Pith. sign in

REVIEW 4 major objections 7 minor 51 references

SparseCtrl-HOI: Sparse Temporal Control for Human-Object Interaction Video Generation

T0 review · 4 major / 7 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Few keyframes replace dense pose tracks for human-object video

desk verdict Sparse keyframe control for HOI video generation is a real idea; superiority claim is thin on the evidence read the letter →

arxiv 2607.05994 v1 pith:2DCFAS2N submitted 2026-07-07 cs.CV

Shenbo Xie , Mingrui Cai , Xu Yang , Yifei Liu , Changxing Ding This is my paper
classification cs.CV
keywords human-objectinteractionvideogenerationsparsetemporalcontrolrotarypositionalembeddingmotionpriorinjectiondiffusiontransformerkeyframe
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

This paper claims that Human-Object Interaction (HOI) video generation can be controlled precisely using only a handful of interaction keyframes rather than dense frame-by-frame pose annotations. The authors propose SparseCtrl-HOI, which pairs two mechanisms: TiRoPE (Time-Controlled Rotary Positional Embedding), which assigns each keyframe the temporal position of its target frame so the model knows exactly when each interaction state should appear, and a Motion Prior Injection Module, which uses a multimodal large language model to extract high-level descriptions of how hands and objects move between keyframes and feeds those descriptions into the video generator to fill in plausible intermediate motion. The paper also introduces SparseHOI-5K, a dataset of 4,850 live-commerce video clips with hand masks, object masks, and object-removed reference frames. The central claim is that this sparse paradigm reduces annotation cost, produces more natural and diverse motion than dense-control methods, and achieves superior or comparable visual quality on both the authors' dataset and an external benchmark.

What carries the argument

TiRoPE (Time-Controlled Rotary Positional Embedding): assigns keyframe latents the temporal positional index of their target frames within the DiT's rotary embedding space, directly aligning them with the frames they should influence. Motion Prior Injection Module: feeds interaction keyframes through a frozen MLLM (Qwen2.5-VL) to produce textual descriptions of hand-object pose changes, compresses these through a Q-Former into 64 learnable query tokens, and injects them into DiT blocks via dedicated cross-attention layers. Decoupled training: Stage 1 trains appearance transfer with LoRA on self-attention and text cross-attention; Stage 2 freezes Stage 1 parameters and trains only the Q-Form

What would settle it

If videos generated without the Motion Prior Injection Module were judged by human evaluators (not the Qwen2.5-VL-based HOI-VLM metric) to have equally natural transitions, the claimed necessity of MLLM-derived priors would be undermined.

Watch

Extended reading notes

Core claim

The paper's central technical contribution is showing that temporal anchoring of sparse keyframes can be achieved by directly assigning each keyframe the rotary positional index of its target video frame (TiRoPE), rather than requiring a learned mapping or dense interpolation. Combined with MLLM-derived motion priors compressed through a Q-Former and injected via cross-attention, this suffices to generate coherent HOI video with physically plausible transitions. The ablation shows that removing TiRoPE collapses timestamp alignment (Ti-SSIM drops from 0.66 to 0.52), and removing the Motion Prior Injection degrades temporal smoothness (MS-RAFT worsens from 0.45 to 0.53) and video quality (FVD从

Load-bearing premise

The same multimodal language model (Qwen2.5-VL) is used both to extract motion priors during generation and to evaluate output quality via the custom HOI-VLM metric, which could bias evaluation toward outputs consistent with that model's internal representations.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper introduces SparseCtrl-HOI, a framework for human-object interaction (HOI) video generation that replaces dense frame-wise guidance with a few sparse interaction keyframes. The core technical contributions are: (1) TiRoPE, which assigns temporal positional indices to keyframe latents so they align with designated output timestamps during DiT self-attention; (2) a Motion Prior Injection Module that uses Qwen2.5-VL to extract high-level motion descriptions from keyframes, compresses them via a Q-Former, and injects them through dedicated cross-attention layers; and (3) a two-stage decoupled training strategy. The authors also introduce SparseHOI-5K, a dataset of 4,850 live-streaming e-commerce clips with hand/object masks and object-free inpainted references. Evaluations are conducted on SparseHOI-5K (50 test clips, identity-disjoint) and the AnchorCrafter test set, with ablations on TiRoPE, Motion Prior Injection, condition injection design, and keyframe count.

Significance. The problem is well-motivated: dense pose-guided HOI video generation is annotation-heavy and produces rigid motion, and a sparse keyframe-driven alternative is a reasonable direction. The TiRoPE mechanism is a clean, parameter-free idea (assigning the target temporal RoPE index to keyframe latents), and the ablation in Table 3 and Supplementary Table 3 shows it substantially outperforms alternatives (additive bias, learnable embedding, concatenated positional embedding) on the Ti-SSIM metric (0.66 vs. 0.41–0.65). The Motion Prior Injection ablation shows FVD degradation from 629.6 to 651.2 when removed, and the interpretability analysis (Supplementary Fig. 1) provides useful evidence that the Q-Former attention concentrates on hand-object contact regions. The dataset contribution (SparseHOI-5K with multi-modal annotations including object-free inpainted videos) is a tangible community resource, and the authors commit to releasing code and data. The framework is built on established components (Wan2.1-DiT, Qwen2.5-VL, Q-Former), which aids reproducibility.

major comments (4)
  1. Section 4.4, Table 2 (AnchorCrafter dataset): The FID margin between SparseCtrl-HOI (91.4) and AnchorCrafter (91.9) is 0.5, and the FVD margin (725.8 vs. 726.4) is 0.6. These differences are well within the known noise floor of FID/FVD metrics, especially on a small test set (8 celebrity images, 10 source videos per Section 4.1). The paper states 'best performance across all metrics' on AnchorCrafter, but the evidence does not support a meaningful superiority claim. The authors should either (a) report confidence intervals or multiple random seeds to establish that these margins are statistically meaningful, or (b) soften the language to 'comparable' rather than 'superior.' This is load-bearing because the cross-dataset comparison is the only evaluation not on the authors' own dataset.
  2. Section 4.2 (Baselines): On the SparseHOI-5K dataset, AnchorCrafter is run without its required ±45° viewpoint object images—the authors substitute by 'duplicat[ing] the image of a single available object three times.' This places AnchorCrafter at a known disadvantage on the authors' own dataset, where the larger FID/FVD margins (89.1 vs. 258.4 FID) are reported. The paper should either (a) obtain the required multi-view object images for the SparseHOI-5K test set to run AnchorCrafter under its intended configuration, or (b) explicitly flag that the SparseHOI-5K comparison with AnchorCrafter is not apples-to-apples and avoid drawing superiority conclusions from it. The cross-dataset evaluation on AnchorCrafter's own test set (where inputs are presumably correct) is the fairer comparison, and there the margins are negligible (see comment above).
  3. Section 4.1 (HOI-VLM metric): The HOI-VLM metric uses Qwen2.5-VL to score physical plausibility and interaction diversity, but Qwen2.5-VL is also used in the Motion Prior Injection Module (Section 3.3) and in the dataset construction pipeline (Section 3.5, Phase III). This creates a risk of evaluator bias: the same model family that contributes to generation also judges output quality. On AnchorCrafter, the HOI-VLM scores are 3.00 (ours) vs. 2.90 (AnchorCrafter)—a 0.1-point difference on a 1–5 scale. The authors should either (a) use a different MLLM (e.g., GPT-4o or InternVL) for evaluation to break the circularity, (b) report inter-rater agreement between multiple MLLM judges, or (c) at minimum acknowledge this limitation and note that the HOI-VLM metric should be interpreted alongside the user study results (Supplementary Table 1), which show a larger margin (30.1% vs. 6.7% preference
  4. Section 4.1 (SparseHOI-5K test set): The test set contains only 50 clips. FID and FVD are known to be unstable at small sample sizes, and the paper's central quantitative claims on SparseHOI-5K (FID 89.1 vs. 129.3 for OmniAvatar-BL, FVD 629.6 vs. 1110.9) rest on this small set. The authors should report variance across multiple random subsets or seeds, or expand the test set, to establish that the reported margins are not artifacts of small-sample FID/FVD instability.
minor comments (7)
  1. Section 3.3: The prompt to Qwen2.5-VL is given in full, but the temperature, decoding strategy, and whether outputs are deterministic across runs are not specified. Since the Motion Prior Injection depends on these textual outputs, reproducibility would benefit from stating these parameters.
  2. Table 2: Sync-C scores are reported as '-' for AnchorCrafter, VACE, and Phantom on both datasets. If these models produce audio-aligned output, Sync-C should be computable; if not, a brief note explaining why it is omitted would be preferable to leaving it blank.
  3. Section 3.5, Phase III: The IoU threshold for excluding clips where 'the model may misidentify the subject as the object' is described qualitatively as 'excessive IoU' but the exact threshold value is not provided. Please state the threshold.
  4. Supplementary Table 2: The Q-Former ablation reports FID 86.86 for N=128 (better than the full model's 91.28) but FVD 640.50 (worse than 607.13). The text explains this as overfitting to frame-level details, but the FID improvement is somewhat at odds with the 'overfitting' framing. A brief clarification would help.
  5. Figure 2: The architecture diagram shows 'Audio' entering the pipeline, but the role of audio in the HOI generation task is not emphasized in the method sections (3.1–3.4). If audio is a required input (as implied by the OmniAvatar backbone and Sync-C metric), its contribution to the HOI generation pipeline should be clarified.
  6. Section 4.5 (Decoupled Training): The justification for the two-stage strategy is qualitative only ('standard quantitative metrics often fail to capture subtle foreground anomalies'). A quantitative comparison, even on MS-RAFT or a foreground-specific metric, would strengthen this claim. If none is available, the limitation of the qualitative-only evidence should be acknowledged.
  7. Reference [5] cites 'Nano Banana' as 'Gemini 2.5' (arXiv:2507.06261). The reference title and the tool name do not match, which may cause confusion. Please verify this is the correct citation for the privacy-preserving avatar reference image generation tool used.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises four major comments, all of which concern the rigor of our quantitative evaluation. We agree with the substance of each point and will revise the manuscript accordingly. Below we address each comment in turn.

read point-by-point responses

  1. Referee: Section 4.4, Table 2 (AnchorCrafter dataset): FID/FVD margins (0.5 and 0.6) are within noise floor on a small test set. 'Best performance across all metrics' claim is not supported.

    Authors: The referee is correct. The FID margin of 0.5 and FVD margin of 0.6 on the AnchorCrafter test set (8 celebrity images × 10 source videos) are well within the known instability range of these metrics at small sample sizes, and we should not have claimed 'best performance across all metrics' on this basis. We will make two changes: (1) We will re-run evaluation on the AnchorCrafter test set with 3 random seeds and report mean ± standard deviation for FID and FVD. (2) We will revise the language in Section 4.4 from 'best performance across all metrics' to 'comparable performance, with improvements on HOI-VLM and MS-RAFT' — the latter two metrics (3.00 vs. 2.90 on HOI-VLM; 0.27 vs. 0.43 on MS-RAFT) show larger and more interpretable margins, though we acknowledge the HOI-VLM concern separately (see below). We agree that the cross-dataset comparison should be framed as demonstrating competitive generalization rather than superiority. revision: yes

  2. Referee: Section 4.2 (Baselines): AnchorCrafter run without required ±45° viewpoint object images on SparseHOI-5K; single object image duplicated three times places it at a disadvantage.

    Authors: The referee is correct that this is not an apples-to-apples comparison. AnchorCrafter's pipeline is designed around multi-view object images (front, ±45°), and substituting three copies of a single view degrades its object-aware conditioning. We were transparent about this substitution in Section 4.2, but we did not adequately flag its implications for the comparison. We will revise the manuscript to: (1) Add an explicit caveat in Section 4.4 stating that the SparseHOI-5K comparison with AnchorCrafter is not under its intended configuration, and that the large FID/FVD margins on SparseHOI-5K should be interpreted with this caveat in mind. (2) Reframe the SparseHOI-5K results as demonstrating that our method is better suited to the sparse-keyframe, single-view-object setting, rather than claiming outright superiority over AnchorCrafter. (3) Emphasize that the fairer cross-method comparison is on the AnchorCrafter test set (where inputs follow the original configuration), where the results are comparable — consistent with our revised framing in the previous comment. We note that obtaining genuine multi-view object images for the diverse commodities in SparseHOI-5K (1,000+ products from live-streaming e-commerce) is practically infeasible, as these are not available in the source videos. We will state this explicitly. revision: yes

  3. Referee: Section 4.1 (HOI-VLM metric): Qwen2.5-VL used both in generation pipeline and as evaluator, creating circularity risk. 0.1-point difference on AnchorCrafter is not convincing.

    Authors: The referee raises a valid concern about evaluator bias. Qwen2.5-VL is used in three roles: (1) Motion Prior Injection (Section 3.3), (2) dataset interaction validation (Section 3.5, Phase III), and (3) the HOI-VLM evaluation metric. While the Motion Prior Injection uses Qwen2.5-VL to extract textual motion descriptions from keyframes (not to generate or judge video frames directly), we agree that using the same model family for both generation and evaluation introduces a potential circularity. We will address this as follows: (1) We will re-run the HOI-VLM evaluation on the AnchorCrafter test set using GPT-4o as an alternative judge and report both sets of scores. If the margins remain consistent, this will mitigate the circularity concern; if they change, we will report that honestly. (2) We will add an explicit limitation statement in Section 6 (Limitations) acknowledging the shared-model concern and recommending that HOI-VLM scores be interpreted alongside the user study results (Supplementary Table 1), which use human judges and show a larger preference margin (30.1% vs. 6.7% for AnchorCrafter). (3) We agree that the 0.1-point HOI-VLM difference on the AnchorCrafter test set is not by itself convincing and will not rely on it as a superiority claim. revision: partial

  4. Referee: Section 4.1 (SparseHOI-5K test set): Only 50 clips; FID/FVD unstable at small sample sizes. Report variance or expand test set.

    Authors: The referee is correct that 50 clips is small for FID/FVD, which are known to be sensitive to sample size. We will address this by: (1) Expanding the test set from 50 to 100 clips (drawing from the remaining identity-disjoint held-out pool) and re-running all SparseHOI-5K quantitative evaluations. (2) Reporting FID and FVD variance across 5 random 50-clip subsets of the expanded test set to demonstrate stability. (3) If the expanded test set reveals materially different margins, we will update the reported numbers and adjust our claims accordingly. We believe the core findings (substantial FID/FVD improvements over OmniAvatar-BL, VACE, and Phantom) will hold, as the current margins are large (e.g., FID 89.1 vs. 129.3; FVD 629.6 vs. 1110.9), but we commit to verifying this rigorously. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; derivation chain is self-contained.

full rationale

The paper's core derivation chain is not circular. TiRoPE (Section 3.2) is a parameter-free assignment of temporal positional indices to keyframe tokens—a straightforward architectural design with no self-referential dependency. The Motion Prior Injection Module (Section 3.3) uses Qwen2.5-VL (an external, third-party model) to extract textual motion descriptions from keyframes, compresses them via a Q-Former, and injects them into the DiT through cross-attention; the Q-Former and cross-attention layers are trained on ground-truth video data via flow matching (Eq. 1), not on the model's own outputs. The ablation studies (Table 3) remove components and measure degradation on held-out test data, which is standard methodology. The only minor concern is that Qwen2.5-VL is used both for motion prior extraction (Section 3.3) and for the HOI-VLM evaluation metric (Section 4.1), creating a potential evaluator bias. However, this is a correctness/evaluation concern, not circularity: the HOI-VLM score does not reduce to the model's training inputs by construction, and the paper also reports standard metrics (FID, FVD, VBench, Sync-C) that are computed independently. No self-citation chains are load-bearing for the central claims—the cited works (OmniAvatar [9], Wan2.1 [41], Qwen2.5-VL [38], Q-Former/BLIP-2 [22]) are external to the authors. Score 1 reflects the minor evaluator overlap, which does not rise to the level of a circular derivation.

Assumptions & free parameters 5 free parameters · 3 assumptions · 3 invented entities

The paper introduces three new entities (dataset, TiRoPE, Motion Prior Injection Module) and relies on three key axioms. The free parameters are mostly justified by ablation studies. The main concern is the axiom that Qwen2.5-VL reliably extracts accurate motion priors, which is only qualitatively demonstrated, and the axiom that decoupled training is superior, which lacks quantitative support.

free parameters (5)
  • Q-Former learnable queries (N=64) = 64
    Chosen by ablation (Supplementary Table 2) as optimal balance between FVD and temporal coherence.
  • Q-Former transformer depth (4 layers) = 4
    Selected by ablation (Supplementary Table 2) as optimal representational capacity.
  • Number of training keyframes (K=4) = 4
    Fixed for training; inference-time K is ablated (Table 3).
  • CFG scale = 3.0
    Selected empirically (Supplementary Table 5) as the most balanced configuration.
  • LoRA rank = Not specified
    Used in Stage 1 training but rank not stated in the paper.
assumptions (3)
  • domain assumption Standard 3D RoPE can be extended to assign arbitrary temporal indices to keyframe tokens without destabilizing the positional encoding space.
    Section 3.2 and Supplementary 5.1. The TiRoPE mechanism assumes that assigning the target frame's temporal index to a keyframe token will encourage high attention scores during self-attention, which is a reasonable but unproven assumption about the behavior of RoPE in this modified context.
  • domain assumption Qwen2.5-VL can reliably extract physically accurate motion priors from a sequence of keyframes.
    Section 3.3. The Motion Prior Injection Module depends on the MLLM correctly inferring hand-object pose changes, contact regions, and physical plausibility from sparse visual input. The paper provides one qualitative example (Supplementary Figure 2) but does not systematically verify the accuracy of these priors.
  • ad hoc to paper A two-stage decoupled training strategy is superior to joint training for this architecture.
    Section 3.4 and 4.5. The paper claims decoupled training prevents feature entanglement but only provides qualitative evidence (Figure 5), stating standard metrics fail to capture the difference.
invented entities (3)
  • SparseHOI-5K dataset independent evidence
    purpose: Training and evaluation dataset for sparse temporal control HOI video generation
    The dataset is a new constructed artifact with a described pipeline (Section 3.5). It will be publicly released, allowing independent verification of its properties.
  • TiRoPE (Time-Controlled Rotary Positional Embedding) independent evidence
    purpose: Temporal anchoring of sparse keyframes in DiT self-attention
    A new mechanism introduced by the paper. Its effectiveness is tested via ablation (Table 3, Supplementary Table 3) against alternatives, providing falsifiable evidence.
  • Motion Prior Injection Module independent evidence
    purpose: Compressing MLLM-derived motion semantics into tokens for DiT cross-attention
    A new architectural component. Its contribution is ablated (Table 3) and its interpretability is analyzed (Supplementary Figure 1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SparseCtrl-HOI: Sparse Temporal Control for Human-Object Interaction Video Generation." pith.science (2026). https://pith.science/paper/2DCFAS2N

@misc{pith2026260705994,
  author       = {Pith},
  title        = {Pith review of: SparseCtrl-HOI: Sparse Temporal Control for Human-Object Interaction Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DCFAS2N}},
  note         = {Machine review of arXiv:2607.05994}
}
read the original abstract

Human-Object Interaction (HOI) video generation aims to synthesize realistic videos of humans manipulating diverse objects, serving as a promising avenue for AI-driven live streaming e-commerce. A primary obstacle in this domain lies in the complexity of modeling fine-grained physical dynamics and the intricate spatial-temporal coordination between human hands and objects. Existing approaches to this problem typically rely on dense temporal guidance, e.g., frame-wise hand-object pose sequences, to strictly control the interaction process. However, such dense guidance incurs high annotation costs and affects motion synthesis diversity. To overcome these limitations, we introduce SparseCtrl-HOI, a novel sparse temporal control framework for HOI video generation. It requires only a few keyframes that capture interaction states at designated timestamps. Specifically, we employ a Time-Controlled Rotary Positional Embedding (TiRoPE) mechanism to temporally anchor these keyframes while preserving their spatial integrity. Subsequently, to govern the dynamics across intermediate frames, we propose a Motion Prior Injection Module that leverages Multimodal Large Language Models (MLLMs) to extract high-level motion priors. This empowers the model to hallucinate logically and physically plausible transitions. Furthermore, we build SparseHOI-5K, a high-quality and richly annotated dataset for HOI video generation with sparse temporal control. Comprehensive evaluations confirm that our method substantially reduces annotation overhead while synthesizing superior live-streaming e-commerce videos. Both our code and dataset are publicly available at https://mpi-lab.github.io/SparseCtrl-HOI.

Figures

Figures reproduced from arXiv: 2607.05994 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our SparseCtrl-HOI. We adopt a two-stage training strategy built upon the Wan2.1-DiT backbone. Given a reference image Iref, an audio sequence, and sparse interaction keyframes I 1 ho, . . . , IK ho at specified timestamps, the pipeline operates as follows. Motion Prior Encoding (Top): A frozen Qwen2.5-VL [38] extracts high-level motion semantics from the prompt and keyframes. These are then… view at source ↗
Figure 3
Figure 3. An overview of SparseHOI-5K dataset collection pipeline. the patch embedding and the audio pack [9], and employ Low-Rank Adaptation (LoRA) [15] on the pre-trained DiT blocks. Stage 2: Motion Transition Inference. This stage envisions physically plau￾sible motion transitions between two nearby interaction keyframes. We freeze all LoRA parameters optimized in Stage 1, thus preserving the visual synthesis abil￾ity of t… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overall qualitative comparison [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Qualitative ablations. 4.5 Ablation Study We perform ablation studies on SparseHOI-5K to justify the contribution of each proposed component. The results are detailed in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 1
Figure 1. Figure 1: Interpretability of the Motion Prior Injection (M.P.I.) module. Warmer colors in (a) denote higher attention; the maps in (b) are normalized to [0, 1] per frame. where d represents the feature dimension of the queries and keys. Motion-Prior Interpretability [PITH_FULL…
Figure 2
Figure 2. Figure 2: The four input interaction keyframes used to query Qwen2.5-VL for motion￾prior extraction. ### Keyframe 1: - **Hand Pose:** The left hand is holding the shoe with the thumb and index finger gripping the side near the heel. - **Object Pose:** The shoe is positioned hori…
Figure 3
Figure 3. Figure 3: A failure case of our work. scripts [3, 9] to batch-download the videos from their lists. To strictly adhere to privacy regulations, portrait rights, and platform copyright policies, we will only publicly release the exact Video IDs (VIDs) alongside our automated downl…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages

  1. [1]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , articleno =

    Huang, Ziyao and Zhou, Zixiang and Cao, Juan and Ma, Yifeng and Chen, Yi and Rao, Zejing and Xu, Zhiyong and Wang, Hongmei and Lin, Qin and Zhou, Yuan and Lu, Qinglin and Tang, Fan , title =. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , articleno =. 2025 , isbn =. doi:10.1145/3757377.3763861 , abstract =

    work page doi:10.1145/3757377.3763861 2025

  2. [2]

    Wang, Lizhen and Xia, Zhurong and Hu, Tianshu and Wang, Pengrui and Wei, Pengfei and Zheng, Zerong and Zhou, Ming and Zhang, Yuan and Gao, Mingyuan , journal=

    work page

  3. [3]

    Zhang, Wanyue and Foo, Lin Geng and Dabral, Rishabh and Beeler, Thabo and Theobalt, Christian , journal =

    work page

  4. [4]

    Liu, Bangya and Gong, Xinyu and Zhao, Zelin and Song, Ziyang and Lu, Yulei and Wu, Suhui and Zhang, Jun and Banerjee, Suman and Zhang, Hao , journal=

    work page

  5. [5]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Effective whole-body pose estimation with two-stages distillation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  6. [6]

    Chung, J. S. and Zisserman, A. Out of time: automated lip sync in the wild. Workshop on Multi-view Lip-reading, ACCV. 2016

    work page 2016

  7. [7]

    International Conference on Learning Representations , volume=

    Ravi, Nikhila and Gabeur, Valentin and Hu, Yuan-Ting and Hu, Ronghang and Ryali, Chaitanya and Ma, Tengyu and Khedr, Haitham and R. International Conference on Learning Representations , volume=

    work page

  8. [8]

    Zhou, Shangchen and Li, Chongyi and Chan, Kelvin C.K and Loy, Chen Change , booktitle=

    work page

Show all 51 references
  1. [9]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and Advanced Large-Scale Video Generative Models , author=. arXiv preprint arXiv:2503.20314 , year=

    arXiv

  2. [10]

    Edward J Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , url=

    2022

  3. [11]

    International Journal of Computer Vision , volume=

    A survey on human interaction motion generation , author=. International Journal of Computer Vision , volume=. 2026 , publisher=

    2026

  4. [12]

    arXiv preprint arXiv:2407.08428 , year=

    A Comprehensive Survey on Human Video Generation: Challenges, Methods, and Insights , author=. arXiv preprint arXiv:2407.08428 , year=

    arXiv

  5. [13]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Human motion generation: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  6. [14]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Animate anyone 2: High-fidelity character image animation with environment affordance , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  7. [15]

    Men, Yifang and Yao, Yuan and Cui, Miaomiao and Bo, Liefeng , booktitle=

  8. [16]

    Fan, Yingying and Yang, Quanwei and Wang, Kaisiyuan and Zhou, Hang and Li, Yingying and Feng, Haocheng and Ding, Errui and Wu, Yu and Wang, Jingdong , booktitle=

  9. [17]

    Pang, Youxin and Shao, Ruizhi and Zhang, Jiajun and Tu, Hanzhang and Liu, Yun and Zhou, Boyao and Zhang, Hongwen and Liu, Yebin , booktitle=

  10. [18]

    Xue, Zihui Sherry and Luo, Romy and Chen, Changan and Grauman, Kristen , journal=

  11. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , year =

    Hogue, Steven and Zhang, Chenxu and Daruger, Hamza and Tian, Yapeng and Guo, Xiaohu , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , year =

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Co-Speech Gesture Video Generation via Motion-Decoupled Diffusion Model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [21]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

    Democratizing High-Fidelity Co-Speech Gesture Video Generation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

  14. [22]

    2025 , url=

    Gaojie Lin and Jianwen Jiang and Chao Liang and Tianyun Zhong and Jiaqi Yang and Zerong Zheng and Yanbo Zheng , booktitle=. 2025 , url=

    2025

  15. [23]

    Qijun Gan and Ruizi Yang and Jianke Zhu and Shaofei Xue and Steven Hoi , journal=

  16. [24]

    2026 , publisher=

    Xu, Ziyi and Huang, Ziyao and Cao, Juan and Zhang, Yong and Cun, Xiaodong and Shuai, Qing and Wang, Yuchen and Bao, Linchao and Tang, Fan , journal=. 2026 , publisher=

    2026

  17. [25]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  18. [26]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    arXiv

  19. [27]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  20. [28]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    arXiv

  21. [29]

    2009 , howpublished =

    2009

  22. [30]

    2013 , howpublished =

    2013

  23. [31]

    Teed, Zachary and Deng, Jia , booktitle =

  24. [32]

    and Bjontegaard, Gisle and Luthra, Ajay , journal =

    Wiegand, Thomas and Sullivan, Gary J. and Bjontegaard, Gisle and Luthra, Ajay , journal =. Overview of the. 2003 , publisher =

    2003

  25. [33]

    2024 , howpublished =

    2024

  26. [34]

    2005 , howpublished =

    2005

  27. [35]

    2023 , organization=

    Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven , booktitle=. 2023 , organization=

    2023

  28. [36]

    Jiang, Zeyinzi and Han, Zhen and Mao, Chaojie and Zhang, Jingfeng and Pan, Yulin and Liu, Yu , booktitle=

  29. [37]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Phantom: Subject-consistent video generation via cross-modal alignment , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  30. [38]

    2019 , url=

    Unterthiner, Thomas and van Steenkiste, Sjoerd and Kurach, Karol and Marinier, Raphael and Michalski, Marcin and Gelly, Sylvain , booktitle=. 2019 , url=

    2019

  31. [39]

    2026 , publisher=

    Sun, Yasheng and Xu, Zhiliang and Zhou, Hang and Guan, Jiazhi and Yang, Quanwei and Wang, Kaisiyuan and Liang, Borong and Li, Yingying and Feng, Haocheng and Wang, Jingdong and others , journal=. 2026 , publisher=

    2026

  32. [40]

    Guan, Jiazhi and Wang, Kaisiyuan and Xu, Zhiliang and Yang, Quanwei and Sun, Yasheng and He, Shengyi and Liang, Borong and Cao, Yukang and Li, Yingying and Feng, Haocheng and others , booktitle=

  33. [41]

    Meng, Rang and Zhang, Xingyu and Li, Yuming and Ma, Chenguang , booktitle=

  34. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Kun Liu and Qi Liu and Xinchen Liu and Jie Li and Yongdong Zhang and Jiebo Luo and Xiaodong He and Wu Liu , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  35. [43]

    Huang, Ziqi and He, Yinan and Yu, Jiashuo and Zhang, Fan and Si, Chenyang and Jiang, Yuming and Zhang, Yuanhan and Wu, Tianxing and Jin, Qingyang and Chanpaisit, Nattapol and others , booktitle=

  36. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Corona, Enric and Zanfir, Andrei and Bazavan, Eduard Gabriel and Kolotouros, Nikos and Alldieck, Thiemo and Sminchisescu, Cristian , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  37. [45]

    Liang, Chao and Jiang, Jianwen and Liao, Wang and Yang, Jiaqi and Zeng, Weihong and Liang, Han and others , journal=

  38. [46]

    GitHub repository , howpublished =

    Castellano, Brandon , title =. GitHub repository , howpublished =. 2014 , publisher =

    2014

  39. [47]

    arXiv preprint arXiv:2006.01780 , year=

    A Novel Nudity Detection Algorithm for Web and Mobile Application Development , author=. arXiv preprint arXiv:2006.01780 , year=

    2006 arXiv

  40. [48]

    , title=

    Bradski, G. , title=. Dr. Dobb's Journal of Software Tools , year=

  41. [49]

    Lugaresi, Camillo and Tang, Jiuqiang and Nash, Hadon and McClanahan, Chris and Uboweja, Esha and Hays, Michael and Zhang, Fan and Chang, Chuo-Ling and Yong, Ming Guang and Lee, Juhyun and others , journal=

  42. [50]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Segment anything , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  43. [51]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.