Pith. sign in

REVIEW 3 major objections 5 minor 39 references

TruePose: Human-Parsing-guided Attention Diffusion for Full-ID Preserving Pose Transfer

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Pose-transfer diffusion models lose clothing because their attention ignores garment regions—parsing-guided attention reweights it to keep face and clothing intact.

desk verdict Solid incremental pose-transfer paper with strong reported numbers, but the written condition set omits the target pose, making the central claim unverifiable from the text. read the letter →

arxiv 2502.03426 v1 pith:36KBMCTU submitted 2025-02-05 cs.CV cs.AI

Zhihong Xu , Dongxia Wang , Peng Du , Yang Cao , Qing Guo This is my paper · ORCID
classification cs.CVcs.AI
keywords pose-guidedpersonimagesynthesisdiffusionmodelshumanparsingattentionreweightingclothingpreservationidentitySiamesenetworkslatent
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

The paper sets out to show that poor clothing preservation in diffusion-based pose transfer is not an inevitable limitation of generative models but a fixable attention problem: the diffusion model's attention maps pay too little attention to clothing regions, so those patterns fade out during denoising. To fix it, TruePose introduces a Siamese network of two identical UNets—one denoising the target pose, one encoding the source image—with two attention modules that explicitly steer the generation toward face and garment regions using a person-parsing map. If the claim is right, then pose transfer can preserve not just the face but the full clothing identity, including text, logos, and textures, which matters for fashion e-commerce where reshoots are expensive and garment design is protected. The paper reports that this outperforms 13 prior methods on LPIPS, PSNR, and user-preference metrics on DeepFashion and the in-the-wild WPose dataset.

What carries the argument

The load-bearing object is the human-parsing-aware Siamese network: two identical UNets, SourceNet for source-image embeddings and TargetNet for denoising, connected by two attention modules. HPFA computes a mask from the source parsing map and target skeleton, reweights the source attention logits so masked regions get multiplier 1+δ and outside regions get σ, then fuses through cross-attention; CAA finds the K parsing regions with lowest average attention and applies cross-attention with their region-level semantic embeddings to recover under-represented details such as faces. These modules are inserted at multiple TargetNet layers, and the whole model is trained with classifier-free guidance.

What would settle it

Compute clothing-region LPIPS on DeepFashion test pairs for the full model and for a variant where the parsing mask reweighting (HPFA) is disabled at inference; if the score barely changes, the paper's claim that attention neglect of clothing is the bottleneck is not supported.

Watch

Extended reading notes

Core claim

TruePose claims that full identity—face and clothing—can survive a large pose change if the diffusion model's attention is reweighted by human-parsing maps and refined by region-level semantic embeddings. The paper's view is that prior diffusion baselines (CFLD and PCDM) fail on clothing not because of the generator but because their encoders' attention maps pay little attention to clothing regions, and it supports this with a 50-sample statistical analysis of attention values on head, body, and cloth regions. The proposed architecture then inserts two modules into a latent diffusion UNet: HPFA reweights source embeddings' self-attention using a mask of parsing regions that appear in the target pose, and CAA selects the K lowest-attention parsing regions and aligns them with the semantic embeddings of those regions. On DeepFashion and WPose, the paper reports that this yields the best LPIPS and PSNR among 13 baselines while retaining text and texture patterns.

Load-bearing premise

The paper's whole design assumes that clothing details are lost because the diffusion model's attention pays too little attention to clothing regions; if the true bottleneck is in the image encoder or decoder instead, reweighting attention with parsing masks would not fix it.

Editorial extensions

If this is right

  • Garment text, logos, regular textures, and irregular patterns should survive large pose changes instead of being averaged away, because attention is forced to concentrate on the masked clothing regions.
  • Facial identity and clothing identity are treated as separable but jointly preserved: the parsing map selects face and clothes regions separately, and the CAA module specifically rescues low-attention regions such as the face.
  • The method remains stable as the gap between source and target pose grows, since the mask is built from the target pose's skeleton categories, so only the source regions relevant to the target are reweighted.
  • The modules transfer to in-the-wild images with messy backgrounds (WPose), where the reported gains over the two latest diffusion baselines are larger than on the clean-background benchmark.
  • A mask-based sampling trick enables training-free style transfer, letting a user keep a region of interest from a reference image while the rest follows the source image's pose and appearance.

Reading between the lines

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

  • Editorial inference: the mask-reweighting recipe is not tied to human bodies; applied to any object with a semantic part segmentation, it could improve pattern retention in other conditional diffusion tasks.
  • Editorial inference: if the attention-neglect diagnosis is the true bottleneck, similar gains should be obtainable by inserting HPFA-style reweighting into other latent-diffusion person editors with only lightweight fine-tuning, since the fix acts on attention rather than on the whole generator.
  • Editorial inference: the same K-lowest-region refinement could be used at inference to boost under-represented parts in generic text-to-image or image editing, provided a part segmentation and a region-level semantic encoder exist.
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 / 5 minor

Summary. The paper proposes TruePose, a pose-guided person image synthesis method built on Stable Diffusion, featuring a Siamese pair of UNets (TargetNet and SourceNet), a human-parsing-guided fusion attention module (HPFA), and a CLIP-guided attention alignment module (CAA). The authors argue that existing diffusion-based PGPIS methods lose clothing patterns because their attention modules under-weight clothing regions, and they present quantitative and qualitative comparisons against 13 baselines on DeepFashion (two resolutions) and WPose, plus ablations and a style-transfer application.

Significance. If the reported results are substantiated, TruePose would be a practically useful contribution, particularly for fashion-oriented PGPIS where preserving clothing texture and text patterns matters. The paper's strengths are its broad evaluation across 13 baselines, two resolutions, a held-out in-the-wild dataset, and explicit ablations of the proposed modules; the style-transfer application additionally shows a potentially useful downstream capability. However, the central architectural description currently omits how the target pose actually enters the denoising network, which blocks verification of the paper's core pose-transfer claim. The reported gains on LPIPS and user-study metrics would be meaningful if the architecture were fully and correctly described.

major comments (3)
  1. [Section 4.1, Eq. (5)] The condition set is written as C = {F', Fclip}, where F' is the SourceNet embedding of the source image and Fclip is the set of CLIP region embeddings. The target pose p_tau appears nowhere in C, and Section 4.2 uses p_tau only to select which parsing categories form the mask M; it does not inject any spatial keypoint coordinates or a pose tensor into the denoiser. The implementation details mention a PoseEncoder with four convolutional layers, but do not state where its output is inserted. If Eq. (5) is literal, TargetNet has no information about the target limb positions and the pose-transfer results are unexplainable. If the PoseEncoder output is in fact injected into the UNet, that injection is a central architectural component missing from the stated formulation. Either way, the paper as written does not allow a reader to verify how target pose controls generation, so the reported pose-transfer comparisons cannot be checked without code. This is the load-bearing issue for the paper's central claim and must be fixed by giving the complete condition set and the pose-injection path, with equations or a precise description of where PoseEncoder features enter TargetNet.
  2. [Section 4.2, Eqs. (6)-(7)] There is a dimension mismatch in the definition of the mask-weight matrix M'. The mask M is resized to match F' of shape h*w*c, so M' is indexed as an h-by-w spatial matrix, and Eq. (6) defines M'_{i,j} with i in {0,...,h} and j in {0,...,w}. However, Eq. (7) applies M' by elementwise multiplication with the attention map SoftMax((Q'K'^T)/sqrt(d)), which is an (h*w)-by-(h*w) matrix under standard self-attention. These dimensions are incompatible. The paper needs to clarify whether M' is tiled/broadcast over attention heads and token pairs, whether the attention map is computed in a spatially factorized form, or whether some other reshaping is intended. Without this clarification, the HPFA module cannot be implemented from the text, and the ablations for HPFA cannot be independently reproduced.
  3. [Section 3.2] The paper's motivating claim that poor clothing preservation 'primarily stems from the conditional diffusion model's attention modules failing to adequately capture and preserve clothing patterns' is supported only by a statistical analysis over 50 examples from two baselines, with no error bars, no significance test, and no controlled test of whether the attention deficit is causal or merely correlated. The subsequent ablations show that the proposed modules improve metrics, but they do not isolate whether the bottleneck is the attention mechanism rather than, for example, VAE latent capacity or decoder fidelity. Since every proposed component is designed to fix exactly this supposed deficit, the causal claim is load-bearing for the architecture's motivation. I recommend either strengthening this analysis with a larger sample, confidence intervals, and a baseline that manipulates attention without the parsing mask, or softening the causal wording to a correlational observation.
minor comments (5)
  1. [Throughout] There are several typos and inconsistent terms that should be corrected: 'idenitical' in Section 4.1, 'emebdding' in Section 4.3, 'Classifer-free' in Section 4.4, 'stacket' in Section 5.3, and the user-study metric is written as 'Jab' in Fig. 5 and the text but as 'J2b' in Section 5.1.
  2. [Table 1] The Ground Truth row reports FID of 8.028 on DeepFashion (256x176), which is higher than several generated methods (e.g., PIDM at 6.663). Since lower FID is better, this is surprising and should be explained, for example by clarifying what reference distribution the FID is computed against.
  3. [Section 4.4] The hyperparameters sigma (0.3), delta (0.6), K (2), the classifier-free guidance scale (3.5), and the DDIM sampling step count (35) are stated only as empirical choices. A short sensitivity analysis or at least a statement of how they were selected would help assess robustness, especially because Eq. (6) directly controls the strength of the proposed attention reweighting.
  4. [Section 5.4] In the style-transfer application, the relation y_t = m ⊙ y_t + (1-m) ⊙ y_ref_t mixes notation: y_t is earlier used for the noisy latent, but here it appears to denote the predicted noise or reconstructed latent. The notation should be made consistent and the equation should state which quantity is blended at each sampling step.
  5. [Section 5.2 and Fig. 5] The user study uses only 30 volunteers and reports no confidence intervals or significance tests for the R2G/G2R/J2b differences. Reporting these would strengthen the claim that the improvements are not due to chance.

Circularity Check

1 steps flagged · score 2.0 of 10

Main quantitative evaluation is independent, so no central circularity; one attention-map 'validation' is self-confirming by construction, giving a minor score.

  1. self definitional [Sec. 4.2, Eqs. (6)-(7), and Sec. 4.2 attention-map comparison (Fig. 3)]
    "we propose human-parsing-guided fusion attention, which leverages binary mask M to reweight the embeddings of F′. ... Then, we compute the hidden states via reweighted attention: H′ = RwSelfAtt(Q′, K′, V′, M′), = SoftMax((Q′K′⊤/√d) ⊙ M′)V′. ... we find that the model focuses more on the areas indicated by the mask during the diffusion process, allowing for better feature extraction and effective pose transfer."

    In Eq. (6), inside the mask positive attention is multiplied by 1+δ and negative attention by δ, while outside the mask positive attention is multiplied by σ and negative attention by 1+σ, with δ=0.6 and σ=0.3. Eq. (7) applies these multiplicative weights before softmax, so masked positions receive mechanically larger softmax mass. The claim that the model 'focuses more on the areas indicated by the mask' is therefore an algebraic consequence of the reweighting definition, not an independent empirical observation. Reporting attention maps (Fig. 3 and the attention-score comparison for 'Ours' in Fig. 2) as evidence for the mechanism is self-confirming.

full rationale

The paper's headline comparisons (Table 1) rest on external, held-out metrics — LPIPS, SSIM, PSNR, FID, Q-Align, and a user study — on DeepFashion at two resolutions and on WPose in the wild, so the central claim of improved pose transfer and clothing preservation is not forced by construction. The empirical motivation in Sec. 3.2 is based on only 50 examples and is weakly supported, but that is a statistical strength issue, not circularity. The one genuinely circular element is the attention-map 'validation' in Sec. 4.2/Fig. 3: the reweighting in Eqs. (6)-(7) is defined precisely to boost attention inside the parsing mask, so demonstrating that the model focuses more on those regions is a restatement of the mechanism rather than evidence for it. I also note that Eq. (5) lists conditions C = {F′, Fclip} with no target-pose tensor, and the PoseEncoder output mentioned in Sec. 4.4 is never explicitly inserted into the equations; this is a reproducibility and correctness gap, not a circular reduction, and it does not change the circularity score. No load-bearing self-citations or imported uniqueness arguments are present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard diffusion assumptions, the accuracy of an external human parser, the untested causal hypothesis that attention deficits explain clothing loss, and the design hypothesis that attention reweighting plus CLIP alignment can transfer textures. Five hyperparameters are set by hand or by grid choice, and no new entities are introduced.

free parameters (5)
  • delta = 0.6
    Additive weight applied to masked, high-attention positions in Eq. (6); selected empirically in Sec 4.2 to avoid overfitting.
  • sigma = 0.3
    Additive weight applied to unmasked, high-attention positions in Eq. (6); selected empirically in Sec 4.2.
  • K = 2
    Number of lowest-attention parsing regions selected for CLIP alignment in Sec 4.3; set empirically.
  • classifier-free guidance scale = 3.5
    Inference-time guidance strength in Eq. (11), stated in Sec 4.4.
  • DDIM sampling steps = 35
    Number of denoising steps used at inference, stated in Sec 4.4.
assumptions (5)
  • standard math The diffusion forward process Eq. (1) and noise-prediction objective Eq. (2) define a valid training target.
    The method inherits stable diffusion training as in Ho et al. [7] and Rombach et al. [24]; assumed correct without re-derivation.
  • domain assumption The external human parser of [11] produces parsing maps accurate enough to guide attention.
    HPFA reweights attention according to parsing masks (Sec 4.2); systematic parsing errors would misguide the reweighting.
  • domain assumption Low attention on clothing regions in baseline models is a cause, not merely a correlate, of clothing detail loss.
    Sec 3.2 shows correlation on 50 examples; the design of HPFA assumes this causality.
  • ad hoc to paper Reweighting attention by parsing masks and aligning CLIP regional embeddings transfers clothing patterns into generated images.
    This is the core design hypothesis of the two proposed modules; no formal or independent empirical proof is given beyond the paper's own ablations.
  • standard math Classifier-free guidance, Eq. (11), improves sample quality as used.
    From Ho and Salimans [6]; standard in diffusion generation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TruePose: Human-Parsing-guided Attention Diffusion for Full-ID Preserving Pose Transfer." pith.science (2026). https://pith.science/paper/36KBMCTU

@misc{pith2026250203426,
  author       = {Pith},
  title        = {Pith review of: TruePose: Human-Parsing-guided Attention Diffusion for Full-ID Preserving Pose Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36KBMCTU}},
  note         = {Machine review of arXiv:2502.03426}
}
read the original abstract

Pose-Guided Person Image Synthesis (PGPIS) generates images that maintain a subject's identity from a source image while adopting a specified target pose (e.g., skeleton). While diffusion-based PGPIS methods effectively preserve facial features during pose transformation, they often struggle to accurately maintain clothing details from the source image throughout the diffusion process. This limitation becomes particularly problematic when there is a substantial difference between the source and target poses, significantly impacting PGPIS applications in the fashion industry where clothing style preservation is crucial for copyright protection. Our analysis reveals that this limitation primarily stems from the conditional diffusion model's attention modules failing to adequately capture and preserve clothing patterns. To address this limitation, we propose human-parsing-guided attention diffusion, a novel approach that effectively preserves both facial and clothing appearance while generating high-quality results. We propose a human-parsing-aware Siamese network that consists of three key components: dual identical UNets (TargetNet for diffusion denoising and SourceNet for source image embedding extraction), a human-parsing-guided fusion attention (HPFA), and a CLIP-guided attention alignment (CAA). The HPFA and CAA modules can embed the face and clothes patterns into the target image generation adaptively and effectively. Extensive experiments on both the in-shop clothes retrieval benchmark and the latest in-the-wild human editing dataset demonstrate our method's significant advantages over 13 baseline approaches for preserving both facial and clothes appearance in the source image.

Figures

Figures reproduced from arXiv: 2502.03426 by the authors.

Figure 1
Figure 1. Pose-guided person image synthesis (PGPIS) task and comparison among CFLD [15], PCDM (ICLR’24) [26], and our methods. The two SOTA methods fail to preserve the clothing patterns and textures. The main reason is that the image encoder overlooks the clothing details (See the “Feature Attention Map (Feat. Att. Map)"). In contrast, our method can generate high-quality images with preserved face and clothing patterns. 4 … view at source ↗
Figure 2
Figure 2. Top: compar￾ing average attention scores of dif￾ferent regions within source images. Bottom: visualization results of attention maps of our method and two baseline methods. train the UNet ϵθ(·), the predicted noise should be the same as the sampled noise during the forward process \label {eq:loss_diffusion} \mathcal {L}_\text {mse} = \mathds {E}_{\mathbf {z}_0, \mathcal {C}, \epsilon , t}(\|\epsilon -\epsilon _\thet… view at source ↗
Figure 3
Figure 3. Pipeline of the proposed human-parsing-guided attention diffusion model. transformations, as it can lead to distortion of image gener￾ation(more details in Sec. 5.3). This limitation arises from the inability to effectively guide the model’s focus toward the interest region. In response, we propose human-parsing￾guided fusion attention, which leverages binary mask M to reweight the embeddings of F ′ . Specifically, … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons with PISE [34], ADGAN[19], SPGNet [16], CASD [38], DPTN [35], PIDM [1], NTED [23], PCDM [26] and CFLD [15]. proposed model’s ability to better capture fine-grained cloth￾ing details, resulting in more realistic and higher-quality images. Further…
Figure 7
Figure 7. Figure 7: Style transfer results of our method. We can edit the reference image to incorporate pattern features from the source image while ensuring consistency in appearance and pose. features while minimizing distortions caused by irrelevant regions. Consequently, B3 achieves …
Figure 6
Figure 6. Figure 6: Visual￾ization results of ab￾lation study one set, we select 30 images with complex patterns from each method. In the other set, we randomly select 30 images for fair comparison. Our method significantly outperformed others on both sets, achieving scores of 70.2 and 53…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 25 canonical work pages

  1. [1]

    Person image synthesis via de- noising diffusion model

    Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via de- noising diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5968–5976, 2023. 1, 2, 6, 7

    work page 2023

  2. [2]

    A varia- tional u-net for conditional appearance and shape generation

    Patrick Esser, Ekaterina Sutter, and Björn Ommer. A varia- tional u-net for conditional appearance and shape generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8857–8866, 2018. 1, 2

    work page 2018

  3. [3]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2

    2014

  4. [4]

    Controllable person image synthesis with pose- constrained latent diffusion

    Xiao Han, Xiatian Zhu, Jiankang Deng, Yi-Zhe Song, and Tao Xiang. Controllable person image synthesis with pose- constrained latent diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22768– 22777, 2023. 2, 3, 6, 7

    work page 2023

  5. [5]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bern- hard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6

    2017

  6. [6]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 6

    arXiv 2022

  7. [7]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 1, 2, 3, 6

    work page 2020

  8. [8]

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

    Li Hu. Animate anyone: Consistent and controllable image- to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153–8163, 2024. 4, 8

    work page 2024

Show all 39 references
  1. [9]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. arXiv preprint arXiv:1312.6114, 2013. 3

    2013 arXiv

  2. [10]

    Unihuman: A unified model for editing human images in the wild

    Nannan Li, Qing Liu, Krishna Kumar Singh, Yilin Wang, Jianming Zhang, Bryan A Plummer, and Zhe Lin. Unihuman: A unified model for editing human images in the wild. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2039–2048, 2024. 6

    2024

  3. [11]

    Self- correction for human parsing

    Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self- correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 4

    2020

  4. [12]

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

    Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5904–5913, 2019. 1

    2019

  5. [13]

    Deepfashion: Powering robust clothes recognition and retrieval with rich annotations

    Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR),

  6. [14]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 3

    2021

  7. [15]

    Coarse-to-fine latent diffusion for pose-guided person image synthesis

    Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jian- Huang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In CVPR, 2024. 1, 2, 3, 6, 7

    2024

  8. [16]

    Learning semantic per- son image generation by region-adaptive normalization

    Zhengyao Lv, Xiaoming Li, Xin Li, Fu Li, Tianwei Lin, Dongliang He, and Wangmeng Zuo. Learning semantic per- son image generation by region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10806–10815, 2021. 2, 6, 7

    2021

  9. [17]

    Pose guided person image genera- tion

    Liqian Ma, Xu Jia, Qianru Sun, Bernt Schiele, Tinne Tuyte- laars, and Luc Van Gool. Pose guided person image genera- tion. Advances in neural information processing systems, 30,

  10. [18]

    Disentangled person image generation

    Liqian Ma, Qianru Sun, Stamatios Georgoulis, Luc Van Gool, Bernt Schiele, and Mario Fritz. Disentangled person image generation. In Proceedings of the IEEE conference on com- puter vision and pattern recognition , pages 99–108, 2018. 2

    2018

  11. [19]

    Controllable person image synthesis with attribute-decomposed gan

    Yifang Men, Yiming Mao, Yuning Jiang, Wei-Ying Ma, and Zhouhui Lian. Controllable person image synthesis with attribute-decomposed gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5084–5093, 2020. 2, 6, 7

    2020

  12. [20]

    Conditional generative adversarial nets

    Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2, 3

    2014 arXiv

  13. [21]

    T2i- adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i- adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3

    2023 arXiv

  14. [22]

    Deep image spatial transformation for person image generation

    Yurui Ren, Xiaoming Yu, Junming Chen, Thomas H Li, and Ge Li. Deep image spatial transformation for person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7690–7699,

  15. [23]

    Neural texture extraction and distribution for controllable person image synthesis

    Yurui Ren, Xiaoqing Fan, Ge Li, Shan Liu, and Thomas H Li. Neural texture extraction and distribution for controllable person image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13535–13544, 2022. 2, 6, 7

    2022

  16. [24]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 3

    2022

  17. [25]

    Style and pose control for image synthesis of humans from a single monocular view

    Kripasindhu Sarkar, Vladislav Golyanik, Lingjie Liu, and Christian Theobalt. Style and pose control for image synthesis of humans from a single monocular view. arXiv preprint arXiv:2102.11263, 2021. 1, 2

    2021 arXiv

  18. [26]

    Advancing pose-guided image synthesis with progressive conditional diffusion models

    Fei Shen, Hu Ye, Jun Zhang, Cong Wang, Xiao Han, and Yang Wei. Advancing pose-guided image synthesis with progressive conditional diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7

    2024

  19. [27]

    Deformable gans for pose-based human image generation

    Aliaksandr Siarohin, Enver Sangineto, Stéphane Lathuiliere, and Nicu Sebe. Deformable gans for pose-based human image generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3408–3416,

  20. [28]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502,

    2010 arXiv

  21. [29]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3

    2011 arXiv

  22. [30]

    Xinggan for person image generation

    Hao Tang, Song Bai, Li Zhang, Philip HS Torr, and Nicu Sebe. Xinggan for person image generation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV 16, pages 717–734. Springer, 2020. 1

    2020

  23. [31]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 6

    2004

  24. [32]

    Q-align: Teaching lmms for vi- sual scoring via discrete text-defined levels

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for vi- sual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 6

    2023 arXiv

  25. [33]

    Magicanimate: Temporally consistent human im- age animation using diffusion model

    Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human im- age animation using diffusion model. arXiv preprint arXiv:2311.16498, 2023. 4

    2023 arXiv

  26. [34]

    Pise: Person image synthesis and editing with decoupled gan

    Jinsong Zhang, Kun Li, Yu-Kun Lai, and Jingyu Yang. Pise: Person image synthesis and editing with decoupled gan. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 7982–7990, 2021. 1, 2, 6, 7

    2021

  27. [35]

    Exploring dual-task correlation for pose guided person image generation

    Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 7713–7722, 2022. 1, 2, 6, 7

    2022

  28. [36]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018. 6

    2018

  29. [37]

    Cocos- net v2: Full-resolution correspondence learning for image translation

    Xingran Zhou, Bo Zhang, Ting Zhang, Pan Zhang, Jianmin Bao, Dong Chen, Zhongfei Zhang, and Fang Wen. Cocos- net v2: Full-resolution correspondence learning for image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11465–...

    2021

  30. [38]

    Cross attention based style distribution for controllable person image synthesis

    Xinyue Zhou, Mingyu Yin, Xinyuan Chen, Li Sun, Changxin Gao, and Qingli Li. Cross attention based style distribution for controllable person image synthesis. In European Conference on Computer Vision, pages 161–178. Springer, 2022. 1, 2, 6, 7

    2022

  31. [39]

    Progressive pose attention transfer for person image generation

    Zhen Zhu, Tengteng Huang, Baoguang Shi, Miao Yu, Bofei Wang, and Xiang Bai. Progressive pose attention transfer for person image generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 2347–2356, 2019. 1, 6

    2019

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.