semantic reasoning.

Editorial extensions

If this is right

  • Pose generation systems can accept abstract, user-facing prompts instead of requiring joint-angle specifications, removing a key usability barrier for animation and embodied-AI applications.
  • Explicit chain-of-thought supervision becomes a necessary component rather than a prompt trick: dropping the reasoning loss degrades semantic generalization even when training-prompt metrics look similar.
  • The automatic data-synthesis pipeline can generate aligned training triplets for new action vocabularies without manual annotation, making it feasible to scale beyond 550 action labels.
  • Models built on discrete pose tokens, such as UniPose, can be upgraded for abstract-prompt generation by fine-tuning with LoRA on synthesized triplets rather than redesigning the architecture.

Reading between the lines

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

  • Editorial inference: If the CoT bridge is the active ingredient, the same reason-then-decode formulation should transfer to other structured outputs with an abstract-to-geometric gap—hand poses, facial expressions, or object layouts—where a short verb hides a joint-level configuration.
  • Editorial inference: The evaluation uses prompts constructed from the same 550-label taxonomy used in training, with ground truth estimated by SMPLest-X rather than measured motion capture; a held-out test with action labels never seen in training and MoCap ground truth would separate genuine reasoning from memorization.
  • Editorial inference: A testable extension is to use the generated detailed prompts as editing handles—modify one clause, such as "right arm bent at 90 degrees," and re-decode—since the model already maps that text through the same pose decoder; the paper does not present this capability.
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

4 major / 5 minor

Summary. The paper proposes CoT-Pose, a method for generating 3D SMPL poses from abstract text prompts by fine-tuning UniPose with LoRA on automatically synthesized triplets of (abstract prompt, detailed prompt, 3D pose). The data synthesis pipeline uses FLUX to generate images, SMPLest-X to estimate 3D poses, Pose2Text to produce detailed descriptions, and ChatGPT to refine them. At inference, the model first autoregressively generates a detailed pose description (the chain-of-thought) and then decodes 80 discrete pose tokens from learnable pose queries. Experiments compare with PoseScript, ChatPose, and UniPose on 50 prompts and include ablations of the reasoning loss and prompt refinement. The paper claims state-of-the-art PFD and MPJPE, and argues that the reasoning objective is necessary for generalization to unseen prompts.

Significance. If the central claim were established, the paper would make a useful contribution: an automatic pipeline for constructing abstract-to-detailed-to-pose training data and a concrete instantiation of chain-of-thought reasoning for structured 3D pose generation. The idea is timely and the writing is clear. However, the paper's own quantitative evidence does not support the claim that the CoT reasoning objective is beneficial: the no-reasoning ablation outperforms the full model on the same benchmark. The evaluation is also self-referential and too small to support generalization claims. As presented, the significance of the method is not demonstrated.

major comments (4)
  1. [Section 4.6.2, Table 3] Ablation 3, trained only with the pose token generation objective (no L_text), achieves better PFD (0.5701 vs 0.6162) and MPJPE (115.14 vs 124.91) than the full CoT-Pose on the same real-image benchmark. This directly contradicts the central claim that explicit chain-of-thought reasoning improves pose generation. The paper dismisses the result as 'competitive' and appeals to Figure 3, but Figure 3 contains only four selected qualitative examples on unseen prompts, with no quantitative metric, no baseline comparison in the same figure, and no independent ground truth. Since the central contribution is the reasoning objective, the authors need to show on a properly held-out benchmark, with statistical testing, that the full model outperforms Ablation 3; otherwise the quantitative evidence points in the opposite direction.
  2. [Section 4.2, Section 4.6.2] The evaluation set is described as '50 real-world images from online sources, each paired with a corresponding abstract description used in the synthetic data generation,' and the ablation section explicitly states these are 'prompts seen during training.' The paper's stated motivation is generalization to abstract prompts and unseen wording, but the reported benchmark cannot measure that: the test prompts are drawn from the same 550-label taxonomy used to create training data. Evaluation on held-out prompts, ideally from a different source or constructed by independent annotators, is required to support the generalization claim.
  3. [Section 3.1, Section 4.2] The evaluation loop is circular. Training pose labels are SMPLest-X estimates from FLUX-generated images, and the 50 evaluation ground-truth poses are also SMPLest-X estimates from online images. The detailed prompts used in both training and evaluation are produced by Pose2Text followed by ChatGPT refinement, i.e., the same machinery. Therefore the reported PFD/MPJPE numbers may simply measure how well the model reproduces the synthesis/estimation pipeline, rather than how well it understands abstract action semantics. Independent ground truth (e.g., MoCap data, manual pose annotation, or at least a cross-check against a different pose estimator) is needed to validate the quantitative comparisons.
  4. [Section 4.3, Tables 1-3] All the reported comparisons use only 50 examples with no error bars, no confidence intervals, and no significance tests. The headline improvements are small; for example, Ours versus ChatPose gives PFD 0.6162 vs 0.6769 and MPJPE 124.91 vs 126.63. It is impossible to tell whether these differences are meaningful. In addition, the normalization note in Section 4.3 refers to 'pose FID' and 'text FID' although the reported metrics are PFD, TFD, and MFD; the text should be corrected and the scaling factors justified.
minor comments (5)
  1. [Figure 3] The qualitative evidence for unseen-prompt generalization is anecdotal: only four examples are shown, with no sampling protocol, no multiple random seeds, and no side-by-side baselines on the same prompts. The figure alone cannot compensate for the negative quantitative result in Table 3.
  2. [Section 4.2] The construction of the aligned dataset (3,374 PoseScript samples after deduplication, 239 synthesized samples) is described only briefly. Please clarify how BABEL abstracts, PoseScript details, and AMASS frames are aligned, how duplicates are identified, and what the manual filtering criteria are.
  3. [Figure 2 and Section 3.2] The role of the PQ tokens is not fully specified. Please state explicitly whether the 80 pose-query tokens are learned parameters or fixed special tokens, and clarify the notation '<SPQ>', '<EPQ>', and '<space>' in the token sequence diagram.
  4. [Section 4.6.2] Ablation 3 is presented only in a separate table and is not included in the main ablation table (Table 2). Including it in the same table would make the comparison across ablations clearer.
  5. [References] References [13] and [22] are to a model card/technical report and a GitHub repository, respectively; please provide versioned citations or release identifiers to support reproducibility.

Circularity Check

3 steps flagged · score 6.0 of 10

Quantitative validation loop is self-referential: evaluation ground truth comes from the same SMPLest-X/Pose2Text/ChatGPT pipeline that manufactured the training labels, and the paper's own Ablation 3 (no reasoning loss) outperforms the full CoT-Pose on that benchmark.

  1. other [Section 4.2 (Dataset), evaluation paragraph]
    "For evaluation, we collect a set of 50 real-world images from online sources, each paired with a corresponding abstract description used in the synthetic data generation. The 3D SMPL poses are estimated from these images using SMPLest-X [26], and subsequently passed to the Pose2Text [4] and ChatGPT [1] to generate and refine detailed prompts, as described in Section 3.1."

    The evaluation ground truth is generated by the same FLUX/SMPLest-X/Pose2Text/ChatGPT chain that manufactured the training triplets (Section 3.1). The only difference is that the 50 images are 'real-world' rather than FLUX renders; the abstract prompts are the same ones used in synthetic data generation and are admitted to be 'seen during training.' Therefore Table 1's PFD/MPJPE compare the model against the same estimator whose outputs are the training targets. By construction, a model that memorizes SMPLest-X's mapping from these prompts will score well; the benchmark cannot validate semantic understanding or generalization. This is a self-referential evaluation loop, not an independent test.

  2. other [Section 4.6.2 (Fine-tuning Strategy), after Table 3]
    "Remarkably, even without fine-tuning the text token generation loss Ltext (Ablation 3), which implies the absence of explicit reasoning supervision, the model achieves competitive quantitative results. However, as illustrated in Figure 3, it generates semantically inconsistent or implausible poses when evaluated on novel abstract prompts unseen during training."

    The paper's own quantitative comparison contradicts the central premise: Ablation 3 (no Ltext, no CoT supervision) has better PFD (0.5701 vs 0.6162) and MPJPE (115.14 vs 124.91) than the full CoT-Pose. The only counter-evidence is Figure 3, four qualitative examples on unseen prompts with no metric or systematic sampling. The manuscript itself labels Ablation 3's result 'competitive' and then asserts necessity of reasoning from examples. That asserted necessity is therefore unsupported by the quantitative benchmark; the claimed benefit of CoT reduces to anecdote on the paper's own numbers.

1 more flagged steps
  1. other [Section 4.6.2, parenthetical describing the benchmark]
    "using our collected real-image dataset with prompts seen during training"

    This sentence explicitly concedes that the quantitative benchmark (Tables 1-3) uses abstract prompts already present in training. Consequently, the reported improvements do not measure generalization to novel abstract language, which is what the abstract and introduction claim ('abstract textual inputs', 'generalization'). The 50-prompt evaluation cannot distinguish semantic reasoning from memorization of the same 550-label taxonomy; only the unquantified Figure 3 addresses unseen prompts.

full rationale

The derivation chain is: (1) ChatGPT builds a 550-label abstract-action taxonomy; FLUX renders images; SMPLest-X estimates SMPL poses; Pose2Text captions them; ChatGPT refines the captions into training triplets. (2) UniPose is fine-tuned with Ltext + Lpose to generate a detailed prompt and then 80 pose tokens from an abstract prompt. (3) Evaluation uses 50 online images whose abstract prompts come from the same taxonomy and are explicitly described as 'seen during training,' with ground-truth poses produced by the same SMPLest-X estimator and detailed prompts by the same Pose2Text+ChatGPT refinement used in training. Thus the benchmark is self-referential: the model is trained to reproduce a synthetic pipeline's outputs and then scored against that same pipeline's outputs on memorized prompts. Table 3 compounds the problem by showing that removing the reasoning objective improves both PFD and MPJPE on this benchmark, so the headline claim that CoT reasoning is beneficial is not supported by the paper's quantitative evidence; the paper falls back to four qualitative examples. There are no load-bearing self-citations by the paper's own authors, so the circularity is not a citation-chain issue but a partial validation-loop circularity: the quantitative 'prediction' reduces to agreement with the estimator that defined the training targets. Score 6 reflects that partial, by-construction circularity of the central quantitative claim.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The central claim rests on hand-set hyperparameters, a synthetic data pipeline whose semantic fidelity is assumed, and an evaluation protocol that reuses the same pose estimator and prompt taxonomy as training. There are no new physical entities; the SPQ/EPQ/PQ tokens are model-internal constructs rather than independent theoretical postulates.

free parameters (2)
  • LoRA and training hyperparameters = r=64, alpha=16, dropout=0.05, lr=5e-5, batch=8, 5 epochs
    Chosen by hand rather than derived; they define the fine-tuning regime that the central results depend on.
  • Taxonomy and dataset scale = 11 categories x 50 sub-actions = 550 labels; 3374 PoseScript samples + 239 synthetic samples; 50 evaluation prompts
    These hand-set scales control the training distribution and the evaluation set; the paper does not test sensitivity to them.
assumptions (5)
  • domain assumption SMPLest-X pose estimates are accurate enough to serve as pseudo ground truth for both training and evaluation.
    Section 3.1 uses SMPLest-X to label synthetic training poses, and Section 4.2 uses SMPLest-X to generate evaluation ground truth from 50 real images. No validation against motion capture or manual 3D annotation is reported.
  • domain assumption FLUX-generated images faithfully realize the abstract action labels.
    The image is the only semantic bridge between the abstract prompt and the estimated pose (Section 3.1); if the image depicts the wrong action or an unrealistic body, the training triplet is corrupted.
  • domain assumption ChatGPT refinement of Pose2Text descriptions preserves and improves pose-prompt alignment.
    Raw Pose2Text outputs are acknowledged to have orientation errors (Section 3.1); ChatGPT is used to fix them without an automatic consistency check between refined text and estimated joint positions.
  • domain assumption The UniPose/VQ-VAE discrete pose token space is expressive enough for the evaluated pose range.
    Sections 3.2 and 3.3 inherit the 80-token pose representation from UniPose; reconstruction fidelity on this dataset is not measured.
  • ad hoc to paper The 550-action taxonomy and the 50 evaluation prompts are representative of abstract user language.
    Evaluation prompts are drawn from the same taxonomy used to generate training data and are seen during training (Section 4.2, Table 3 caption), so generalization to novel user wording is not tested quantitatively.
invented entities (1)
  • SPQ, EPQ, and PQ pose-query/reasoning tokens
    purpose: Mark the start and end of pose queries and enable joint text and pose decoding in UniPose's discrete token space.
    These are architectural token delimiters adapted from UniPose's tokenization scheme, not independently falsifiable entities with predictions outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoT-Pose: Chain-of-Thought Reasoning for 3D Pose Generation from Abstract Prompts." pith.science (2026). https://pith.science/paper/3XMYIHH6

@misc{pith2026250807540,
  author       = {Pith},
  title        = {Pith review of: CoT-Pose: Chain-of-Thought Reasoning for 3D Pose Generation from Abstract Prompts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XMYIHH6}},
  note         = {Machine review of arXiv:2508.07540}
}
read the original abstract

Recent advances in multi-modal large language models (MLLMs) and chain-of-thought (CoT) reasoning have led to significant progress in image and text generation tasks. However, the field of 3D human pose generation still faces critical limitations. Most existing text-to-pose models rely heavily on detailed (low-level) prompts that explicitly describe joint configurations. In contrast, humans tend to communicate actions and intentions using abstract (high-level) language. This mismatch results in a practical challenge for deploying pose generation systems in real-world scenarios. To bridge this gap, we introduce a novel framework that incorporates CoT reasoning into the pose generation process, enabling the interpretation of abstract prompts into accurate 3D human poses. We further propose a data synthesis pipeline that automatically generates triplets of abstract prompts, detailed prompts, and corresponding 3D poses for training process. Experimental results demonstrate that our reasoning-enhanced model, CoT-Pose, can effectively generate plausible and semantically aligned poses from abstract textual inputs. This work highlights the importance of high-level understanding in pose generation and opens new directions for reasoning-enhanced approach for human pose generation.

Figures

Figures reproduced from arXiv: 2508.07540 by the authors.

Figure 1
Figure 1. From abstract to detailed prompts for 3D poses. We illustrate (a) abstract prompts, (b) the corresponding detailed prompts, and (c) the ground-truth 3D pose mesh aligned with the abstract and detailed prompts. This highlights our objective of rea￾soning detailed prompts from abstract ones to enable accurate 3D human pose generation. 3D human poses from text prompts. Some approaches [2, 9] rely on abstract prompts, s… view at source ↗
Figure 2
Figure 2. Overview of the proposed CoT-pose framework and examples. The top section presents the overall pipeline of CoT-Pose, which leverages large language models with chain-of-thought (CoT) reasoning to interpret abstract descriptions and generate semantically coherent 3D human poses. Given a prompt, the model first performs reasoning to decompose the abstract intent into detailed pose semantics, which are then used to gen… view at source ↗
Figure 3
Figure 3. Qualitative comparison of the generation results for detailed prompts and 3D poses from (a) Ablation 1 (w/o pose supervision), [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of data synthesis process for the action playing bass drum, using our synthesis pipeline based on FLUX [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of 3D pose generation results from (a) PoseScript [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 19 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  2. [2]

    Towards Better Adversarial Synthesis of Human Images from Text

    Rania Briq, Pratika Kochar, and Juergen Gall. Towards bet- ter adversarial synthesis of human images from text. arXiv preprint arXiv:2107.01869, 2021. 1

    work page Pith review arXiv 2021

  3. [3]

    Visual chain- of-thought prompting for knowledge-based visual reasoning

    Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Zhiqing Sun, Dan Gutfreund, and Chuang Gan. Visual chain- of-thought prompting for knowledge-based visual reasoning. In AAAI, 2024. 2

    work page 2024

  4. [4]

    Posescript: 3d human poses from natural language

    Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Gr ´egory Rogez. Posescript: 3d human poses from natural language. In ECCV, 2022. 1, 2, 3, 4, 5, 6, 7, 8

    work page 2022

  5. [5]

    Emerging properties in unified multimodal pretraining

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025. 2

    arXiv 2025

  6. [6]

    Tokenhmr: Advancing human mesh recov- ery with a tokenized pose representation

    Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J Black. Tokenhmr: Advancing human mesh recov- ery with a tokenized pose representation. In CVPR, 2024. 4

    work page 2024

  7. [7]

    Chatpose: Chatting about 3d human pose

    Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In CVPR, 2024. 1, 2, 3, 5, 6, 8

    work page 2024

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2

    arXiv 2025

Show all 27 references
  1. [9]

    Avatarclip: Zero-shot text- driven generation and animation of 3d avatars

    Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text- driven generation and animation of 3d avatars. ACM Trans- actions on Graphics (TOG), 2022. 1

    2022

  2. [10]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR,

  3. [11]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 2

    2024 arXiv

  4. [12]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. NIPS, 2022. 2

    2022

  5. [13]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arX...

    arXiv

  6. [14]

    Unipose: A unified multimodal framework for human pose comprehension, generation and editing

    Yiheng Li, Ruibing Hou, Hong Chang, Shiguang Shan, and Xilin Chen. Unipose: A unified multimodal framework for human pose comprehension, generation and editing. In CVPR, 2025. 1, 2, 3, 4, 5, 6, 8

    2025

  7. [15]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 2023. 4

    2023

  8. [16]

    Smpl: A skinned multi- person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi- person linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2. 2023. 2, 3, 4

    2023

  9. [17]

    Troje, Ger- ard Pons-Moll, and Michael J

    Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Ger- ard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 5

    2019

  10. [18]

    Huang, Joachim Tesch, David T

    Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regres- sion analysis. In CVPR, 2021. 2

    2021

  11. [19]

    Expressive body capture: 3d hands, face, and body from a single image

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 4

    2019

  12. [20]

    Babel: Bodies, action and behavior with english la- bels

    Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english la- bels. In CVPR, 2021. 5

    2021

  13. [21]

    Cogcom: A visual language model with chain-of- manipulations reasoning

    Ji Qi, Ming Ding, Weihan Wang, Yushi Bai, Qingsong Lv, Wenyi Hong, Bin Xu, Lei Hou, Juanzi Li, Yuxiao Dong, et al. Cogcom: A visual language model with chain-of- manipulations reasoning. In ICLR, 2025. 2

    2025

  14. [22]

    Flux-lora-dlc

    Prithiv Sakthi U R. Flux-lora-dlc. https://github. com/PRITHIVSAKTHIUR/Flux-LoRA-DLC , 2024. 3, 6, 7

    2024

  15. [23]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NIPS, 2017. 4

    2017

  16. [24]

    Black, Ivan Laptev, and Cordelia Schmid

    G ¨ul Varol, Javier Romero, Xavier Martin, Naureen Mah- mood, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In CVPR, 2017. 2

    2017

  17. [25]

    Chain-of-thought prompting elicits reasoning in large lan- guage models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan- guage models. NIPS, 2022. 2

    2022

  18. [26]

    Smplest-x: Ultimate scal- ing for expressive human pose and shape estimation

    Wanqi Yin, Zhongang Cai, Ruisi Wang, Ailing Zeng, Chen Wei, Qingping Sun, Haiyi Mei, Yanjun Wang, Hui En Pang, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Atsushi Ya- mashita, Lei Yang, and Ziwei Liu. Smplest-x: Ultimate scal- ing for expressive human pose and shape estimation...

    2025 arXiv

  19. [27]

    Least-to-most prompting enables complex reasoning in large language models

    Denny Zhou, Nathanael Sch ¨arli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 2

    2022 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.