Pith. sign in

REVIEW 5 major objections 5 minor 69 references

A single rendered visual proxy is enough to control both human motion and camera in video generation.

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 · deepseek-v4-flash

2026-08-04 17:51 UTC pith:A7HQP5XH

load-bearing objection Good idea, under-supported evaluation: the unified visual proxy is worth attending to, but the 'substantial gains' need external benchmarks and error bars. the 5 major comments →

arxiv 2608.01944 v1 pith:A7HQP5XH submitted 2026-08-03 cs.CV

UniMoCa: Unifying Motion and Camera Controls as Visual Proxies for Faithful Human Video Generation

Liming Tan , Ye Chen , Hao Zhang , Lirong Qian , Feifei Li , Bingbing Ni This is my paper
classification cs.CV
keywords human video generationmotion controlcamera controlvisual proxydiffusion transformermulti-person scenescamera trajectoryvideo conditioning
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.

UniMoCa is a representation-driven framework for making video generation obey both a chosen human motion and a chosen camera movement. Its central idea is to stop feeding the model two different kinds of control signals—pixel-aligned pose maps for the person and parametric camera embeddings for the camera—and instead render both into one identity-neutral 'motion-camera visual proxy' sequence. The authors argue that this unified visual interface removes the ambiguity a model faces when deciding whether an observed change comes from the subject or from the viewpoint. If correct, the same proxy lets a single diffusion transformer handle multi-person motion, occlusions, and moving cameras with less need for complex control modules. The paper backs this with a new curated dataset and experiments comparing against prior visual-plus-parametric conditioning.

Core claim

On the paper's own terms, the central claim is that motion control and camera control can be expressed as two distinguishable visual factors composed into one proxy frame, rather than as a visual signal plus a geometric embedding. By extracting 3D human motion and camera trajectories from a driving video, inverse-projecting the motion to a global 3D+T space, smoothing it, and then rendering the smoothed human geometry while replaying the recovered camera path, MCVP makes both controls visible in the same pixel space. The addition of explicit camera trajectory markers lets the model separate subject-induced from camera-induced changes. The authors show that this design yields better motion ad

What carries the argument

Motion-Camera Visual Proxy (MCVP): a rendered, identity-neutral image sequence that fuses a human visual factor (articulated body geometry rendered under the recovered camera trajectory) with a camera visual factor (explicit trajectory markers), so that both controls live in the same visual token space. The proxy is fed as additional tokens to a flow-matching image-to-video transformer, with a shifted rotary positional encoding separating proxy tokens from generated tokens, and random proxy dropout during training for robustness. This mechanism converts both motion and camera estimation errors into a single image-space residual, which the model can absorb more gracefully than a parametric ca

Load-bearing premise

The claimed gains are measured on a held-out subset of the authors' own 80k-clip dataset, created by discarding clips with reconstruction failures, implausible camera jumps, or proxy-video misalignment; if these discarded clips are exactly the hard cases that real driving videos contain, the reported improvements may not hold in general.

What would settle it

Train or evaluate the same framework on a set of driving videos that deliberately includes large camera jumps, occlusions, and incomplete body reconstructions, and compare MCVP against a parametric-camera baseline (e.g., Plücker embeddings). If the parametric baseline matches or beats MCVP on those clips, the claimed robustness would be a dataset artifact rather than a property of the representation.

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

If this is right

  • If MCVP works as claimed, future human video generators can be controlled through one visual interface instead of needing separate camera-embedding branches.
  • Motion and camera can be recombined freely: a stored 3D motion sequence can be paired with different camera paths or different first-frame identities without retraining.
  • The unified proxy may reduce sensitivity to camera estimation errors, making control pipelines more robust under noisy monocular reconstruction.
  • Because the proxy is identity-neutral, appearance contamination from intermediate geometry errors is reduced, improving character consistency across generations.
  • The approach adds minimal architecture complexity, so the same proxy could be dropped into other image-to-video diffusion models.

Where Pith is reading between the lines

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

  • An immediate testable extension is to evaluate MCVP on unfiltered driving videos with large camera jumps and partial occlusions; the current benchmark is built from a filtered subset, so the robustness claim may be optimistic.
  • The proxy representation is model-agnostic; probing whether the same MCVP conditioning transfers to a different backbone would separate the gain of the representation from the gain of the specific training pipeline.
  • Because motion is stored in global 3D+T space, MCVP could serve as an authoring format in game engines; this workflow is described but not measured, so a user study on authored proxies would be a natural next step.
  • A comparative study with purely parametric camera control on the same data, while holding the human visual factor fixed, would isolate whether the benefit comes from the visual markers or from the rendering under the camera trajectory.

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 / 5 minor

Summary. The paper proposes UniMoCa, a human video generation framework built on the Wan2.2 I2V backbone, together with a new conditioning representation called MCVP (Motion-Camera Visual Proxy). MCVP converts recovered 3D human motion and camera trajectories from a driving video into an identity-neutral visual proxy: a rendered human geometry sequence plus explicit camera trajectory markers. Both motion and camera controls are thereby injected as visual tokens via a trainable proxy encoder, with Shifted RoPE to distinguish proxy tokens from video tokens and MCVP dropout as regularization. The authors also introduce the MCVP-Video dataset, curated from 200k candidate clips to 80k video–MCVP training pairs. Experiments compare UniMoCa with MultiAnimate, SCAIL, SCAIL-2, Seedance 2.0, and ReCamMaster on self-constructed benchmarks, reporting better motion and camera control metrics, with ablations on the unified visual conditioning, Shifted RoPE, and dropout.

Significance. The central idea is attractive and potentially useful: replacing heterogeneous visual-plus-parametric camera controls with a unified visual proxy could simplify joint motion-camera conditioning and reduce sensitivity to estimation errors. The design is lightweight and builds on a strong open backbone. However, the empirical evidence as presented is not yet convincing enough to establish the paper's main claim of 'substantial gains.' The evaluation is performed on a held-out subset of the authors' own filtered dataset, baselines are not shown to have been adapted to that distribution, reported margins are small and lack error bars, and the most relevant joint-control baseline (Uni3C) is not compared. The paper also does not provide a direct robustness experiment for the 'camera-aware robustness' claim. If the method were evaluated more fairly and reproducibly, the MCVP representation could be a valuable contribution; at present the significance is conditional.

major comments (5)
  1. [Sec. 5.1 and Sec. 5.2 (Evaluation Setup)] Need no extra text.
  2. [Sec. 4.4 (MCVP-Video Data Curation)]
  3. [Sec. 5 (Comparisons) and Related Works]
  4. [Abstract and Sec. 5]
  5. [Sec. 3.2 (MCVP Construction)]
minor comments (5)
  1. [Tables 1 and 2] No error bars, confidence intervals, or significance tests are reported for any automatic or human evaluation metric. Human evaluation details (number of raters, instructions, agreement) are also missing. Given the small differences, these are essential for interpretation.
  2. [Figures 3 and 4 captions] Some qualitative comparisons explicitly use 'Seedance 2.0 (with MCVP)', while Table 1 lists only 'Seedance 2.0'. It is unclear whether the tabulated Seedance 2.0 results use MCVP as an input. If they do, the comparison is not a test of heterogeneous vs. unified control; if they do not, the qualitative and quantitative settings are inconsistent. Please clarify exactly what inputs were given to each baseline.
  3. [Eq. (11) and Sec. 5.3] The Shifted RoPE spatial offset is fixed to Δ=120, but no sensitivity ablation is reported. Since the offset is a free design choice central to the token-role separation, a small ablation (e.g., Δ=0, 60, 120, 240) would strengthen the paper.
  4. [Sec. 5.1] The claim of 'minimal additional complexity' is not quantified. Please report the number of added parameters, training FLOPs, and inference latency relative to the Wan2.2 baseline.
  5. [Eqs. (4) and (9)] The first-order perturbation expressions are illustrative but not formal bounds. The paper should label them as motivation rather than proof, since no dominance result is shown between heterogeneous and unified residual paths.

Circularity Check

0 steps flagged

No significant circularity: the central MCVP claim is an empirical, externally-benchmarked result, not an input-equivalent derivation.

full rationale

UniMoCa's central claim is that replacing heterogeneous visual motion maps plus parametric camera embeddings with the unified visual proxy MCVP improves joint human-motion and camera control. The paper's derivation chain is constructive rather than tautological: MCVP is defined as a rendered combination of recovered human geometry and camera-trajectory markers, and the subsequent experiments measure whether this representation actually improves control accuracy. No equation reduces a predicted quantity to a fitted input by construction: the camera-accuracy metrics (RotErr/TransErr) measure deviations from the prescribed trajectory in the generated video, not the marker positions used as input; the flow-matching loss in Eq. (10) is the standard objective, and the ablation in Table 3 tests the components empirically. The paper does rely on its own filtered MCVP-Video dataset for training and evaluation, and baselines are not retrained on that distribution, which is a legitimate benchmarking concern (a correctness/external-validity risk) but not circularity: the reported gains are not forced by the construction of MCVP or by any fitted parameter being renamed as a prediction. Self-citations appear only as general references to prior video-generation work and are not load-bearing for the novelty claim. There is no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result presented as a derivation. Accordingly, the paper is self-contained with respect to the circularity criteria defined here.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claim rests on the upstream reliability of SAM3DBody and Depth Anything 3, on the representational compatibility of MCVP with the Wan2.2 transformer, and on the validity of the paper's self-curated evaluation. Free parameters are limited to hand-picked design choices such as the RoPE offset, dropout rate, and dataset filtering criteria. No new physical entities are proposed; the invented entities are representation-level constructs.

free parameters (3)
  • Shifted RoPE spatial offset Δ = 120
    Hand-picked constant used to mark MCVP proxy tokens as control context rather than generated content. The optimal value is not derived or swept.
  • MCVP dropout rate = 0.15
    Chosen by hand as a stochastic conditioning probability. Affects the robustness-controllability tradeoff and is only evaluated as on/off in ablations.
  • Dataset filtering thresholds = Unspecified; 120k of 200k candidate clips discarded
    Qualitative criteria (incomplete reconstruction, discontinuous tracks, implausible camera jumps, misalignment) determine the final 80k training and evaluation clips. These thresholds affect the reported performance and are not precisely specified.
axioms (4)
  • domain assumption Wan2.2 I2V backbone can accept concatenated proxy tokens and learn from them without architectural changes beyond a trainable encoder.
    The entire method rests on the compatibility of the Wan2.2 denoising transformer with MCVP token concatenation and Shifted RoPE.
  • domain assumption SAM3DBody and Depth Anything 3 recover sufficiently accurate 3D human states and camera trajectories from driving videos.
    The MCVP construction, Eq. (5) and Eq. (6), depends on these upstream estimators being accurate enough for the proxy to be a faithful control signal.
  • standard math The first-order Jacobian perturbation analysis in Eqs. (4) and (9) is a valid model of error propagation in flow-matching training.
    The paper assumes small errors and smoothness of the denoising network to claim that unified visual residuals are easier to absorb. This is a heuristic approximation, not a proven theorem.
  • ad hoc to paper A held-out subset of the authors' own filtered MCVP-Video dataset is representative enough for evaluating generalization.
    The evaluation benchmark is constructed from the same curated dataset used for training, with no external standard benchmark; this is a load-bearing assumption for the validity of the empirical claims.
invented entities (2)
  • MCVP (Motion-Camera Visual Proxy) no independent evidence
    purpose: Unified visual conditioning representation that encodes human motion and camera trajectory into one visual signal.
    This is a new representation introduced by the paper. Its utility is demonstrated only in the paper's own experiments; there is no external falsifiable handle beyond the paper's benchmark.
  • Camera visual factor F_C (camera trajectory markers) no independent evidence
    purpose: Explicit visual markers that encode camera dynamics in the proxy frame.
    The design of the markers is not specified in detail, and their effectiveness is only evaluated indirectly through camera-accuracy metrics within the paper.

pith-pipeline@v1.3.0-daily-deepseek · 12631 in / 10959 out tokens · 117266 ms · 2026-08-04T17:51:08.177050+00:00 · methodology

0 comments
read the original abstract

Controlling human motion and camera movement is essential for faithful human-oriented video generation, yet remains challenging in multi-person scenes with large body motions, occlusions, and dynamic cameras. Existing pipelines typically rely on visual motion sequences, such as skeleton maps, pose maps, or rendered body representations, for motion control, while using camera embeddings for camera control. Such heterogeneous control interfaces force video generation models to reconcile pixel-aligned visual cues with non-visual geometric embeddings, making motion-camera attribution difficult and sensitive to camera estimation errors. We propose \textbf{UniMoCa}, a representation-driven framework that unifies motion and camera controls in visual space. At the core of UniMoCa is \textbf{Motion-Camera Visual Proxy} (\textbf{MCVP}), a mutually-sharable novel representation that converts 3D human motion and camera trajectories extracted from driving videos into an identity-neutral visual proxy. MCVP renders temporally aligned human geometry under the recovered camera trajectory and augments it with explicit camera trajectory markers, replacing heterogeneous visual-parametric controls with distinguishable visual cues. As both control factors are represented in the same visual space, they become mutually compatible rather than heterogeneous, enabling consistent joint reasoning and editing during video generation. We further curate a \textbf{MCVP-Video} dataset covering complex actions, multi-person interactions, and diverse camera trajectories. Experiments based on the Wan2.2 I2V show that UniMoCa achieves substantial gains in human motion control, camera control, temporal consistency, and camera-aware robustness with minimal additional complexity. More details are shown in our Project page: https://tanliming-daniel.github.io/UniMoCa/.

Figures

Figures reproduced from arXiv: 2608.01944 by Bingbing Ni, Feifei Li, Hao Zhang, Liming Tan, Lirong Qian, Ye Chen.

Figure 1
Figure 1. Figure 1: Illustration of our Motivation. Prior methods combine visual motion cues with geometric camera codes, creating ambiguous motion-camera attribution across het￾erogeneous control spaces (top). UniMoCa instead expresses both controls as MCVP in a shared visual space for faithful motion and camera control (bottom). Wang et al. 2025a). This requirement becomes particularly challenging in scenes involving multip… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of UniMoCa. Top: we recover camera-view human motion and camera trajectories from a driving video, transform and smooth the motion in global coordinates, and render the human and camera visual factors into MCVP. Bottom: a proxy encoder maps MCVP to visual tokens with Shifted RoPE; these tokens are concatenated with noisy video and appearance/text condition tokens and processed by the diffusion tra… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of motion control. Note that we also feed MCVP to Seedance 2.0, as it supports this input and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of camera control. UniMoCa follows the prescribed reframing more accurately while main [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Heterogeneous motion-camera controls degrade [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: We can see that UniMoCa better follows the pre [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

69 extracted references · 18 linked inside Pith

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    ICLR workshop on deep generative models for highly structured data , year=

    Video diffusion models , author=. ICLR workshop on deep generative models for highly structured data , year=

  13. [13]

    arXiv preprint arXiv:2311.15127 , year=

    Stable video diffusion: Scaling latent video diffusion models to large datasets , author=. arXiv preprint arXiv:2311.15127 , year=

    Pith/arXiv arXiv

  14. [14]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Align your latents: High-resolution video synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  15. [15]

    arXiv preprint arXiv:2501.11325 , year=

    Catv2ton: Taming diffusion transformers for vision-based virtual try-on with temporal concatenation , author=. arXiv preprint arXiv:2501.11325 , year=

    Pith/arXiv arXiv

  16. [16]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , month =

    Zou, Yuliang and Yang, Jimei and Ceylan, Duygu and Zhang, Jianming and Perazzi, Federico and Huang, Jia-Bin , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , month =

  17. [17]

    arXiv preprint arXiv:2504.14977 , year=

    Realisdance-dit: Simple yet strong baseline towards controllable character animation in the wild , author=. arXiv preprint arXiv:2504.14977 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2409.06202 , year=

    Realisdance: Equip controllable character animation with realistic hands , author=. arXiv preprint arXiv:2409.06202 , year=

    Pith/arXiv arXiv

  19. [19]

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

    Reconstructing hands in 3d with transformers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  20. [20]

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

    MultiAnimate: Pose-Guided Image Animation Made Extensible , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  21. [21]

    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=

  22. [22]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Expressive body capture: 3d hands, face, and body from a single image , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  23. [23]

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

    Scail: Towards studio-grade character animation via in-context learning of 3d-consistent pose representations , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  24. [24]

    arXiv preprint arXiv:2606.10804 , year=

    SCAIL-2: Unifying Controlled Character Animation with End-to-end In-Context Conditioning , author=. arXiv preprint arXiv:2606.10804 , year=

    Pith/arXiv arXiv

  25. [25]

    Advances in Neural Information Processing Systems , volume=

    Cameras as relative positional encoding , author=. Advances in Neural Information Processing Systems , volume=

  26. [26]

    arXiv preprint arXiv:2404.02101 , year=

    Cameractrl: Enabling camera control for text-to-video generation , author=. arXiv preprint arXiv:2404.02101 , year=

    Pith/arXiv arXiv

  27. [27]

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

    Recammaster: Camera-controlled generative rendering from a single video , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  28. [28]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    Uni3c: Unifying precisely 3d-enhanced camera and human motion controls for video generation , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  29. [29]

    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

  30. [30]

    arXiv preprint arXiv:2104.09864 , year=

    Roformer: Enhanced transformer with rotary position embedding , author=. arXiv preprint arXiv:2104.09864 , year=

    Pith/arXiv arXiv

  31. [31]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Koala-36m: A large-scale video dataset improving consistency between fine-grained conditions and video content , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  32. [32]

    arXiv preprint arXiv:2501.05098 , year=

    Motion-x++: A large-scale multimodal 3d whole-body human motion dataset , author=. arXiv preprint arXiv:2501.05098 , year=

    Pith/arXiv arXiv

  33. [33]

    The eleventh international conference on learning representations , year=

    Flow matching for generative modeling , author=. The eleventh international conference on learning representations , year=

  34. [34]

    Computational Visual Media , year=

    Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models , author=. Computational Visual Media , year=

  35. [35]

    Advances in Neural Information Processing Systems , volume=

    Humanvid: Demystifying training data for camera-controllable human image animation , author=. Advances in Neural Information Processing Systems , volume=

  36. [36]

    arXiv preprint arXiv:2511.10647 , year=

    Depth anything 3: Recovering the visual space from any views , author=. arXiv preprint arXiv:2511.10647 , year=

    Pith/arXiv arXiv

  37. [37]

    The journal of machine learning research , volume=

    Dropout: a simple way to prevent neural networks from overfitting , author=. The journal of machine learning research , volume=. 2014 , publisher=

    2014

  38. [38]

    arXiv preprint arXiv:2604.14148 , year=

    Seedance 2.0: Advancing video generation for world complexity , author=. arXiv preprint arXiv:2604.14148 , year=

    Pith/arXiv arXiv

  39. [39]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    Camclonemaster: Enabling reference-based camera control for video generation , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  40. [40]

    arXiv preprint arXiv:2601.05138 , year=

    Versecrafter: Dynamic realistic video world model with 4d geometric control , author=. arXiv preprint arXiv:2601.05138 , year=

    arXiv

  41. [41]

    IEEE Transactions on Visualization and Computer Graphics , year=

    Vidcraft3: Camera, object, and lighting control for image-to-video generation , author=. IEEE Transactions on Visualization and Computer Graphics , year=

  42. [42]

    arXiv preprint arXiv:2602.15989 , year=

    Sam 3d body: Robust full-body human mesh recovery , author=. arXiv preprint arXiv:2602.15989 , year=

    arXiv

  43. [43]

    China University of Mining Technology Beijing Graduate School , volume=

    Frechet inception distance (fid) for evaluating gans , author=. China University of Mining Technology Beijing Graduate School , volume=

  44. [44]

    FVD: A new metric for video generation , author=

  45. [45]

    On the content bias in fr

    Ge, Songwei and Mahapatra, Aniruddha and Parmar, Gaurav and Zhu, Jun-Yan and Huang, Jia-Bin , booktitle=. On the content bias in fr

  46. [46]

    arXiv preprint arXiv:2406.19680 , year=

    Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance , author=. arXiv preprint arXiv:2406.19680 , year=

    Pith/arXiv arXiv

  47. [47]

    Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

    Cinemaster: A 3d-aware and controllable framework for cinematic text-to-video generation , author=. Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

  48. [48]

    Science China Information Sciences , volume=

    Unianimate: Taming unified video diffusion models for consistent human image animation , author=. Science China Information Sciences , volume=. 2025 , publisher=

    2025

  49. [49]

    arXiv preprint arXiv:2509.14055 , year=

    Wan-animate: Unified character animation and replacement with holistic replication , author=. arXiv preprint arXiv:2509.14055 , year=

    arXiv

  50. [50]

    International Conference on Learning Representations , volume=

    Cogvideox: Text-to-video diffusion models with an expert transformer , author=. International Conference on Learning Representations , volume=

  51. [51]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Animate anyone: Consistent and controllable image-to-video synthesis for character animation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  52. [52]

    Controlnet++: Improving conditional controls with efficient consistency feedback: Project page: liming-ai. github. io/controlnet\_plus\_plus , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  53. [53]

    arXiv preprint arXiv:2410.10306 , year=

    Animate-x: Universal character image animation with enhanced motion representation , author=. arXiv preprint arXiv:2410.10306 , year=

    Pith/arXiv arXiv

  54. [54]

    Seminal Graphics Papers: Pushing the Boundaries, Volume 2 , pages=

    SMPL: A skinned multi-person linear model , author=. Seminal Graphics Papers: Pushing the Boundaries, Volume 2 , pages=

  55. [55]

    European conference on computer vision , pages=

    Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image , author=. European conference on computer vision , pages=. 2016 , organization=

    2016

  56. [56]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  57. [57]

    International Conference on Learning Representations , volume=

    Trajectory attention for fine-grained video motion control , author=. International Conference on Learning Representations , volume=

  58. [58]

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

    Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  59. [59]

    International Conference on Learning Representations , volume=

    Vd3d: Taming large video diffusion transformers for 3d camera control , author=. International Conference on Learning Representations , volume=

  60. [60]

    arXiv preprint arXiv:2412.00131 , year=

    Open-sora plan: Open-source large video generation model , author=. arXiv preprint arXiv:2412.00131 , year=

    Pith/arXiv arXiv

  61. [61]

    European Conference on Computer Vision , pages=

    Rotary position embedding for vision transformer , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    Let them talk: Audio-driven multi-person conversational video generation , author=. Advances in Neural Information Processing Systems , volume=

  63. [63]

    Proceedings of the computer vision and pattern recognition conference , pages=

    Prompthmr: Promptable human mesh recovery , author=. Proceedings of the computer vision and pattern recognition conference , pages=

  64. [64]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  65. [65]

    arXiv preprint arXiv:2606.31946 , year=

    World Narrative Model for Highly Controllable Video Generation: A Paradigm Shift from Pixel Sampling to Physical World Orchestration , author=. arXiv preprint arXiv:2606.31946 , year=

    Pith/arXiv arXiv

  66. [66]

    arXiv preprint arXiv:2510.12256 , year=

    Vectorized video representation with easy editing via hierarchical spatio-temporally consistent proxy embedding , author=. arXiv preprint arXiv:2510.12256 , year=

    arXiv

  67. [67]

    arXiv preprint arXiv:2505.04512 , year=

    Hunyuancustom: A multimodal-driven architecture for customized video generation , author=. arXiv preprint arXiv:2505.04512 , year=

    Pith/arXiv arXiv

  68. [68]

    arXiv preprint arXiv:2606.02753 , year=

    MetaWorld: Scaling Multi-Agent Video World Model from Single-view Video Data , author=. arXiv preprint arXiv:2606.02753 , year=

    Pith/arXiv arXiv

  69. [69]

    arXiv preprint arXiv:2404.15789 , year=

    Motionmaster: Training-free camera motion transfer for video generation , author=. arXiv preprint arXiv:2404.15789 , year=

    Pith/arXiv arXiv