Pith. sign in

REVIEW 5 major objections 10 minor 71 references

Joint continuous multi-scale human motion control and discrete contact commands can generate real-time upper-body videos where a person both moves and controllably changes nearby object states.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 20:24 UTC pith:KXL3LTR7

load-bearing objection Solid real-time upper-body HOI video system with a clean continuous/discrete control split; the “world model” label and same-pipeline synthetic eval are the main overreaches. the 5 major comments →

arxiv 2607.23517 v1 pith:KXL3LTR7 submitted 2026-07-26 cs.CV

Real-Time Human-Centric World Modeling for Upper-Body Human-Object Interaction

Chaonan Ji , Jinwei Qi , Peng Zhang , Bang Zhang This is my paper
classification cs.CV
keywords video generationworld modelhuman-object interactionupper-body animationmulti-scale motion latentdiscrete interaction statereal-time streaming
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most human video generators either animate a person in isolation or let free-form text vaguely describe interaction, so motion control, contact outcomes, and real-time use stay poorly joined. This paper argues that upper-body interactive generation is better cast as a small human-centric local world: continuous human state plus a discrete interaction state jointly determine the next frames. Human state is a single implicit latent built by encoding upper-body, hand, and face regions at multiple scales and fusing them without skeleton retargeting; interaction state is a short language command such as “no contact” or “grasp,” not an open prompt. Supervised on a dedicated synthetic interaction pipeline and distilled for streaming, the system runs at 25 FPS and reports better fine-grained motion fidelity and hand–object coordination than text-only or animation baselines. A sympathetic reader cares because this is a concrete path from pure motion replay toward interactive agents and telepresence that can actually pick things up or leave them alone on command.

Core claim

Person-centered upper-body scene evolution can be synthesized in real time by jointly conditioning a video generator on a unified multi-scale implicit human-state latent and a language-encoded discrete interaction-state command. The human state captures coordinated body, hand, and face dynamics without explicit retargeting; the interaction state switches contact relations such as no contact versus grasp under similar motion, so the model controls both how the person moves and how nearby objects change.

What carries the argument

Continuous–discrete joint control: a multi-scale implicit human state s_h formed by fusing upper-body, left-hand, right-hand, and face motion latents with reference features, plus a discrete interaction state s_o drawn from a tiny language-encoded set {no contact, grasp}, both fed to the generator as ŷ = G(x_ref, x_obj, s_h, s_o).

Load-bearing premise

That training and testing mostly on videos and labels produced by the same synthetic rendering pipeline, with only two scene-agnostic contact labels, is enough to claim realistic controllable hand–object coordination and local scene-state changes in the wild.

What would settle it

Run the same driving motion and object reference on held-out real (not pipeline-synthesized) upper-body videos with both grasp and no-contact labels: if hand–object contact, object identity, and user-controllable state switches collapse relative to the paper’s synthetic test set, the central claim does not hold outside that generative distribution.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Interactive agents and telepresence can switch grasp versus no-contact online without rewriting free-form prompts.
  • Fine hand and face motion can be driven cross-identity from multi-scale latents without skeleton retargeting errors.
  • Streaming 25 FPS generation on two H100-class GPUs becomes a practical target for person-centered local world models.
  • A small closed set of interaction-state commands can replace entangled text for contact-aware video control.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extending the discrete state set beyond two labels (place, push, release) is the natural next stress test of whether language-as-command stays stable.
  • If real-video fine-tuning fails to match synthetic held-out scores, the method may be more a controllable renderer of generative data than a world model of physical contact.
  • The same continuous–discrete split could transfer to egocentric or full-body settings where viewpoint and locomotion also change scene state.

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

5 major / 10 minor

Summary. The manuscript presents a video generation system for upper-body human–object interaction framed as a "human-centric local world model." Control is factored into (i) a continuous human-state latent built by multi-scale motion encoding — separate ViT encoders for upper body, left/right hand (shared weights), and face regions, fused and implicitly retargeted to the reference identity via an attention decoder (§3.2, Eqs. 3–4) — and (ii) a discrete interaction state s_o ∈ {no contact, grasp} encoded as fixed T5 text phrases (§3.3). Interaction supervision comes from a fully synthetic pipeline (Z-Image references, Qwen-VL object crops, Qwen3 prompts, Wan2.2-14B-I2V videos, VL filtering; ~30K labeled samples). Training is two-stage (§3.5): human-state pretraining on >3M upper-body videos with flow matching + RePA + keypoint auxiliary losses, then interaction learning with the motion encoder frozen. The model is distilled via CausVid + Self-Forcing to 4-step causal sampling, claiming 25 FPS streaming at 512×512 with ~1000 ms latency on two H100s. Experiments (Tab. 1) compare against Wan-Animate, Wan2.2-TI2V-5B, Humo, MAI, VACE on FVD/tLPIPS/CLIP-I/OQ/HandSIM/FaceSIM/IdentitySIM plus a user study.

Significance. If the claims hold, this is a useful systems contribution: streaming human-driven interaction generation at 25 FPS with ~1 s latency and explicit, switchable contact control is a practical step toward interactive avatars and telepresence, and most prior work offers either motion control or text prompting, not both. Specific strengths worth crediting: the discrete-state-as-command design is simple, scene-agnostic, and directly falsifiable via a compliance rate; the multi-scale implicit latent with a shared hand encoder and implicit retargeting is a coherent extension of prior latent-motion work; training is unusually well specified (λ_rep=0.1, λ_kp=10, stage step counts, GPU counts, 45-frame clips); and the promised release of the 30K state-labeled interaction dataset would benefit the community. The "world model" framing, however, currently exceeds what a two-state contact controller demonstrates.

major comments (5)
  1. [§4.2, Tab. 1] The 100-clip interaction test set is produced by the same generative pipeline (Z-Image references, Qwen-VL crops, Qwen3 prompts, Wan2.2-14B-I2V, VL filtering) that created the Stage-II training data. Disjoint identities/objects/prompts do not make this an independent sample: the FVD/tLPIPS/CLIP-I/OQ ground truth is Wan2.2-14B output, i.e., the teacher distribution the model was trained to match, while the baselines are not so matched. Gains may reflect distribution matching rather than more realistic interaction. Please add at least one off-pipeline evaluation: real upper-body HOI footage, held-out clips from a different video generator, or human/VLM contact-plausibility ratings on real reference images.
  2. [§4.1, Tab. 1] (i) HandSIM/FaceSIM are reported only against Wan-Animate ('–' for all other baselines); if these require motion-GT alignment that text-driven baselines cannot satisfy, state this and consider per-task tables. (ii) Each baseline's conditioning (driving video? object image? prompt source?) is unspecified, so the user-study axes TF/AF are not interpretable. (iii) The text claims 'favorable performance,' but MAI beats Ours on FVD (152.7 vs 154.1), tLPIPS (0.176 vs 0.178), and VQ (3.91 vs 3.82); this should be acknowledged rather than elided. (iv) The user study reports no rater count, clip count, protocol (Likert vs pairwise), or significance testing.
  3. [§4.4, Fig. 8] The discrete interaction state is the paper's central control contribution, yet its effect is shown only qualitatively in Fig. 8. A state-compliance rate (commanded s_o vs realized contact, using the authors' own Qwen-VL filter or human raters) over the test set would directly test the claim and is cheap to compute. Mid-stream state switching — the motivating real-time use case — is unquantified (frames-to-effect, failure modes). The 'grasp-by-default' bias of the w/o-text variant should also be quantified; the label balance of the 30K dataset is not reported.
  4. [§3.4, Tab. 1] The manuscript never states whether Tab. 1 reports the 20-step teacher or the 4-step distilled student. Since the headline claim couples real-time streaming with improved fidelity, the distillation cost must be measured: report FVD/tLPIPS/HandSIM before and after CausVid + Self-Forcing, plus resolution, sliding-window size, and how the two H100s are used (single stream vs sharded). If Tab. 1 is the teacher, the deployed system's quality at 25 FPS is currently unverified.
  5. [§4.2] The 100 reenactment driving videos come from SpeakerVid-5M, which also supplies Stage-I training clips, and no decontamination statement is given. If any test video or near-duplicate appears in training, the HandSIM/FaceSIM/IdentitySIM gains over Wan-Animate partly reflect memorization. Please confirm exclusion or re-evaluate on a disjoint motion source.
minor comments (10)
  1. [Title, §1, §5] Title/abstract/§5: the 'world model' framing is stronger than what is demonstrated — the interaction state has cardinality 2, and there is no persistence, memory, or scene-state evolution beyond contact switching. Either demonstrate additional scene states (e.g., object displacement persisting after release) or temper the framing to 'controllable interaction generation.'
  2. [§1] Second paragraph ends with a dangling citation list ('...contact and object manipulation. [Bhatnagar et al. 2022; ...]') not attached to any grammatical claim.
  3. [§4.1, Tab. 1] HandSIM/FaceSIM are defined as per-frame MSE distances to tracked coefficients but named 'SIM' (similarity) and tabulated with ↓; rename (e.g., HandMSE) or define the direction explicitly.
  4. [§4.1] OQ is a custom composite (product of object dynamics score and object DINO consistency). Justify the product form, report the two components separately, and give the metric's range; absolute differences (0.147 vs 0.151) are small and no variance is given.
  5. [§3.5] L_kp 'predicts target 2D keypoints from the implicit motion variable' — specify which tensor is decoded (z_b/z_lh/z_rh/z_f or s_h), which keypoints (body/hand/face), and which detector supplies supervision.
  6. [§3.4, Fig. 1] Give a breakdown of the ~1000 ms end-to-end latency (motion encoding, 4 DiT steps, VAE decode) and distinguish first-frame from steady-state latency; also reconcile '25 FPS on two H100 GPUs' (abstract/§3.4) with 'on H100 GPUs' (contributions).
  7. [Fig. 10 vs §3.3] Streaming demos use Gemini-generated references while the training pipeline uses Z-Image; comment on robustness to reference-image source, since references in-the-wild will differ from both.
  8. [§3.2, Eq. (4), Fig. 2] State the shape/token layout of s_h as injected into the DiT (Fig. 2 shows in-context tokens; Eq. 4 gives a spatial feature map), and how noised target tokens attend to it.
  9. [passim] Typographical: 'changes, Overall' (comma splice, §1); 'Discrete Interaction State' heading lacks a period (§4.4); inconsistent 'Fig.7'/'Fig. 8' spacing; abstract rendering of the emphasized states lost spacing ('asno contactorgrasp').
  10. [§3.3] The dataset release promise (§3.3) is welcome; please also state whether code, the distilled model, and the evaluation splits will be released, and report the 30K dataset's grasp/no-contact label balance.

Circularity Check

0 steps flagged

No derivation circularity: empirical video-generation method with standard train/eval, not a self-reducing prediction chain.

full rationale

This is a systems/method paper for controllable upper-body human–object video generation. Its load-bearing claims are architectural and empirical (multi-scale implicit human-state latents fused from body/hands/face; language-encoded discrete interaction commands {no contact, grasp}; two-stage training; CausVid/Self-Forcing distillation to 25 FPS), not first-principles derivations or parameter fits renamed as predictions. Equations (1)–(4) define an encoder–generator pipeline s_h = E(d), ŷ = G(x_ref, x_obj, s_h, s_o) and a multi-scale fusion s_h = D(f_ref, z_b, z_lh, z_rh, z_f); they do not equate a reported metric to a free parameter by construction. Stage-I/II objectives (L_FM + auxiliaries, then L_FM) are ordinary supervised/flow-matching losses. Baselines (Wan-Animate, Wan2.2, Humo, MAI, VACE) and metrics (FVD, tLPIPS, CLIP-I, HandSIM, FaceSIM, user study) are external comparisons, not self-justifying uniqueness theorems. Overlap with Wan-Animate authorship is normal prior-work citation and does not force the central continuous–discrete control claim. Training and held-out evaluation both use the authors’ synthetic HOI pipeline (§3.3, §4.2) is a generalization/validity risk, not circularity in the derivation sense: no quoted step reduces a “prediction” to its fitted input. No steps match the enumerated circularity kinds.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

Load-bearing content is mostly engineering assumptions and training choices, not physical axioms. The claim rests on: (i) adequacy of a 4-region implicit motion bottleneck plus frozen-T5 discrete phrases as a complete control interface; (ii) synthetic I2V supervision as a proxy for real contact; (iii) standard flow-matching DiT generation and known distillation methods behaving as in cited work. Free parameters are loss weights and train/distill schedules. Invented entities are framing constructs (human-centric local world model, continuous–discrete joint control) rather than new physical objects.

free parameters (5)
  • lambda_rep (RePA loss weight) = 0.1
    Set to 0.1 in Stage I to align DiT features with DINOv3; chosen for training stability, not derived.
  • lambda_kp (keypoint loss weight) = 10
    Set to 10 for geometric guidance via lightweight keypoint decoder in Stage I.
  • Stage I / II step counts and data mix = 100K then 10K steps
    100K steps on ~3M motion clips (128 A100) then 10K steps on ~30K interaction clips (32 A100) with frozen motion encoder; schedule is empirical.
  • Distilled sampling steps and windowing = 4 steps; 25 FPS on 2×H100 at 512²
    CausVid reduces sampling from 20 to 4 steps with causal/sliding-window attention; Self-Forcing settings affect long-horizon stability and the reported 25 FPS / ~1000 ms latency.
  • Motion latent bottleneck dimension
    Low-dimensional 1D latents per region justified by information-bottleneck identity leakage reduction; exact size not derived from theory.
axioms (6)
  • ad hoc to paper Upper-body interactive generation is adequately factored into a continuous human-state latent s_h = E(d) and a discrete interaction command s_o ∈ {no contact, grasp} with ŷ = G(x_ref, x_obj, s_h, s_o).
    Problem formulation §3.1; defines the paper's control interface rather than a proved completeness result over real HOI.
  • domain assumption Multi-scale region crops (body, left/right hand, face via MediaPipe) encoded by separate ViTs and fused with reference patches capture coordinated fine-grained dynamics better than a single global latent, without explicit retargeting.
    §3.2 and ablation Fig. 7; standard CV assumption extended from cited implicit motion work (e.g., X-UniMotion).
  • ad hoc to paper Language-encoded fixed phrases are sufficient interaction-state commands and are not entangled open-ended semantic prompts.
    §3.3; motivates discreteness from grasp vs non-contact under similar motion, but vocabulary size is chosen by authors.
  • domain assumption Synthetic videos from Z-Image + Qwen-VL/Qwen3 + Wan2.2-14B-I2V with VL filtering provide valid supervision for discrete contact control.
    §3.3 data pipeline; central to Stage II and interaction evaluation.
  • domain assumption Flow-matching DiT training plus CausVid/Self-Forcing distillation preserves interaction controllability at 4-step streaming inference.
    §3.4–3.5; relies on cited distillation literature behaving as claimed under HOI conditioning.
  • standard math Standard deep learning / transformer / flow-matching optimization mathematics holds for the reported objectives L_FM, L_RePA, L_kp.
    Training objectives in §3.5; routine.
invented entities (3)
  • Human-centric local world model (upper-body interactive generation framing) no independent evidence
    purpose: Positions the system as modeling person-centered local scene evolution rather than isolated reenactment or general world simulation.
    Introduced in §1 and Abstract; operational content is still conditional video generation with two contact flags—no independent physics or persistent world state beyond the video model.
  • Continuous–discrete joint control scheme (unified multi-scale human state + language discrete interaction state) no independent evidence
    purpose: Separates motion control from contact control for online state switching and finer HOI generation.
    Core method object in §3; evidenced only by this paper's ablations and baselines, not an external construct with independent measurements.
  • Dedicated HOI rendering/supervision pipeline with aligned discrete state labels (~30K samples) no independent evidence
    purpose: Supplies Stage-II training signal for grasp vs no-contact under matched motion-like conditions.
    §3.3; valuable artifact if released, but currently author-constructed synthetic distribution.

pith-pipeline@v1.2.0-grok45-kimik3 · 24811 in / 4739 out tokens · 99624 ms · 2026-07-30T20:24:23.750467+00:00 · methodology

0 comments
read the original abstract

We present a real-time human-centric world model for upper-body interactive generation, aiming to synthesize coherent local world dynamics centered on a person, where coordinated body, hand, and facial motions evolve jointly with controllable human-object discrete interaction. To this end, we adopt a continuous-discrete joint control scheme with two complementary components: a continuous human state and a discrete interaction state. For continuous human-state control, we introduce a unified implicit representation based on multi-scale motion encoding, in which motion latents from the upper body, hands, and face are fused into a shared latent space. This multi-scale design improves expressiveness across different spatial scales, captures fine-grained human dynamics more effectively, and enables direct control without explicit retargeting. For discrete object interaction-state control, we represent object contact using a small set of language-encoded discrete interaction states, where text serves as an explicit interaction-state command, such as \emph{no contact} or \emph{grasp}, rather than an open-ended generation prompt, and we further construct a dedicated rendering pipeline for human-object interaction data to supervise such discrete interaction states. By combining continuous implicit human-state control with discrete interaction-state control, our model enables precise modeling of how a person moves and interacts with the local environment, including controllable changes to nearby scene states. Finally, we distill the model for efficient streaming real-time inference, achieving 25 FPS on two H100 GPUs. Experiments demonstrate improved fine-grained motion fidelity, more realistic hand-object coordination, and effective real-time interaction, establishing a practical step beyond motion reproduction toward real-time human-centric world modeling.

Figures

Figures reproduced from arXiv: 2607.23517 by Bang Zhang, Chaonan Ji, Jinwei Qi, Peng Zhang.

Figure 1
Figure 1. Figure 1: Visual results. Conditioned on a reference image, an object image, a driving input, and a discrete interaction-state command, our method synthesizes [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An overview of our framework. Given a reference image, an object reference image, a driving signal, and a discrete interaction-state command, the Human State module captures the person’s multi-scale implicit motion states, while the Interaction State module models the human–object contact states. These two states are then jointly fed into DiT to synthesize the final result. crop the corresponding image reg… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of Human State generation, which decomposes the driving [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pipeline for constructing a high-quality human–object interaction [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison with state-of-the-art methods. Please refer to the supplementary material for more results. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study of the multi-scale Human State representation.. w/o multi-scale does not inject local hand and facial regions, and w/ multi￾scale is Human State modeling. Red boxes mark local regions. Reference Driving w/o text w/ text State=No contact State=No contact [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Human State control. Compared with Wan-Animate, our method achieves more accurate cross-identity driving and better local motion control, especially for complex hand movements. MAI sometimes produces unstable contact, such as floating objects or implausible grasps. Wan2.2 also struggles to preserve object iden￾tity consistently. In addition, these baselines mainly rely on text￾based conditioning, making pr… view at source ↗
Figure 9
Figure 9. Figure 9: Extra visual results. , Vol. 1, No. 1, Article . Publication date: July 2026 [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional results of streaming real-time driving. Given live driving input, our method generates temporally coherent upper-body interaction videos [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 4 canonical work pages

  1. [1]

    Crandall and Chen Yu , title =

    Sven Bambach and Stefan Lee and David J. Crandall and Chen Yu , title =. 2015. 2015 , url =. doi:10.1109/ICCV.2015.226 , timestamp =

    work page doi:10.1109/iccv.2015.226 2015

  2. [3]

    Arpit Mittal and Andrew Zisserman and Philip H. S. Torr , editor =. Hand detection using multiple proposals , booktitle =. 2011 , url =. doi:10.5244/C.25.75 , timestamp =

    work page doi:10.5244/c.25.75 2011

  3. [6]

    GanHand: Predicting Human Grasp Affordances in Multi-Object Scenes , booktitle =

    Enric Corona and Albert Pumarola and Guillem Aleny. GanHand: Predicting Human Grasp Affordances in Multi-Object Scenes , booktitle =. 2020 , url =. doi:10.1109/CVPR42600.2020.00508 , timestamp =

    arXiv 2020

  4. [7]

    Black and Krikamol Muandet and Siyu Tang , editor =

    Korrawe Karunratanakul and Jinlong Yang and Yan Zhang and Michael J. Black and Krikamol Muandet and Siyu Tang , editor =. Grasping Field: Learning Implicit Representations for Human Grasps , booktitle =. 2020 , url =. doi:10.1109/3DV50981.2020.00043 , timestamp =

    arXiv 2020

  5. [8]

    Computer Vision -

    Jiaxin Lu and Hao Kang and Haoxiang Li and Bo Liu and Yiding Yang and Qixing Huang and Gang Hua , editor =. Computer Vision -. 2024 , url =. doi:10.1007/978-3-031-72855-6\_24 , timestamp =

    doi:10.1007/978-3-031-72855-6 2024

  6. [10]

    2023 , url =

    Yufei Ye and Xueting Li and Abhinav Gupta and Shalini De Mello and Stan Birchfield and Jiaming Song and Shubham Tulsiani and Sifei Liu , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.02153 , timestamp =

    arXiv 2023

  7. [11]

    2024 , url =

    Mengqi Zhang and Yang Fu and Zheng Ding and Sifei Liu and Zhuowen Tu and Xiaolong Wang , title =. 2024 , url =. doi:10.1109/CVPR52733.2024.00814 , timestamp =

    arXiv 2024

  8. [13]

    Karen Liu , editor =

    Jiaman Li and Alexander Clegg and Roozbeh Mottaghi and Jiajun Wu and Xavier Puig and C. Karen Liu , editor =. Controllable Human-Object Interaction Synthesis , booktitle =. 2024 , url =. doi:10.1007/978-3-031-72940-9\_4 , timestamp =

    work page doi:10.1007/978-3-031-72940-9 2024

  9. [14]

    Grasp as You Say: Language-guided Dexterous Grasp Generation , booktitle =

    Yi. Grasp as You Say: Language-guided Dexterous Grasp Generation , booktitle =. 2024 , url =

    2024

  10. [16]

    CoRR , volume =

    Jianzhu Guo and Dingyun Zhang and Xiaoqiang Liu and Zhizhou Zhong and Yuan Zhang and Pengfei Wan and Di Zhang , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2407.03168 , eprinttype =. 2407.03168 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.03168 2024

  11. [17]

    The Thirteenth International Conference on Learning Representations,

    Xiaochen Zhao and Hongyi Xu and Guoxian Song and You Xie and Chenxu Zhang and Xiu Li and Linjie Luo and Jinli Suo and Yebin Liu , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  12. [18]

    Follow-Your-Emoji: Fine-Controllable and Expressive Freestyle Portrait Animation , booktitle =

    Yue Ma and Hongyu Liu and Hongfa Wang and Heng Pan and Yingqing He and Junkun Yuan and Ailing Zeng and Chengfei Cai and Heung. Follow-Your-Emoji: Fine-Controllable and Expressive Freestyle Portrait Animation , booktitle =. 2024 , url =. doi:10.1145/3680528.3687587 , timestamp =

    arXiv 2024

  13. [19]

    X-UniMotion: Animating Human Images with Expressive, Unified and Identity-Agnostic Motion Latents , booktitle =

    Guoxian Song and Hongyi Xu and Xiaochen Zhao and You Xie and Tianpei Gu and Zenan Li and Chenxu Zhang and Linjie Luo , editor =. X-UniMotion: Animating Human Images with Expressive, Unified and Identity-Agnostic Motion Latents , booktitle =. 2025 , url =. doi:10.1145/3757377.3763952 , timestamp =

    arXiv 2025

  14. [20]

    CoRR , volume =

    Li Hu and Guangyuan Wang and Zhen Shen and Xin Gao and Dechao Meng and Lian Zhuo and Peng Zhang and Bang Zhang and Liefeng Bo , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.06145 , eprinttype =. 2502.06145 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.06145 2025

  15. [21]

    2025 , eprint=

    Wan-Animate: Unified Character Animation and Replacement with Holistic Replication , author=. 2025 , eprint=

    2025

  16. [22]

    2026 , eprint=

    Kling-MotionControl Technical Report , author=. 2026 , eprint=

    2026

  17. [23]

    IM-Animation: An Implicit Motion Representation for Identity-decoupled Character Animation , journal =

    Zhufeng Xu and Xuan Gao and Feng. IM-Animation: An Implicit Motion Representation for Identity-decoupled Character Animation , journal =. 2026 , url =. doi:10.48550/ARXIV.2602.07498 , eprinttype =. 2602.07498 , timestamp =

    work page doi:10.48550/arxiv.2602.07498 2026

  18. [24]

    CoRR , volume =

    Zhixue Fang and Xu He and Songlin Tang and Haoxian Zhang and Qingfeng Li and Xiaoqiang Liu and Pengfei Wan and Kun Gai , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.03796 , eprinttype =. 2602.03796 , timestamp =

    doi:10.48550/arxiv.2602.03796 2026

  19. [25]

    CoRR , volume =

    Lizhen Wang and Yongming Zhu and Zhipeng Ge and Youwei Zheng and Longhao Zhang and Tianshu Hu and Shiyang Qin and Mingshuang Luo and Jiaxu Zhang and Xin Chen and Yulong Wang and Zerong Zheng and Jianwen Jiang and Chao Liang and Weifeng Chen and Xing Wang and Yuan Zhang and Mingyuan Gao , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2601.1010...

    doi:10.48550/arxiv.2601.10103 2026

  20. [26]

    CoRR , volume =

    Yutong Bai and Danny Tran and Amir Bar and Yann LeCun and Trevor Darrell and Jitendra Malik , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.21552 , eprinttype =. 2506.21552 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.21552 2025

  21. [27]

    2025 , eprint=

    Re-HOLD: Video Hand Object Interaction Reenactment via adaptive Layout-instructed Diffusion Model , author=. 2025 , eprint=

    2025

  22. [28]

    Dennis and Ashley Edwards and Jack Parker

    Jake Bruce and Michael D. Dennis and Ashley Edwards and Jack Parker. Genie: Generative Interactive Environments , booktitle =. 2024 , url =

    2024

  23. [29]

    Matrix-Game 2.0: An Open-Source, Real-Time, and Streaming Interactive World Model , journal =

    Xianglong He and Chunli Peng and Zexiang Liu and Boyang Wang and Yifan Zhang and Qi Cui and Fei Kang and Biao Jiang and Mengyin An and Yangyang Ren and Baixin Xu and Hao. Matrix-Game 2.0: An Open-Source, Real-Time, and Streaming Interactive World Model , journal =. 2025 , url =. doi:10.48550/ARXIV.2508.13009 , eprinttype =. 2508.13009 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.13009 2025

  24. [30]

    CoRR , volume =

    Jiaqi Li and Junshu Tang and Zhiyong Xu and Longhuang Wu and Yuan Zhou and Shuai Shao and Tianbao Yu and Zhiguo Cao and Qinglin Lu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.17201 , eprinttype =. 2506.17201 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.17201 2025

  25. [31]

    CoRR , volume =

    Yuanpeng Tu and Hao Luo and Xi Chen and Xiang Bai and Fan Wang and Hengshuang Zhao , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.09995 , eprinttype =. 2506.09995 , timestamp =

    doi:10.48550/arxiv.2506.09995 2025

  26. [32]

    CoRR , volume =

    Hao He and Yinghao Xu and Yuwei Guo and Gordon Wetzstein and Bo Dai and Hongsheng Li and Ceyuan Yang , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2404.02101 , eprinttype =. 2404.02101 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.02101 2024

  27. [33]

    CoRR , volume =

    Zeyinzi Jiang and Zhen Han and Chaojie Mao and Jingfeng Zhang and Yulin Pan and Yu Liu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2503.07598 , eprinttype =. 2503.07598 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.07598 2025

  28. [34]

    2026 , eprint=

    LTX-2: Efficient Joint Audio-Visual Foundation Model , author=. 2026 , eprint=

    2026

  29. [35]

    2025 , eprint=

    HunyuanVideo: A Systematic Framework For Large Video Generative Models , author=. 2025 , eprint=

    2025

  30. [36]

    CoRR , volume =

    Youliang Zhang and Zhengguang Zhou and Zhentao Yu and Ziyao Huang and Teng Hu and Sen Liang and Guozhen Zhang and Ziqiao Peng and Shunkai Li and Yi Chen and Zixiang Zhou and Yuan Zhou and Qinglin Lu and Xiu Li , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.01538 , eprinttype =. 2602.01538 , timestamp =

    doi:10.48550/arxiv.2602.01538 2026

  31. [37]

    The Thirteenth International Conference on Learning Representations,

    Zhuoyi Yang and Jiayan Teng and Wendi Zheng and Ming Ding and Shiyu Huang and Jiazheng Xu and Yuanming Yang and Wenyi Hong and Xiaohan Zhang and Guanyu Feng and Da Yin and Yuxuan Zhang and Weihan Wang and Yean Cheng and Bin Xu and Xiaotao Gu and Yuxiao Dong and Jie Tang , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  32. [38]

    2025 , url =

    Amir Bar and Gaoyue Zhou and Danny Tran and Trevor Darrell and Yann LeCun , title =. 2025 , url =. doi:10.1109/CVPR52734.2025.01472 , timestamp =

    arXiv 2025

  33. [39]

    CoRR , volume =

    Ziyi Xu and Ziyao Huang and Juan Cao and Yong Zhang and Xiaodong Cun and Qing Shuai and Yuchen Wang and Linchao Bao and Jintao Li and Fan Tang , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2411.17383 , eprinttype =. 2411.17383 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.17383 2024

  34. [40]

    CoRR , volume =

    Zhenzhi Wang and Jiaqi Yang and Jianwen Jiang and Chao Liang and Gaojie Lin and Zerong Zheng and Ceyuan Yang and Dahua Lin , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.09984 , eprinttype =. 2506.09984 , timestamp =

    doi:10.48550/arxiv.2506.09984 2025

  35. [41]

    CoRR , volume =

    Liyang Chen and Tianxiang Ma and Jiawei Liu and Bingchuan Li and Zhuowei Chen and Lijie Liu and Xu He and Gen Li and Qian He and Zhiyong Wu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2509.08519 , eprinttype =. 2509.08519 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.08519 2025

  36. [42]

    2026 , eprint=

    Advancing Open-source World Models , author=. 2026 , eprint=

    2026

  37. [43]

    CoRR , volume =

    Tianxing Xu and Zixuan Wang and Guangyuan Wang and Li Hu and Zhongyi Zhang and Peng Zhang and Bang Zhang and Song. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.22960 , eprinttype =. 2602.22960 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2602.22960 2026

  38. [44]

    2026 , eprint=

    PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation , author=. 2026 , eprint=

    2026

  39. [45]

    2025 , eprint=

    From Slow Bidirectional to Fast Autoregressive Video Diffusion Models , author=. 2025 , eprint=

    2025

  40. [46]

    2024 , eprint=

    Improved Distribution Matching Distillation for Fast Image Synthesis , author=. 2024 , eprint=

    2024

  41. [47]

    2025 , eprint=

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion , author=. 2025 , eprint=

    2025

  42. [48]

    2026 , eprint=

    Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation , author=. 2026 , eprint=

    2026

  43. [49]

    First Order Motion Model for Image Animation , booktitle =

    Aliaksandr Siarohin and St. First Order Motion Model for Image Animation , booktitle =. 2019 , url =

    2019

  44. [50]

    International Conference on 3D Vision (3DV) , year=

    Collaborative Regression of Expressive Bodies using Moderation , author=. International Conference on 3D Vision (3DV) , year=

  45. [51]

    T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models , booktitle =

    Chong Mou and Xintao Wang and Liangbin Xie and Yanze Wu and Jian Zhang and Zhongang Qi and Ying Shan , editor =. T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models , booktitle =. 2024 , url =. doi:10.1609/AAAI.V38I5.28226 , timestamp =

    doi:10.1609/aaai.v38i5.28226 2024

  46. [54]

    2018 , eprint=

    Learning to Detect Human-Object Interactions , author=. 2018 , eprint=

    2018

  47. [55]

    Petrov and Cristian Sminchisescu and Christian Theobalt and Gerard Pons

    Bharat Lal Bhatnagar and Xianghui Xie and Ilya A. Petrov and Cristian Sminchisescu and Christian Theobalt and Gerard Pons. 2022 , url =. doi:10.1109/CVPR52688.2022.01547 , timestamp =

    arXiv 2022

  48. [56]

    Computer Vision -

    Xianghui Xie and Bharat Lal Bhatnagar and Gerard Pons. Computer Vision -. 2022 , url =. doi:10.1007/978-3-031-20086-1\_8 , timestamp =

    doi:10.1007/978-3-031-20086-1 2022

  49. [57]

    DexMV: Imitation Learning for Dexterous Manipulation from Human Videos , booktitle =

    Yuzhe Qin and Yueh. DexMV: Imitation Learning for Dexterous Manipulation from Human Videos , booktitle =. 2022 , url =. doi:10.1007/978-3-031-19842-7\_33 , timestamp =

    doi:10.1007/978-3-031-19842-7 2022

  50. [58]

    Yu. DexYCB:. 2021 , url =. doi:10.1109/CVPR46437.2021.00893 , timestamp =

    arXiv 2021

  51. [59]

    Guibas and Mustafa Mukadam and Abhinav Gupta and Shubham Tulsiani , title =

    Kaichun Mo and Leonidas J. Guibas and Mustafa Mukadam and Abhinav Gupta and Shubham Tulsiani , title =. 2021. 2021 , url =. doi:10.1109/ICCV48922.2021.00674 , timestamp =

    arXiv 2021

  52. [60]

    2018 , copyright =

    Ha, David and Schmidhuber, Jürgen , title =. 2018 , copyright =. doi:10.5281/ZENODO.1207631 , url =

    doi:10.5281/zenodo.1207631 2018

  53. [61]

    CoRR , volume =

    Anthony Hu and Lloyd Russell and Hudson Yeo and Zak Murez and George Fedoseev and Alex Kendall and Jamie Shotton and Gianluca Corrado , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2309.17080 , eprinttype =. 2309.17080 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.17080 2023

  54. [62]

    2023 , eprint=

    Consistency Models , author=. 2023 , eprint=

    2023

  55. [63]

    2026 , eprint=

    LOME: Learning Human-Object Manipulation with Action-Conditioned Egocentric World Model , author=. 2026 , eprint=

    2026

  56. [64]

    Follow Your Pose: Pose-Guided Text-to-Video Generation Using Pose-Free Videos , booktitle =

    Yue Ma and Yingqing He and Xiaodong Cun and Xintao Wang and Siran Chen and Xiu Li and Qifeng Chen , editor =. Follow Your Pose: Pose-Guided Text-to-Video Generation Using Pose-Free Videos , booktitle =. 2024 , url =. doi:10.1609/AAAI.V38I5.28206 , timestamp =

    doi:10.1609/aaai.v38i5.28206 2024

  57. [65]

    CoRR , volume =

    Lizhen Wang and Zhurong Xia and Tianshu Hu and Pengrui Wang and Pengfei Wang and Zerong Zheng and Ming Zhou , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2506.10568 , eprinttype =. 2506.10568 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.10568 2025

  58. [66]

    2019 , eprint=

    MediaPipe: A Framework for Building Perception Pipelines , author=. 2019 , eprint=

    2019

  59. [67]

    2021 , eprint=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. 2021 , eprint=

    2021

  60. [68]

    2021 , eprint=

    Masked Autoencoders Are Scalable Vision Learners , author=. 2021 , eprint=

    2021

  61. [69]

    2025 , eprint=

    Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer , author=. 2025 , eprint=

    2025

  62. [70]

    2025 , eprint=

    Qwen3-VL Technical Report , author=. 2025 , eprint=

    2025

  63. [71]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  64. [72]

    arXiv preprint arXiv:2503.20314 , year=

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

    Pith/arXiv arXiv

  65. [73]

    2025 , eprint=

    DINOv3 , author=. 2025 , eprint=

    2025

  66. [74]

    2025 , eprint=

    SpeakerVid-5M: A Large-Scale High-Quality Dataset for Audio-Visual Dyadic Interactive Human Generation , author=. 2025 , eprint=

    2025

  67. [75]

    2026 , eprint=

    Hand2World: Autoregressive Egocentric Interaction Generation via Free-Space Hand Gestures , author=. 2026 , eprint=

    2026

  68. [76]

    2023 , eprint=

    DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation , author=. 2023 , eprint=

    2023

  69. [77]

    Ni and Heung

    Hao Zhang and Feng Li and Shilong Liu and Lei Zhang and Hang Su and Jun Zhu and Lionel M. Ni and Heung. The Eleventh International Conference on Learning Representations,. 2023 , url =

    2023

  70. [78]

    Reconstructing Hands in 3

    Pavlakos, Georgios and Shan, Dandan and Radosavovic, Ilija and Kanazawa, Angjoo and Fouhey, David and Malik, Jitendra , booktitle=. Reconstructing Hands in 3

  71. [79]

    ArcFace: Additive Angular Margin Loss for Deep Face Recognition , volume=

    Deng, Jiankang and Guo, Jia and Yang, Jing and Xue, Niannan and Kotsia, Irene and Zafeiriou, Stefanos , year=. ArcFace: Additive Angular Margin Loss for Deep Face Recognition , volume=. IEEE Transactions on Pattern Analysis and Machine Intelligence , publisher=. doi:10.1109/tpami.2021.3087709 , number=

    arXiv 2021