Pith. sign in

REVIEW 3 major objections 4 minor 65 references

One-shot Human Motion Transfer via Occlusion-Robust Flow Prediction and Neural Texturing

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper claims that fusing 2D appearance flow warping with 2.5D neural texture mapping in a single jointly trained model lets one-shot human motion transfer keep the correct pose through heavy self-occlusion, even when the DensePose…

desk verdict Solid if incremental systems paper: pose-transfer numbers are genuinely strong, but the occlusion-robustness headline outruns the evidence, and the whole method sits on DensePose pseudo-labels the authors admit can break it. read the letter →

arxiv 2412.06174 v1 pith:HJIG7LIH submitted 2024-12-09 cs.CV

classification cs.CV
keywords humanmotiontransferone-shotanimationDensePoseIUVneuraltexturemappingmulti-modalfeaturefusionself-occlusionhandlingvideosynthesis
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 tries to solve one-shot human motion transfer - animating a single static photo of a person with a driving video - by treating appearance and geometry as parallel streams rather than competing alternatives. It argues that 2D landmark and skeleton methods lose appearance under large pose changes and self-occlusion, while DensePose-based neural rendering suffers from the IUV maps' noise and gaps, and that a unified framework with multi-scale feature warping plus neural texture mapping can get both right. The paper reports that the model achieves the best pose-accuracy scores (AKD, MKR) against ten prior methods on the TED-Talks, TaiChiHD and iPER benchmarks, and that human evaluators rate its geometry the highest. A reader should care because keeping the correct pose under occlusion is the main bottleneck to believable, controllable character animation from a single image.

What carries the argument

The load-bearing object is the DensePose IUV map, a per-pixel 2.5D parametrization of the SMPL body surface. The paper treats it as a noisy but usable anchor: a MotionNet outputs residual dense flow plus occlusion maps for the 2D branch and two latent translation signals ($\alpha$, $\rho$) for the 2.5D branch. The texture branch separates the neural texture atlas translation from DensePose IUV translation with two small encoders, then uses differentiable bilinear sampling to gather texels at predicted UV coordinates and weights the per-body-part results by a soft semantic score map, producing a geometry feature that is fused with the warped appearance features in a BlenderNet. This decoupling is what lets the network repair some broken IUV maps instead of inheriting their holes.

What would settle it

Take a fixed set of driving videos whose DensePose IUV maps have large missing-body-part holes, and measure AKD, MKR and per-pixel error on the missing regions; if output pose errors grow with the missing area and the generated frames contain the same holes, the claim of occlusion-robust flow and texture repair would be disproved.

Watch

Extended reading notes

Core claim

The central discovery is that the two standard families of human motion transfer - 2D appearance warping and 2.5D geometry rendering - are complementary, and a network that fuses them at training time can compensate for each one's weakness: the flow-warping branch preserves identity and fine appearance, while the neural texture branch stabilizes geometry even when DensePose is wrong. The key architectural move is to decouple translation of the neural texture atlas from translation of the DensePose IUV map, so neither needs ground truth; both are learned end-to-end through a differentiable texture-mapping sampler that weights each body part by its predicted soft mask. On three benchmark datasets, the fused model reaches the lowest average keypoint distance and missing-keypoint rate among ten state-of-the-art methods, and it visibly handles turning-around and front-to-back motions.

Load-bearing premise

The method assumes frame-by-frame DensePose IUV maps, though noisy, are dependable enough to supervise both the UV regression and the dense flow; the paper admits that when IUV correspondences are severely missing the model cannot correct them.

Editorial extensions

If this is right

  • On the three benchmarks, the fused model obtains lower AKD and MKR than ten prior methods, so pose alignment in one-shot animations improves by a large margin on cross-video tests.
  • The model keeps competitive identity preservation (AED) and image quality (L1/FID) while running at about 58 FPS, orders of magnitude faster than diffusion-based animators.
  • The ablation shows that removing the occlusion maps or reducing warping to a single scale raises AKD/MKR, so multi-scale flow with occlusion filtering is necessary for the appearance branch.
  • The 2.5D branch can complete partially broken driving IUV maps, but when correspondence is severely missing (arms, hands or feet absent) the model fails, limiting its robustness ceiling.
  • With about 34.8M parameters, the approach fits on lower-end GPUs, whereas the diffusion baselines the paper compares against cannot run at interactive speed.

Reading between the lines

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

  • The decoupled texture/geometry translation is representation-agnostic, so the same two-stream fusion could animate other articulated objects whenever a UV-like correspondence map is available.
  • Because the pose-accuracy advantage is largest on the full-body, high-occlusion datasets (TaiChiHD and iPER), a natural extension is to stress-test the architecture on in-the-wild footage with deliberately poor DensePose coverage.
  • The reported failure on subtle facial motion suggests a cheap extension: keep the two-stream fusion but replace DensePose-only conditioning with a parametric face/hand model, which may recover lip and eyelid movement the current model misses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a one-shot human motion transfer framework that combines multi-scale dense flow warping (2D appearance branch) with a neural texture mapping branch conditioned on DensePose IUV maps (2.5D geometry branch). MotionNet predicts dense flow, occlusion maps, and latent translation signals; the 2.5D branch translates a neural texture atlas and DensePose geometry in a differentiable texture-mapping framework; a BlenderNet fuses the two feature streams. Experiments on TED-Talks, TaiChiHD, and iPER compare against ten prior methods for self-reconstruction, cross-video animation, human evaluation, and runtime, plus ablations on TED-Talks. The clearest quantitative strength is pose transfer: AKD and MKR are consistently among the best, while appearance metrics (L1, FID, AED) are mixed. The paper claims particular effectiveness under substantial self-occlusion and front-to-back view changes, but this claim is supported mainly by selected qualitative examples.

Significance. If the central claims held, the paper would make a useful contribution: a single architecture that exploits DensePose geometry without being destroyed by DensePose errors, with efficient inference and strong pose-transfer metrics. The proposed decoupling of texture-atlas translation and DensePose translation within a differentiable neural texture mapper is a reasonable and interesting mechanism. Strengths of the empirical work include comparisons on three datasets with ten baselines, both self-reconstruction and cross-video settings, a clear ablation structure, and an unusually explicit limitations section. However, two load-bearing gaps remain: the ablation does not show that the 2.5D branch improves the full model on most metrics, and the occlusion-robustness claim is not tested in the corrupted-DensePose regime that the paper itself identifies as its failure mode.

major comments (3)
  1. [§V-A; Eq. (11)] The abstract and Section I claim that the method is 'particularly effective in handling challenging cases such as those with substantial self-occlusions', but no experiment isolates that regime. Since Eq. (11) trains the UV regression against DensePose pseudo-labels and the MotionNet consumes the same frame-by-frame IUV maps, the model's geometric anchor is the pseudo-label quality. Section V-A concedes that 'the model will fail to correct geometry errors when the correspondence in the IUV map is severely missing', which is exactly the self-occlusion regime. The paper should add a quantitative evaluation (e.g., corrupting IUV maps by erasing parts or masking low-confidence regions, then reporting AKD/MKR/L1) and/or stratify the test set by DensePose quality or occlusion severity. Without this, the central occlusion-robustness claim is supported only by selected qualitative examples.
  2. [Table III] The ablation does not support the claim in Section IV-B that the full model 'can take advantage of both 2D and 2.5D branches'. Comparing '2D baseline+MotionNet+BlenderNet' with the full model, cross-video AED worsens from 0.3481 to 0.3515, MKR from 0.0176 to 0.0186, and AKD from 3.3809 to 3.6755; in self-reconstruction, FID worsens from 22.5662 to 26.3862 and AKD from 1.9833 to 2.1255. Only MKR and AED improve slightly in self-reconstruction. The authors should either report repeated runs with confidence intervals to show these differences are not noise, or revise the claim that the 2.5D branch improves the unified framework; currently the load-bearing contribution of the neural texture mapping branch is not quantitatively established.
  3. [§IV-A; Tables I and II] The paper should be more precise about what 'competitive results' means. Table I shows that on L1, FID, and AED the method is often worse than MRAA, DAM, or TPSMotion (e.g., TED-Talks L1 0.0317 vs. 0.0260 for DAM and FID 20.15 vs. 17.70 for MRAA), while Table II shows higher AED in cross-video animation. The gains are concentrated in AKD/MKR, which are pose metrics. This is a legitimate strength, but the paper should state it explicitly and provide statistical significance or confidence intervals so readers can judge whether the appearance deficit is within noise; the current explanation in terms of 128×256 training and unstable DensePose is plausible but not verified.
minor comments (4)
  1. [§IV-A, Table II] The human evaluation was conducted with 12 participants on 30 pairs; report inter-rater agreement or confidence intervals before claiming consistent outperformance in Geometry and Overall.
  2. [§IV-C] The FPS comparison excludes DensePose extraction, and the text notes DensePose alone runs at 20 FPS; the end-to-end throughput should be reported so readers can compare fairly with methods that do not require a separate DensePose forward pass.
  3. [§III-B, Eq. (3)] The occlusion map semantics are unclear: the text says that when the occlusion is 0 source-aligned features are passed, but the equation gives Warp(...) * o + previous * (1-o), which passes the previous feature when o=0; clarify whether o is an occlusion map or a visibility map.
  4. [Table III] The abbreviation 'w SSFW' appears in the table without definition; expand it in the caption or text.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claims are validated on external benchmarks; the DensePose pseudo-label loop is a stated limitation, not a forced prediction.

full rationale

This is an empirical systems paper rather than a formal derivation, and the core claims are anchored to external evidence. The headline metrics (AKD, MKR, AED, FID) are computed with external pose estimators and feature extractors on held-out generated frames, and the compared SOTA methods are retrained or run under fixed protocols, so the reported competitive results are not fitted to the evaluation quantities by construction. No parameter is fit to a subset of the evaluation data and then renamed as a prediction, and no load-bearing step is justified by a self-citation chain: the loss components (perceptual correctness, TV, LSGAN, neural texture priors from [17,40,52]) are standard external components, and the paper invokes no uniqueness theorem or prior-work ansatz as authority. The one self-referential element is that DensePose IUV maps are used both as network input (Eqs. 1 and 5) and as pseudo-ground-truth in the UV regression loss (Eq. 11): the geometry branch is supervised with the same DensePose estimator it is supposed to correct. The paper itself acknowledges this in Section V-A, stating that the model 'will fail to correct geometry errors when the correspondence in the IUV map is severely missing.' This is a genuine limitation and a correctness risk for the occlusion-robustness claim, but it is not a circular derivation of the paper's empirical claims, because the final outputs are additionally constrained by masked L1, perceptual, mask, and adversarial losses against real frames, and because the quantitative comparison is anchored to external evaluators rather than to the DensePose pseudo-labels. No specific equation can be exhibited in which a stated prediction equals its input by construction, so the appropriate finding is a low non-circularity score.

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

No physical entities are invented; the model is a learned composition of known modules. The key external dependencies are pretrained DensePose, video matting, VGG-19, re-ID, and pose estimators. The listed free parameters are hand-set training hyperparameters and architecture choices that affect the reported results but are not fitted constants in a derivation.

free parameters (3)
  • Loss weights (lambda_cor, lambda_reg, lambda_tv, lambda_con, lambda_uv, lambda_ce, lambda_p, lambda_1) = 5.0, 0.01, 1.0, 5.0, 5.0, 1.0, 10.0, 1.0
    Chosen by hand in Section III-E; not tuned per dataset, but they shape the joint training and no sensitivity analysis is reported.
  • Latent translation signal lengths (alpha, rho) = 384, 256
    Set in Section IV Experimental Setup; capacity of the texture and DensePose translation branches depends on these lengths, and the paper states single-branch performance is limited by them.
  • Multi-stage training schedule and iteration counts = 100k + 100k + 200k + 200k (total 600k)
    Multi-stage schedule in Section IV; the 256x256 variant removes several losses and is trained end-to-end from scratch, so reported gains conflate architecture and schedule choices.
assumptions (4)
  • domain assumption DensePose IUV maps provide sufficiently accurate 2.5D correspondence for supervision and conditioning.
    The 2.5D branch, Eq. (5)-(7) and Eq. (11), trains UV regression against DensePose pseudo-labels; Section V-A documents failures when these labels have holes or missing parts.
  • domain assumption Learning residual translation signals (alpha, rho) and residual flow from source to driving pose is easier and more stable than direct prediction.
    Section III-A states this as the key assumption behind the MotionNet design; no independent evidence is provided.
  • domain assumption Pretrained DensePose, video matting, VGG-19, re-ID, and pose estimators are accurate enough for pseudo-labels and metrics.
    Used in Section IV for data preparation (DensePose, matting) and evaluation (VGG, re-ID, pose keypoints); errors in these propagate to both training and measured scores.
  • ad hoc to paper Neural texture mapping with predicted DensePose translation can correct geometric errors in corrupted IUV maps.
    Central to the 2.5D branch; Section V-A concedes the model will fail to correct geometry errors when correspondence is severely missing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of One-shot Human Motion Transfer via Occlusion-Robust Flow Prediction and Neural Texturing." pith.science (2026). https://pith.science/paper/HJIG7LIH

@misc{pith2026241206174,
  author       = {Pith},
  title        = {Pith review of: One-shot Human Motion Transfer via Occlusion-Robust Flow Prediction and Neural Texturing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJIG7LIH}},
  note         = {Machine review of arXiv:2412.06174}
}
read the original abstract

Human motion transfer aims at animating a static source image with a driving video. While recent advances in one-shot human motion transfer have led to significant improvement in results, it remains challenging for methods with 2D body landmarks, skeleton and semantic mask to accurately capture correspondences between source and driving poses due to the large variation in motion and articulation complexity. In addition, the accuracy and precision of DensePose degrade the image quality for neural-rendering-based methods. To address the limitations and by both considering the importance of appearance and geometry for motion transfer, in this work, we proposed a unified framework that combines multi-scale feature warping and neural texture mapping to recover better 2D appearance and 2.5D geometry, partly by exploiting the information from DensePose, yet adapting to its inherent limited accuracy. Our model takes advantage of multiple modalities by jointly training and fusing them, which allows it to robust neural texture features that cope with geometric errors as well as multi-scale dense motion flow that better preserves appearance. Experimental results with full and half-view body video datasets demonstrate that our model can generalize well and achieve competitive results, and that it is particularly effective in handling challenging cases such as those with substantial self-occlusions.

Figures

Figures reproduced from arXiv: 2412.06174 by the authors.

Figure 1
Figure 1. Broad view of our approach for human motion transfer allows for good fidelity in 2D appearance transfer by estimating 2D motion flow, while establishing pose accuracy through 2.5D geometric reasoning. See our comparative results in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed pipeline. The MotionNet (a) produces the dense motion flow and translation signals for the appearance and geometry translation branches. We combine multi-scale feature warping (b) and neural texture mapping (c) into a unified framework for motion transfer. The translated appearance and geometry features are integrated by BlenderNet(d) for image refinement. Different from the above diffusion-… view at source ↗
Figure 3
Figure 3. Qualitative results in comparisons with state-of-the-art. We show the results from TED-Talks (the first four sets), TaichiHD (the forth and fifth sets), and iPER dataset (the last two sets). It illustrates our model can animate both half and full human body images with (1) better geometry and details (see the first and second examples), (2) large variations in pose for front-to-back view and self-occlusion (see the … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons for ablation study. We show the final generated images, dense motion flow and translated DensePose IUV map for animation. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Failure cases produced by our model with (1) broken DensePose [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 59 canonical work pages

  1. [1]

    Neural rendering and reenactment of human actor videos,

    L. Liu, W. Xu, M. Zollhoefer et al., “Neural rendering and reenactment of human actor videos,” ACM TOG, vol. 38, no. 5, pp. 1–14, 2019

  2. [2]

    State of the art on neural rendering,

    A. Tewari, O. Fried, J. Thies et al., “State of the art on neural rendering,” Comput. Graph. Forum, vol. 39, no. 2, pp. 701–727, 2020

  3. [3]

    Deformable gans for pose-based human image generation,

    A. Siarohin, E. Sangineto, S. Lathuili `ere et al. , “Deformable gans for pose-based human image generation,” in CVPR, 2018, pp. 3408–3416

  4. [4]

    First order motion model for image animation,

    A. Siarohin, S. Lathuili `ere, S. Tulyakov et al., “First order motion model for image animation,” in NIPS, 2019, pp. 7135–7145. THIS ARTICLE HAS BEEN ACCEPTED FOR PUBLICATION IN IEEE TRANSACTIONS ON MULTIMEDIA. DOI:XX.XXX/XXX.XX.XXXX 13

  5. [5]

    Motion representations for articulated animation,

    A. Siarohin, O. J. Woodford, J. Ren et al., “Motion representations for articulated animation,” in CVPR, 2021, pp. 13 653–13 662

  6. [6]

    Learning motion refinement for unsupervised face animation,

    J. Tao, S. Gu, W. Li et al., “Learning motion refinement for unsupervised face animation,” in NeurIPS, 2023

  7. [7]

    Progressive pose attention transfer for person image generation,

    Z. Zhu, T. Huang, B. Shi et al., “Progressive pose attention transfer for person image generation,” in CVPR, 2019, pp. 2347–2356

  8. [8]

    Neural texture extraction and distribution for controllable person image synthesis,

    Y . Ren, X. Fan, G. Li et al., “Neural texture extraction and distribution for controllable person image synthesis,” in CVPR, 2022, pp. 13 525– 13 534

Show all 65 references
  1. [9]

    Person image synthesis via denoising diffusion model,

    A. K. Bhunia, S. H. Khan, H. Cholakkal et al., “Person image synthesis via denoising diffusion model,” in CVPR, 2023, pp. 5968–5976

  2. [10]

    Cross-domain correspondence learning for exemplar-based image translation,

    P. Zhang, B. Zhang, D. Chen et al. , “Cross-domain correspondence learning for exemplar-based image translation,” in CVPR, 2020, pp. 5143–5153

  3. [11]

    Cocosnet v2: Full-resolution correspondence learning for image translation,

    X. Zhou, B. Zhang, T. Zhang et al. , “Cocosnet v2: Full-resolution correspondence learning for image translation,” in CVPR, 2021, pp. 11 465–11 475

  4. [12]

    Video-to-video synthesis,

    T. Wang, M. Liu, J. Zhu et al., “Video-to-video synthesis,” in NeurIPS, 2018, pp. 1152–1164

  5. [13]

    Everybody dance now,

    C. Chan, S. Ginosar, T. Zhou et al., “Everybody dance now,” in ICCV, 2019, pp. 5932–5941

  6. [14]

    Few-shot video-to-video synthesis,

    T. Wang, M. Liu, A. Tao et al., “Few-shot video-to-video synthesis,” in NeurIPS, 2019, pp. 5014–5025

  7. [15]

    Robust pose transfer with dynamic details using neural video rendering,

    Y . Sun, H. Huang, X. Wang et al., “Robust pose transfer with dynamic details using neural video rendering,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 2, pp. 2660–2666, 2023

  8. [16]

    Dense pose transfer,

    N. Neverova, R. A. G ¨uler, and I. Kokkinos, “Dense pose transfer,” in ECCV, vol. 11207, 2018, pp. 128–143

  9. [17]

    Neural re-rendering of humans from a single image,

    K. Sarkar, D. Mehta, W. Xu et al., “Neural re-rendering of humans from a single image,” in ECCV, vol. 12356, 2020, pp. 596–613

  10. [19]

    Poxture: Human posture imitation using neural texture,

    C. Yang, S. Yao, Z. Zhou et al. , “Poxture: Human posture imitation using neural texture,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 12, pp. 8537–8549, 2022

  11. [20]

    A 3d mesh-based lifting-and-projection network for human pose transfer,

    J. Liu, Y . Zhao, S. Chen et al., “A 3d mesh-based lifting-and-projection network for human pose transfer,” IEEE Trans. Multim. , vol. 24, pp. 4314–4327, 2022

  12. [21]

    Dreampose: Fashion image-to- video synthesis via stable diffusion,

    J. Karras, A. Holynski, T. Wang et al., “Dreampose: Fashion image-to- video synthesis via stable diffusion,” CoRR, vol. abs/2304.06025, 2023

  13. [22]

    Densepose: Dense human pose estimation in the wild,

    R. A. G ¨uler, N. Neverova, and I. Kokkinos, “Densepose: Dense human pose estimation in the wild,” in CVPR, 2018, pp. 7297–7306

  14. [23]

    SMPL: a skinned multi- person linear model,

    M. Loper, N. Mahmood, J. Romero et al. , “SMPL: a skinned multi- person linear model,” ACM Trans. Graph. , vol. 34, no. 6, pp. 248:1– 248:16, 2015

  15. [24]

    Pose-driven realistic 2-d motion synthesis,

    G. Xia, F. Ma, Q. Liu et al., “Pose-driven realistic 2-d motion synthesis,” IEEE Trans. Cybern. , vol. 53, no. 4, pp. 2412–2425, 2023

  16. [25]

    High-fidelity neural human motion transfer from monocular video,

    M. Kappel, V . Golyanik, M. Elgharib et al., “High-fidelity neural human motion transfer from monocular video,” in CVPR, 2021, pp. 1541–1550

  17. [26]

    Video reenactment as inductive bias for content-motion disentanglement,

    J. F. H. Albarrac ´ın and A. R. Rivera, “Video reenactment as inductive bias for content-motion disentanglement,” IEEE Trans. Image Process. , vol. 31, pp. 2365–2374, 2022

  18. [27]

    GAC-GAN: A general method for appearance-controllable human video motion transfer,

    D. Wei, X. Xu, H. Shen et al. , “GAC-GAN: A general method for appearance-controllable human video motion transfer,” IEEE Trans. Multim., vol. 23, pp. 2457–2470, 2021

  19. [28]

    PISE: person image synthesis and editing with decoupled GAN,

    J. Zhang, K. Li, Y . Lai et al., “PISE: person image synthesis and editing with decoupled GAN,” in CVPR, 2021, pp. 7982–7990

  20. [29]

    Soft-gated warping-gan for pose- guided person image synthesis,

    H. Dong, X. Liang, K. Gong et al., “Soft-gated warping-gan for pose- guided person image synthesis,” in NIPS, 2018, pp. 472–482

  21. [30]

    Deep spatial transformation for pose-guided person image generation and animation,

    Y . Ren, G. Li, S. Liu et al., “Deep spatial transformation for pose-guided person image generation and animation,” IEEE Trans. Image Process. , vol. 29, pp. 8622–8635, 2020

  22. [31]

    Motion transformer for unsupervised image animation,

    J. Tao, B. Wang, T. Ge et al. , “Motion transformer for unsupervised image animation,” in ECCV, vol. 13676, 2022, pp. 702–719

  23. [32]

    Controllable person image synthesis with attribute-decomposed GAN,

    Y . Men, Y . Mao, Y . Jiang et al. , “Controllable person image synthesis with attribute-decomposed GAN,” in CVPR, 2020, pp. 5083–5092

  24. [33]

    C2F-FWN: coarse-to-fine flow warping network for spatial-temporal consistent motion transfer,

    D. Wei, X. Xu, H. Shen et al., “C2F-FWN: coarse-to-fine flow warping network for spatial-temporal consistent motion transfer,” in AAAI, 2021, pp. 2852–2860

  25. [34]

    REMOT: A region-to-whole framework for realistic human motion transfer,

    Q. Yang, X. Liu, W. Liu et al., “REMOT: A region-to-whole framework for realistic human motion transfer,” in ACM MM, 2022, pp. 1128–1137

  26. [35]

    Pose-guided human animation from a single image in the wild,

    J. S. Yoon, L. Liu, V . Golyanik et al. , “Pose-guided human animation from a single image in the wild,” in CVPR, 2021, pp. 15 039–15 048

  27. [36]

    Dwnet: Dense warp-based network for pose-guided human video generation,

    P. Zablotskaia, A. Siarohin, B. Zhao et al., “Dwnet: Dense warp-based network for pose-guided human video generation,” in BMVC, 2019, p. 51

  28. [37]

    Humangan: A generative model of human images,

    K. Sarkar, L. Liu, V . Golyanik et al., “Humangan: A generative model of human images,” in 3DV, 2021, pp. 258–267

  29. [38]

    Dense intrinsic appearance flow for human pose transfer,

    Y . Li, C. Huang, and C. C. Loy, “Dense intrinsic appearance flow for human pose transfer,” in CVPR, 2019, pp. 3693–3702

  30. [39]

    Liquid warping GAN: A unified framework for human motion imitation, appearance transfer and novel view synthesis,

    W. Liu, Z. Piao, J. Min et al. , “Liquid warping GAN: A unified framework for human motion imitation, appearance transfer and novel view synthesis,” in ICCV, 2019, pp. 5903–5912

  31. [40]

    Few-shot human motion transfer by personalized geometry and texture modeling,

    Z. Huang, X. Han, J. Xu et al. , “Few-shot human motion transfer by personalized geometry and texture modeling,” in CVPR, 2021, pp. 2297– 2306

  32. [41]

    Deferred neural rendering: Image synthesis using neural textures,

    J. Thies, M. Zollh ¨ofer, and M. Nießner, “Deferred neural rendering: Image synthesis using neural textures,” ACM TOG, vol. 38, no. 4, pp. 1–12, 2019

  33. [42]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in NeurIPS, H. Larochelle, M. Ranzato, R. Hadsell et al., Eds., vol. 33, 2020, pp. 6840–6851

  34. [43]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy et al., “Learning transferable visual models from natural language supervision,” in ICML, vol. 139, 2021, pp. 8748–8763

  35. [44]

    Animate anyone: Consistent and controllable image-to-video synthesis for character animation,

    L. Hu, X. Gao, P. Zhang et al. , “Animate anyone: Consistent and controllable image-to-video synthesis for character animation,” CoRR, vol. abs/2311.17117, 2023

  36. [45]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in ICCV, 2023, pp. 3836–3847

  37. [46]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning,

    Y . Guo, C. Yang, A. Rao et al., “Animatediff: Animate your personalized text-to-image diffusion models without specific tuning,” CoRR, vol. abs/2307.04725, 2023

  38. [47]

    Magicanimate: Temporally consistent human image animation using diffusion model,

    Z. Xu, J. Zhang, J. H. Liew et al., “Magicanimate: Temporally consistent human image animation using diffusion model,” in CVPR, 2024, pp. 1481–1490

  39. [48]

    Conditional image-to-video generation with latent flow diffusion models,

    H. Ni, C. Shi, K. Li et al., “Conditional image-to-video generation with latent flow diffusion models,” in CVPR, 2023, pp. 18 444–18 455

  40. [49]

    BERT: pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M. Chang, K. Lee et al. , “BERT: pre-training of deep bidirectional transformers for language understanding,” in NAACL-HLT, 2019, pp. 4171–4186

  41. [50]

    Human motion transfer from poses in the wild,

    J. Ren, M. Chai, S. Tulyakov et al., “Human motion transfer from poses in the wild,” in ECCV Workshops, vol. 12537, 2020, pp. 262–279

  42. [51]

    Flownet 2.0: Evolution of optical flow estimation with deep networks,

    E. Ilg, N. Mayer, T. Saikia et al. , “Flownet 2.0: Evolution of optical flow estimation with deep networks,” in CVPR, 2017, pp. 1647–1655

  43. [52]

    ANR: articulated neural rendering for virtual avatars,

    A. Raj, J. Tanke, J. Hays et al., “ANR: articulated neural rendering for virtual avatars,” in CVPR, 2021, pp. 3722–3731

  44. [53]

    Least squares generative adversarial networks,

    X. Mao, Q. Li, H. Xie et al. , “Least squares generative adversarial networks,” in ICCV, 2017, pp. 2813–2821

  45. [54]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren et al. , “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

  46. [55]

    Rethinking atrous convolution for semantic image segmentation,

    L. Chen, G. Papandreou, F. Schroff et al., “Rethinking atrous convolution for semantic image segmentation,” CoRR, vol. abs/1706.05587, 2017

  47. [56]

    Robust high-resolution video matting with temporal guidance,

    S. Lin, L. Yang, I. Saleemi et al., “Robust high-resolution video matting with temporal guidance,” in WACV, 2022, pp. 3132–3141

  48. [57]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner et al., “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in NeurIPS, 2017, pp. 6626–6637

  49. [58]

    In defense of the triplet loss for person re-identification,

    A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,” CoRR, vol. abs/1703.07737, 2017

  50. [59]

    Realtime multi-person 2d pose estimation using part affinity fields,

    Z. Cao, T. Simon, S. Wei et al. , “Realtime multi-person 2d pose estimation using part affinity fields,” in CVPR, 2017, pp. 1302–1310

  51. [60]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in ICLR, 2015

  52. [61]

    Structure-aware motion transfer with deformable anchor model,

    J. Tao, B. Wang, B. Xu et al. , “Structure-aware motion transfer with deformable anchor model,” in CVPR, 2022, pp. 3627–3636

  53. [62]

    Image animation with perturbed masks,

    Y . Shalev and L. Wolf, “Image animation with perturbed masks,” in CVPR, 2022, pp. 3637–3646

  54. [63]

    Thin-plate spline motion model for image animation,

    J. Zhao and H. Zhang, “Thin-plate spline motion model for image animation,” in CVPR, 2022, pp. 3647–3656

  55. [64]

    Latent image animator: Learning to animate images via latent space navigation,

    Y . Wang, D. Yang, F. Br´emond et al., “Latent image animator: Learning to animate images via latent space navigation,” in ICLR, 2022

  56. [65]

    V oxceleb: Large-scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie et al., “V oxceleb: Large-scale speaker verification in the wild,” Comput. Speech Lang. , vol. 60, 2020

  57. [66]

    Animating arbitrary objects via deep motion transfer,

    A. Siarohin, S. Lathuili `ere, S. Tulyakov et al. , “Animating arbitrary objects via deep motion transfer,” in CVPR, 2019, pp. 2377–2386. THIS ARTICLE HAS BEEN ACCEPTED FOR PUBLICATION IN IEEE TRANSACTIONS ON MULTIMEDIA. DOI:XX.XXX/XXX.XX.XXXX 14 Yuzhu Ji (Member, IEEE) receiv...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.