pith. sign in

arxiv: 2606.19718 · v1 · pith:KRBEP3LFnew · submitted 2026-06-18 · 💻 cs.CV

One-Shot Novel View and Pose Human Image Synthesis via 3D Prior Guided Diffusion Model

Shenjian Gong , Kangkan Wang , Shanshan Zhang , Jian Yang This is my paper

Pith reviewed 2026-06-26 17:52 UTC · model grok-4.3

classification 💻 cs.CV
keywords human image synthesisnovel view synthesispose transferdiffusion model3D priorsone-shot generationconditional denoising
0
0 comments X

The pith

A conditional diffusion model uses 3D normal maps and color prompts from one human image to generate novel poses and views including occluded parts.

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

The paper seeks to improve one-shot synthesis of humans in new poses and viewpoints by framing the task as conditional denoising diffusion steps. Existing 2D keypoint transfer methods fail on ambiguous or complex poses while generalizable NeRF approaches struggle to recover missing regions without reliable point features. By injecting 3D normal maps for geometry and color prompts for appearance as conditions, the model progressively transfers the reference image to the target configuration. A final self-reconstruction refinement step sharpens details for unseen identities. Experiments on public datasets show the approach yields higher quality outputs and stronger cross-dataset generalization than prior techniques.

Core claim

Conditioning a denoising diffusion model on 3D normal maps and color prompts extracted from a single reference image enables high-quality synthesis of arbitrary target poses and novel views, including accurate recovery of occluded or invisible body parts through iterative conditional steps.

What carries the argument

3D prior guided diffusion model that conditions each denoising step on a 3D normal map for geometry and a color prompt for appearance.

If this is right

  • Synthesis of complex and arbitrary poses succeeds without dependence on ambiguous 2D keypoints.
  • Occluded and invisible human parts are recovered more accurately than in keypoint-based or NeRF-based baselines.
  • A self-reconstruction refinement step produces finer details when the model is applied to previously unseen persons.
  • Performance and generalization improve across multiple public human image datasets relative to earlier methods.

Where Pith is reading between the lines

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

  • The same conditioning strategy could be tested on non-human categories if comparable 3D priors can be obtained from single images.
  • Because the generation unfolds over many denoising steps, the framework may support progressive editing or multi-view consistency constraints without retraining.
  • Errors in the initial 3D prior extraction could be mitigated by the diffusion model's ability to sample plausible completions, suggesting robustness to imperfect geometry inputs.

Load-bearing premise

Reliable 3D normal maps and color prompts can be extracted from a single reference human image to serve as effective conditions for the diffusion process across arbitrary poses and occlusions.

What would settle it

A controlled test set of reference images where the extracted 3D normal maps are intentionally perturbed or incomplete, followed by measurement of whether generated occluded regions still match ground-truth target images.

Figures

Figures reproduced from arXiv: 2606.19718 by Jian Yang, Kangkan Wang, Shanshan Zhang, Shenjian Gong.

Figure 1
Figure 1. Figure 1: Overview of our proposed method. Given a reference image x0 and the target pose p 2d , our model predicts the target image y0 via a denoising diffusion model (Sec. 3.1). We introduce 3D prior guidance (Sec. 3.2) as additional conditions and a self-reconstruction based customized refinement (Sec. 3.3), to enhance image fidelity. 3. Proposed Method [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Failure cases of PIDM [5] under complex target poses with self occlusions. vided by x0 and the target 2D pose p 2d is rather limited. To better guide the generation, we explore to utilize 3D human priors produced from a 3D SMPL model, i.e., 3D nor￾mal map and color prompt. Compared with 2D pose, the 3D normal map can better model the human pose and shape and associate human correspondences between the targ… view at source ↗
Figure 3
Figure 3. Figure 3: Comparisons of 3D normal map and 2D pose. The 3D normal map is able to represent complex postures more accurate and better represent the human body shape than 2D pose. To better demonstrate the advantages of using the 3D normal map, we show some examples in [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The motivation and effects of color prompt. The 3D normal map ensures geometric consistency, while the color prompt further preserves appearance and texture details. Zoom in for better visualization. project all vertices of Mr to x0, so as to obtain the image coordinates on x0 and take the pixel colors for these projected vertices. Then, the vertex coordinates of Mr are replaced with the human mesh M gener… view at source ↗
Figure 5
Figure 5. Figure 5: Failure cases of PIDM [5] with wrong or missing details. Zoom in for better view. 1, the predicted face does not match the reference person and the white label on the shirt is missing. Similarly, in row 2, the generated man has a realistic but inaccurate face and loses his watch on his hand. The problem of failing to recover these local or high-frequency details is mainly caused by overfitting on the limit… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison to the state-of-the-art methods. We visualize the novel view (row 1∼ 3) and novel pose (row 4∼ 6) results of MPS-NeRF [2] , SHERF [4], PIDM [5], Disco [41], Champ [45] and our method on the RenderPeople dataset. Zoom in for better visualization. images under the challenging novel pose scenario. Additionally, [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison to the state-of-the-art methods. We visualize the novel view (row 1∼ 3) and novel pose (row 4∼ 6) results of MPS-NeRF [2] , SHERF [4], PIDM [5], Disco [41], Champ [45] and our method on the THuman [16] dataset. Zoom in for better visualization. 4.3. Ablation Studies In the following, we conduct ablation studies to demonstrate the effectiveness of different components of our method. A… view at source ↗
Figure 8
Figure 8. Figure 8: Step-by-step improvement of our method, indicating the effects of both 3D prior guidance and self-reconstruction based customized refinement. strates introducing the 3D normal map can enhance the synthesis quality obviously and confirms its critical role in geometry-aware synthesis. Additionally, one visualization example is shown in [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative results for cross-dataset (RenderPeople → THuman) validation. Better visual results indicate better cross-dataset generalization ability. Zoom in for better visualization. λ Novel View Novel Pose PSNR↑ SSIM↑ LPIPS↓ PSNR↑ SSIM↑ LPIPS↓ 0 21.61 0.89 0.09 20.96 0.87 0.10 50 23.23 0.91 0.07 22.24 0.89 0.08 100 23.29 0.91 0.07 22.36 0.90 0.07 200 23.20 0.91 0.07 22.25 0.90 0.08 [PITH_FULL_IMAGE:figu… view at source ↗
Figure 10
Figure 10. Figure 10: Application of single-image 3D human reconstruction using our method on RenderPeople [15]. For each person, we show the input reference image and reconstruction results in two different views. The 3D human meshes are recovered with about 15 images from different views generated using our method. 5. Conclusion In this work, we aim to synthesize novel view and pose human images of both high visual quality a… view at source ↗
read the original abstract

This paper addresses the challenge of one-shot novel view and pose human image synthesis. The existing methods transfer the reference human image to a target pose using a set of 2D pose keypoints or synthesize human images based on generalizable human NeRF which uses human model priors to extract point-wise features. However, pose transfer based methods can not handle complex human pose using ambiguous 2D pose as the condition, while generalizable human NeRFs may be inaccurate to recover occluded/invisiable human parts without extracted reliable features. To solve these problems, we propose a novel approach for novel view and pose synthesis from a singe human image via conditional denoising diffusion model. Our diffusion model divides the novel view and pose synthesis problem into a sequence of conditional denoising steps. Specifically, to generate humans with complex and arbitrary poses, we introduce 3D human priors, i.e., 3D normal map and color prompt, as geometry and color conditions into the generation process. By transferring the reference human into the target human with a series of diffusion steps, our diffusion model enables high-quality synthesis including the occluded/invisible parts. Further, we propose a self-reconstruction based customized refinement to enhance fine details when tested on novel persons.Experimental results on different public datasets demonstrate that our approach significantly outperforms previous methods and also shows better generalization ability across datasets. The code will be made publicly available at https://github.com/Yankeegsj/3DPGDM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a conditional denoising diffusion model for one-shot novel view and pose human image synthesis from a single reference image. It divides the task into sequential denoising steps conditioned on 3D human priors (normal maps and color prompts) for geometry and appearance, aiming to handle complex/arbitrary poses and occlusions better than 2D-pose transfer or generalizable NeRF methods; a self-reconstruction refinement step is added for novel persons, with claims of superior performance and cross-dataset generalization on public benchmarks.

Significance. If the empirical claims hold and the 3D-prior extraction proves robust, the work would offer a practical advance in conditional diffusion for human synthesis by explicitly injecting 3D geometry cues, potentially improving handling of invisible regions over purely 2D or implicit NeRF baselines. The promised public code release would support reproducibility.

major comments (2)
  1. [Abstract] Abstract: the central claim that 3D normal maps and color prompts extracted from one reference image serve as reliable conditions for arbitrary target poses and occlusions is load-bearing, yet the text provides no description of the extraction/fitting procedure, no ablation on its accuracy for occluded regions, and no preliminary validation that errors in the prior do not propagate through denoising.
  2. [Abstract] Abstract: the assertion that the method 'significantly outperforms previous methods' and shows 'better generalization ability across datasets' is presented without any quantitative tables, metrics, baselines, or experimental setup details, making it impossible to assess whether the reported superiority is supported by the data.
minor comments (1)
  1. [Abstract] Abstract: the phrasing 'transferring the reference human into the target human with a series of diffusion steps' is vague; clarify how the reference image is injected beyond the 3D priors.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate planned revisions to strengthen the abstract.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 3D normal maps and color prompts extracted from one reference image serve as reliable conditions for arbitrary target poses and occlusions is load-bearing, yet the text provides no description of the extraction/fitting procedure, no ablation on its accuracy for occluded regions, and no preliminary validation that errors in the prior do not propagate through denoising.

    Authors: We agree the abstract is a high-level summary and lacks these specifics. The extraction/fitting procedure (SMPL-based normal map generation and reference color prompting) is detailed in the Methods section. To address the concern directly in the abstract, we will add a concise clause describing the prior extraction and note that experiments validate robustness to prior inaccuracies in occluded areas. This will be incorporated in the revised version. revision: yes

  2. Referee: [Abstract] Abstract: the assertion that the method 'significantly outperforms previous methods' and shows 'better generalization ability across datasets' is presented without any quantitative tables, metrics, baselines, or experimental setup details, making it impossible to assess whether the reported superiority is supported by the data.

    Authors: Abstracts conventionally summarize outcomes at a high level without tables or full experimental protocols due to length constraints. The supporting quantitative results (metrics, baselines including 2D pose transfer and generalizable NeRF methods, cross-dataset evaluation) appear in the Experiments section. We will revise the abstract to include a brief qualifier such as 'as shown in our experiments' to better link the claim to the reported data. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical claims with no derivations

full rationale

The manuscript text (abstract and description) presents a conditional diffusion model using 3D normal maps and color prompts extracted from a single reference image, along with a self-reconstruction refinement step, but contains no equations, mathematical derivations, parameter-fitting procedures, or load-bearing self-citations. The central claim of outperforming prior methods rests entirely on experimental results across public datasets rather than any self-referential definitions, fitted inputs renamed as predictions, or ansatzes smuggled via citation. No steps reduce by construction to the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is supplied, so no concrete free parameters, axioms, or invented entities can be extracted from the manuscript text.

pith-pipeline@v0.9.1-grok · 5801 in / 1093 out tokens · 30769 ms · 2026-06-26T17:52:52.898059+00:00 · methodology

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

54 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    Neural human performer: Learning generalizable radiance fields for human performance rendering,

    Y . Kwon, D. Kim, D. Ceylan, and H. Fuchs, “Neural human performer: Learning generalizable radiance fields for human performance rendering,” inNeurIPS, 2021, pp. 24 741–24 752

    2021

  2. [2]

    Mps-nerf: Generalizable 3d human rendering from multiview images,

    X. Gao, J. Yang, J. Kim, S. Peng, Z. Liu, and X. Tong, “Mps-nerf: Generalizable 3d human rendering from multiview images,”IEEE TPAMI, 2022

    2022

  3. [3]

    Mononhr: Monocular neural human renderer,

    H. Choi, G. Moon, M. Armando, V . Leroy, K. M. Lee, and G. Rogez, “Mononhr: Monocular neural human renderer,” in3DV. IEEE, 2022, pp. 242–251

    2022

  4. [4]

    Sherf: Generalizable human nerf from a single image,

    S. Hu, F. Hong, L. Pan, H. Mei, L. Yang, and Z. Liu, “Sherf: Generalizable human nerf from a single image,”arXiv preprint arXiv:2303.12791, 2023

    work page arXiv 2023

  5. [5]

    Person image synthesis via denoising diffusion model,

    A. K. Bhunia, S. Khan, H. Cholakkal, R. M. Anwer, J. Laaksonen, M. Shah, and F. S. Khan, “Person image synthesis via denoising diffusion model,” inCVPR, 2023, pp. 5968–5976

    2023

  6. [6]

    Pose guided person image generation,

    L. Ma, X. Jia, Q. Sun, B. Schiele, T. Tuytelaars, and L. Van Gool, “Pose guided person image generation,” inNeurIPS, 2017

    2017

  7. [7]

    Progressive pose atten- tion transfer for person image generation,

    Z. Zhu, T. Huang, B. Shi, M. Yu, B. Wang, and X. Bai, “Progressive pose atten- tion transfer for person image generation,” inCVPR, 2019, pp. 2347–2356. 25

    2019

  8. [8]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” inECCV, 2020

    2020

  9. [9]

    Smpl: A skinned multi-person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “Smpl: A skinned multi-person linear model,”ACM Transactions on Graphics, vol. 34, no. 6, 2015

    2015

  10. [10]

    A variational u-net for conditional appear- ance and shape generation,

    P. Esser, E. Sutter, and B. Ommer, “A variational u-net for conditional appear- ance and shape generation,” inCVPR, 2018, pp. 8857–8866

    2018

  11. [11]

    Pot-gan: Pose trans- form gan for person image synthesis,

    T. Li, W. Zhang, R. Song, Z. Li, J. Liu, X. Li, and S. Lu, “Pot-gan: Pose trans- form gan for person image synthesis,”IEEE TIP, vol. 30, pp. 7677–7688, 2021

    2021

  12. [12]

    Pona: Pose-guided non-local attention for human pose transfer,

    K. Li, J. Zhang, Y . Liu, Y .-K. Lai, and Q. Dai, “Pona: Pose-guided non-local attention for human pose transfer,”IEEE TIP, vol. 29, pp. 9584–9599, 2020

    2020

  13. [13]

    Pose flow learning from person images for pose guided synthesis,

    H. Zheng, L. Chen, C. Xu, and J. Luo, “Pose flow learning from person images for pose guided synthesis,”IEEE TIP, vol. 30, pp. 1898–1909, 2020

    1909

  14. [14]

    Towards fine-grained human pose transfer with detail replenishing network,

    L. Yang, P. Wang, C. Liu, Z. Gao, P. Ren, X. Zhang, S. Wang, S. Ma, X. Hua, and W. Gao, “Towards fine-grained human pose transfer with detail replenishing network,”IEEE TIP, vol. 30, pp. 2422–2435, 2021

    2021

  15. [15]

    Renderpeople,

    “Renderpeople,” https://renderpeople.com/3d-people, 2018

    2018

  16. [16]

    Deephuman: 3d human recon- struction from a single image,

    Z. Zheng, T. Yu, Y . Wei, Q. Dai, and Y . Liu, “Deephuman: 3d human recon- struction from a single image,” inICCV, 2019, pp. 7739–7749

    2019

  17. [17]

    Humannerf: Free-viewpoint rendering of moving people from monocular video,

    C.-Y . Weng, B. Curless, P. P. Srinivasan, J. T. Barron, and I. Kemelmacher- Shlizerman, “Humannerf: Free-viewpoint rendering of moving people from monocular video,” inCVPR, 2022, pp. 16 210–16 220

    2022

  18. [18]

    Ani- matable neural radiance fields for modeling dynamic human bodies,

    S. Peng, J. Dong, Q. Wang, S. Zhang, Q. Shuai, X. Zhou, and H. Bao, “Ani- matable neural radiance fields for modeling dynamic human bodies,” inICCV, 2021, pp. 14 314–14 323. 26

    2021

  19. [19]

    Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans,

    S. Peng, Y . Zhang, Y . Xu, Q. Wang, Q. Shuai, H. Bao, and X. Zhou, “Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans,” inCVPR, 2021, pp. 9054–9063

    2021

  20. [20]

    Human101: Training 100+fps human gaussians in 100s from 1 view,

    M. Li, J. Tao, Z. Yang, and Y . Yang, “Human101: Training 100+fps human gaussians in 100s from 1 view,”arXiv preprint arXiv:2312.15258, 2023

    work page arXiv 2023

  21. [21]

    Gauhuman: Articulated gaussian splatting from monocular human videos,

    S. Hu and Z. Liu, “Gauhuman: Articulated gaussian splatting from monocular human videos,”arXiv preprint arXiv:2312.02973, 2023

    work page arXiv 2023

  22. [22]

    Hugs: Human gaussian splats,

    M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan, “Hugs: Human gaussian splats,”arXiv preprint arXiv:2311.17910, 2023

    work page arXiv 2023

  23. [23]

    Gaussianbody: Clothed hu- man reconstruction via 3d gaussian splatting,

    M. Li, S. Yao, Z. Xie, K. Chen, and Y .-G. Jiang, “Gaussianbody: Clothed hu- man reconstruction via 3d gaussian splatting,”arXiv preprint arXiv:2401.09720, 2024

    work page arXiv 2024

  24. [24]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, 2023

    2023

  25. [25]

    Svad: From single image to 3d avatar via synthetic data generation with video diffusion and data augmentation,

    Y . Choi, “Svad: From single image to 3d avatar via synthetic data generation with video diffusion and data augmentation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 3137–3147

    2025

  26. [26]

    Disco4d: Disentan- gled 4d human generation and animation from a single image,

    H. E. Pang, S. Liu, Z. Cai, L. Yang, T. Zhang, and Z. Liu, “Disco4d: Disentan- gled 4d human generation and animation from a single image,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 331– 26 344

    2025

  27. [27]

    Sings: Animatable single-image human gaussian splats with kinematic priors,

    Y . Wu, X. Chen, W. Li, S. Jia, H. Wei, K. Feng, J. Chen, Y . Li, A. He, W. Zhang et al., “Sings: Animatable single-image human gaussian splats with kinematic priors,” inProceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 5571–5580. 27

    2025

  28. [28]

    Deep image spatial transformation for person image generation,

    Y . Ren, X. Yu, J. Chen, T. H. Li, and G. Li, “Deep image spatial transformation for person image generation,” inCVPR, 2020, pp. 7690–7699

    2020

  29. [29]

    Pise: Person image synthesis and editing with decoupled gan,

    J. Zhang, K. Li, Y .-K. Lai, and J. Yang, “Pise: Person image synthesis and editing with decoupled gan,” inCVPR, 2021, pp. 7982–7990

    2021

  30. [30]

    Deformable gans for pose-based human image generation,

    A. Siarohin, E. Sangineto, S. Lathuiliere, and N. Sebe, “Deformable gans for pose-based human image generation,” inCVPR, 2018, pp. 3408–3416

    2018

  31. [31]

    Dense intrinsic appearance flow for human pose transfer,

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

    2019

  32. [32]

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

    W. Liu, Z. Piao, J. Min, W. Luo, L. Ma, and S. Gao, “Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis,” inICCV, 2019, pp. 5904–5913

    2019

  33. [34]

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

    L. Hu, “Animate anyone: Consistent and controllable image-to-video synthesis for character animation,” inCVPR, 2024, pp. 8153–8163

    2024

  34. [35]

    Conditional Generative Adversarial Nets

    M. Mirza and S. Osindero, “Conditional generative adversarial nets,”arXiv preprint arXiv:1411.1784, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  35. [36]

    Liquid warping gan with at- tention: A unified framework for human image synthesis,

    W. Liu, Z. Piao, Z. Tu, W. Luo, L. Ma, and S. Gao, “Liquid warping gan with at- tention: A unified framework for human image synthesis,”IEEE TPAMI, vol. 44, no. 9, pp. 5115–5133, 2022

    2022

  36. [37]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in NeurIPS, 2020, pp. 6840–6851

    2020

  37. [38]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” in NeurIPS, 2021, pp. 8780–8794. 28

    2021

  38. [39]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models,

    A. Q. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen, “Glide: Towards photorealistic image generation and editing with text-guided diffusion models,” inICML, 2022, pp. 16 784–16 804

    2022

  39. [40]

    Classifier-Free Diffusion Guidance

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  40. [41]

    Disco: Disentangled control for realistic human dance generation,

    T. Wang, L. Li, K. Lin, Y . Zhai, C.-C. Lin, Z. Yang, H. Zhang, Z. Liu, and L. Wang, “Disco: Disentangled control for realistic human dance generation,” in CVPR, 2024, pp. 9326–9336

    2024

  41. [42]

    Magicanimate: Temporally consistent human image animation using dif- fusion model,

    Z. Xu, J. Zhang, J. H. Liew, H. Yan, J.-W. Liu, C. Zhang, J. Feng, and M. Z. Shou, “Magicanimate: Temporally consistent human image animation using dif- fusion model,” inCVPR, 2024, pp. 1481–1490

    2024

  42. [43]

    Towards multiple character image animation through en- hancing implicit decoupling,

    J. Xue, H. Wang, Q. Tian, Y . Ma, A. Wang, Z. Zhao, S. Min, W. Zhao, K. Zhang, H.-Y . Shumet al., “Towards multiple character image animation through en- hancing implicit decoupling,” inICLR, 2025

    2025

  43. [44]

    360-degree human video gener- ation with 4d diffusion transformer,

    R. Shao, Y . Pang, Z. Zheng, J. Sun, and Y . Liu, “360-degree human video gener- ation with 4d diffusion transformer,”ACM TOG, vol. 43, no. 6, pp. 1–13, 2024

    2024

  44. [45]

    Champ: Controllable and consistent human image animation with 3d paramet- ric guidance,

    S. Zhu, J. L. Chen, Z. Dai, Z. Dong, Y . Xu, X. Cao, Y . Yao, H. Zhu, and S. Zhu, “Champ: Controllable and consistent human image animation with 3d paramet- ric guidance,” inECCV. Springer, 2024, pp. 145–162

    2024

  45. [46]

    Improved denoising diffusion probabilistic mod- els,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic mod- els,” inICML, 2021, pp. 8162–8171

    2021

  46. [47]

    Learning to count everything,

    V . Ranjan, U. Sharma, T. Nguyen, and M. Hoai, “Learning to count everything,” inCVPR, 2021, pp. 3394–3403

    2021

  47. [48]

    One-shot generative domain adaptation,

    C. Yang, Y . Shen, Z. Zhang, Y . Xu, J. Zhu, Z. Wu, and B. Zhou, “One-shot generative domain adaptation,” inICCV, 2023, pp. 7733–7742. 29

    2023

  48. [49]

    Test-time domain adaptation by learning domain-aware batch normalization,

    Y . Wu, Z. Chi, Y . Wang, K. N. Plataniotis, and S. Feng, “Test-time domain adaptation by learning domain-aware batch normalization,” inAAAI, vol. 38, no. 14, 2024, pp. 15 961–15 969

    2024

  49. [50]

    Image quality assessment through fsim, ssim, mse and psnr—a comparative study,

    U. Sara, M. Akter, and M. S. Uddin, “Image quality assessment through fsim, ssim, mse and psnr—a comparative study,”Journal of Computer and Communi- cations, vol. 7, no. 3, pp. 8–18, 2019

    2019

  50. [51]

    Image quality as- sessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality as- sessment: from error visibility to structural similarity,”IEEE TIP, vol. 13, no. 4, pp. 600–612, 2004

    2004

  51. [52]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” inCVPR, 2018, pp. 586– 595

    2018

  52. [53]

    Controllable person image synthesis with pose-constrained latent diffusion,

    X. Han, X. Zhu, J. Deng, Y .-Z. Song, and T. Xiang, “Controllable person image synthesis with pose-constrained latent diffusion,” inICCV, 2023, pp. 22 768– 22 777

    2023

  53. [54]

    Humans in 4d: Reconstructing and tracking humans with transformers,

    S. Goel, G. Pavlakos, J. Rajasegaran, A. Kanazawa, and J. Malik, “Humans in 4d: Reconstructing and tracking humans with transformers,” inICCV, 2023, pp. 14 783–14 794

    2023

  54. [55]

    Human performance modeling and rendering via neural animated mesh,

    F. Zhao, Y . Jiang, K. Yao, J. Zhang, L. Wang, H. Dai, Y . Zhong, Y . Zhang, M. Wu, L. Xuet al., “Human performance modeling and rendering via neural animated mesh,”ACM TOG, vol. 41, no. 6, pp. 1–17, 2022. 30

    2022