Pith. sign in

REVIEW 5 major objections 6 minor 57 references

Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that portrait shadow removal can be cast as diffusion-based inpainting, conditioned on a shadow-independent structure map and refined by gradient-guided detail restoration, and that this pipeline outperforms existing…

desk verdict Solid conditional accept: novel structure-guided inpainting for portrait shadow removal, but the shadow-independence claim rests on synthetic training and a small test set. read the letter →

arxiv 2507.04692 v2 pith:3ZFUR2FM submitted 2025-07-07 cs.CV

Wanchang Yu , Qing Zhang , Rongjia Zheng , Wei-Shi Zheng This is my paper
classification cs.CV
keywords portraitshadowremovaldiffusionmodelimageinpaintingshadow-independentstructureextractionfacerelightingdetailrestorationgenerative
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

Portrait shadow removal is hard because real paired shadow/shadow-free portraits are scarce and faces tolerate no visible artifacts. The paper tries to sidestep the data scarcity by casting shadow removal as diffusion-based inpainting: erase the shadowed region and let a generative model fill it in. The fill-in is conditioned on a shadow-independent structure map produced by a network trained on relighted portraits, keeping facial structure faithful, and a second diffusion step uses gradient information inside the shadow to restore fine details. On the PSM benchmark, the complete pipeline reports SSIM 0.830, LPIPS 0.056, and RMSE 17.162, beating eight prior methods. If correct, the work shows that high-fidelity portrait shadow removal can be achieved without real paired shadow data.

What carries the argument

The key mechanism is a three-stage pipeline built on two diffusion models and one structure extractor. SE-Net, an encoder-decoder trained with reconstruction, LPIPS, and adversarial losses, produces a shadow-independent structure map by learning to ignore illumination discontinuities. The structure-guided inpainting diffusion model, trained with the DDPM objective and sampled with DDIM, generates the shadow-free portrait conditional on the structure map and the masked input. The gradient-guided detail-restoration diffusion model, conditioned on gradient maps from the original shadow regions, recovers fine-scale details that the structure map does not encode. The synthetic data construction of Eq. 1, which blends a relighted portrait with its original using a random mask, is what makes the shadow-independent training possible.

What would settle it

Take real portraits with hard cast shadows and hand-labeled shadow boundaries, run SE-Net on them, and check whether the structure map contains edges along the boundary; if boundary edges persist, the shadow-independence claim and the downstream inpainting quality both collapse.

Watch

Extended reading notes

Core claim

The paper's central claim is that portrait shadow removal should be treated as structure-guided inpainting rather than direct shadow-to-clean regression. A shadow-independent structure extraction network (SE-Net) is trained on synthetic images formed by blending physics-based relighted portraits with random masks, so that the resulting structure map carries facial edges and geometry but excludes shadow boundaries. A DDIM-sampled inpainting diffusion model then fills the masked shadow region conditioned on this structure map and the masked input, generating a coarse shadow-free portrait. Finally, a detail-restoration diffusion model conditioned on the gradient map inside the shadow region refines the coarse result to bring back eyelashes, moles, and spots. The full method reports SSIM 0.830, LPIPS 0.056, and RMSE 17.162 on the PSM dataset, and ablation experiments attribute the gain to each of the three stages.

Load-bearing premise

The whole pipeline rests on the assumption that synthetic shadows produced by relighting well-lit portraits and blending with random masks are representative enough of real cast shadows that a network trained on them learns to ignore real shadow boundaries.

Editorial extensions

If this is right

  • If the method holds, portrait shadow removal becomes an inpainting task, so advances in diffusion inpainting can transfer directly to shadow removal.
  • The shadow-independent structure map prevents identity tampering and structure blurring, two failure modes of previous portrait shadow removal methods.
  • The gradient-guided detail restoration stage recovers fine facial details such as eyelashes, moles, and spots that structure-only conditioning misses, improving perceptual quality.
  • Because the method only needs a mask that covers the shadow region, it is robust to inaccurate masks, which matters for automatic pipelines where shadow detection is imperfect.
  • The inference time of about 9.5 seconds for a 512x512 portrait makes the approach practical for interactive use compared to slower optimization-based methods.

Reading between the lines

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

  • The same decomposition—a shadow-independent structure estimate plus generative inpainting plus detail restoration—could be applied to other illumination-corruption tasks such as specular highlight removal or deraining, where a structure map can be made invariant to the corruption.
  • Because the inpainting model is trained without requiring shadow labels, the approach could be extended to other object classes by swapping the relighting-based synthesis and the structure extractor, a direction the paper mentions only briefly for natural images.
  • A testable extension is to measure how far the input mask can deviate from the true shadow region before quality degrades; the paper's robustness claim is qualitative, so a quantitative mask-perturbation study would pin down the operating envelope.
  • The reliance on a physics-based relighting model for synthesis suggests that improvements in single-image relighting will directly improve SE-Net's shadow independence, so the method's ceiling is partly set by relighting fidelity.
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

5 major / 6 minor

Summary. The manuscript proposes a three-stage diffusion-based pipeline for removing cast shadows from portrait photos. First, a shadow-independent structure extraction network (SE-Net) is trained on synthetic relighted CelebA images (Eq. 1) to predict a structure map that contains facial details but excludes shadow boundaries. Second, a diffusion inpainting model conditioned on this structure map, the masked input, and the shadow mask generates a coarse shadow-free result. Third, a detail-restoration diffusion model, guided by gradients within the shadow regions, refines fine-scale details. The method is evaluated on the PSM real-world dataset (100 images, 9 subjects) and a self-collected dataset of 200 images; Table 1 reports state-of-the-art SSIM/LPIPS/RMSE values, with ablations supporting the proposed components.

Significance. If the reported results hold, the paper makes a useful contribution: it reframes portrait shadow removal as structure-conditioned inpainting, avoids the identity tampering common in GAN-inversion approaches, and demonstrates a plausible way to exploit relighting for synthetic training. The design is ablated carefully (structure guidance, data synthesis strategy, detail restoration), the code is promised, and the paper includes a failure case. The main limitations are the small evaluation set, the unquantified shadow-independence of the learned structure maps, and the need for a shadow mask, which currently prevent the superiority claim from being fully established.

major comments (5)
  1. [Section 3.1, Eq. (1), Fig. 7] The shadow-independence of SE-Net is not directly validated. The synthetic training data uses a random facial mask and a hard blend between the relighted and original portrait, producing artificial hard seams that differ from real cast shadows, which typically have soft penumbrae, attached-shadow geometry, and spatially varying color shifts. The only evidence for shadow-independence is downstream PSM metrics in Table 1 and a visual comparison; these do not isolate whether shadow boundaries leak into the structure map S, because the inpainting diffusion may compensate for such leakage. Please add a quantitative evaluation of S on real shadows, for example by comparing structure maps extracted from the shadowed input against structure maps of the corresponding ground-truth shadow-free image, or by measuring edge overlap between S and the shadow boundary.
  2. [Section 4, Table 1] The central claim of clearly outperforming existing methods rests on a test set of 100 images from 9 subjects, with no confidence intervals or significance tests. Given the small sample, the reported metric gaps may not be statistically robust. Please report per-subject or per-image variance, perform significance tests (e.g., paired tests on SSIM/LPIPS/RMSE), and, if possible, evaluate on a larger or additional benchmark.
  3. [Section 4.1, Table 1] The PSM baseline numbers are quoted from the original paper rather than recomputed by the authors, as acknowledged in the text. Because the original evaluation may use different mask definitions, metric computations, or test protocols, the comparison is not strictly apples-to-apples. Please re-run PSM if its released data/code permit, or document and justify the protocol differences.
  4. [Section 3.2, Eqs. (4)-(6)] The training specification of the inpainting diffusion model is ambiguous. The text says the target image x0 is "namely the masked input," but the surrounding description says the model reconstructs the original unmasked portrait. If x0 is the masked input, the model only denoises holes and does not perform inpainting; if x0 is the clean target, the conditioning on IM and M and the handling of known pixels during DDIM sampling need to be specified. Please clarify the definition of x0, the construction of xt, and the exact sampling procedure that enforces agreement with the known regions.
  5. [Section 4 and Figure 2] The method is given a shadow mask (from the detector of [6] and subsequently refined), while several compared baselines are fully automatic. Providing the mask gives the proposed method additional information that the baselines do not receive, making the comparison asymmetric. Please either evaluate all methods under the same mask/protocol where possible, or ablate the effect of the mask and discuss this asymmetry explicitly.
minor comments (6)
  1. [Section 3.1, Eq. (3)] The hyperparameter sentence reads "we empirically set λ1 = 0.5, λ1 = 0.25"; the second symbol should presumably be λ2. Please correct this typo.
  2. [Section 1, Contributions] The paper describes the approach as "unsupervised," but training uses synthetic paired data and pseudo-ground-truth from the PDG model; a term such as "supervised with synthetic data" or "without real paired shadow data" would be more accurate.
  3. [Section 3.4 and Table 2] The inference time is reported as about 2.5 seconds on an RTX 4090 in Section 3.4 and 9.5 seconds on an A100 in Table 2; please clarify the hardware and settings used for each measurement.
  4. [Figure 2 and Section 3.3] The extraction and use of the gradient map for detail restoration are described only briefly; please specify how the gradient map is computed from the difference image and how it conditions the detail-restoration diffusion model.
  5. [Section 3.4] The paper states that the diffusion models follow the same U-Net architecture as in [15], but does not specify how the structure map, mask, and gradient map are injected into the network; adding these details would improve reproducibility.
  6. [Section 4.3] Robustness to inaccurate masks is demonstrated visually on a single example; a quantitative study with varying mask perturbations would better support the robustness claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is a supervised empirical pipeline trained with external structure supervision and evaluated on held-out real-world data.

full rationale

The paper's claimed derivation chain is an empirical learning pipeline rather than a formal derivation, and no step reduces to its own inputs by construction. SE-Net is trained to map synthetically relighted portraits to structure maps produced by an external PDG model [46] on the corresponding well-lit portraits; the training target G_p(I) is independent of the final shadow-removal output and of the PSM benchmark ground truth. The structure-guided inpainting diffusion model is trained on randomly masked shadow-free FFHQ images, and the detail-restoration diffusion model is trained on the same synthetic relighting data, again without using the test-set labels. Quantitative claims are made against the external PSM real-world paired dataset and against published baseline models, so the headline Table 1 results are not forced by any fitted parameter relabeled as a prediction. The paper does invoke the PDG model as pseudo-ground-truth, but this is an external pretrained model, not a self-citation, and the ablation 'Ours w/ PDG predicted structure' explicitly shows that using PDG structure directly degrades performance, confirming that the learned SE-Net contributes non-trivially. Concerns about the synthetic relighting data (random masks, hard seams, domain gap to real cast shadows) are empirical validity or robustness concerns, not circularity: they do not make any equation equal to its own input. No load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling citations are present. The paper is self-contained against external benchmarks, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central pipeline introduces no new physical entities. The free parameters are common loss weights. The main assumptions concern the realism of synthetic relighting and the transferability of the pretrained structure extractor and inpainting model to real shadow regions.

free parameters (2)
  • lambda1 = 0.5
    Weight for perceptual loss in SE-Net training (Eq. 3), selected empirically for good visual results.
  • lambda2 = 0.25
    Weight for GAN loss in SE-Net training (Eq. 3), selected empirically.
assumptions (3)
  • domain assumption The physics-based relighting method [54] produces realistic illumination changes on well-lit portraits, so the blended synthetic images in Eq. (1) mimic real cast shadows.
    Used to synthesize training data for SE-Net; unrealistic relighting would weaken the shadow-independence of the structure map.
  • domain assumption The PDG model [46] output on well-lit portraits is a valid shadow-free pseudo-ground-truth for structure maps.
    SE-Net is supervised by PDG output; any systematic error in PDG is inherited by SE-Net.
  • domain assumption A diffusion model trained to inpaint random masks on FFHQ portraits transfers to shadow-region inpainting at test time.
    The method assumes that random-mask inpainting learns to reconstruct facial structure in a way that also removes shadows.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal." pith.science (2026). https://pith.science/paper/3ZFUR2FM

@misc{pith2026250704692,
  author       = {Pith},
  title        = {Pith review of: Structure-Guided Diffusion Models for High-Fidelity Portrait Shadow Removal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZFUR2FM}},
  note         = {Machine review of arXiv:2507.04692}
}
read the original abstract

We present a diffusion-based portrait shadow removal approach that can robustly produce high-fidelity results. Unlike previous methods, we cast shadow removal as diffusion-based inpainting. To this end, we first train a shadow-independent structure extraction network on a real-world portrait dataset with various synthetic lighting conditions, which allows to generate a shadow-independent structure map including facial details while excluding the unwanted shadow boundaries. The structure map is then used as condition to train a structure-guided inpainting diffusion model for removing shadows in a generative manner. Finally, to restore the fine-scale details (e.g., eyelashes, moles and spots) that may not be captured by the structure map, we take the gradients inside the shadow regions as guidance and train a detail restoration diffusion model to refine the shadow removal result. Extensive experiments on the benchmark datasets show that our method clearly outperforms existing methods, and is effective to avoid previously common issues such as facial identity tampering, shadow residual, color distortion, structure blurring, and loss of details. Our code is available at https://github.com/wanchang-yu/Structure-Guided-Diffusion-for-Portrait-Shadow-Removal.

Figures

Figures reproduced from arXiv: 2507.04692 by the authors.

Figure 1
Figure 1. Comparison on portrait shadow removal. Our method allows to robustly remove facial cast shadows while faithfully preserving the portrait identities as well as structure details in shadow regions and also ensuring the overall naturalness. Abstract We present a diffusion-based portrait shadow removal ap￾proach that can robustly produce high-fidelity results. Un￾like previous methods, we cast shadow removal as diffusio… view at source ↗
Figure 2
Figure 2. Inference pipeline of our method. Given a input portrait image and the corresponding shadow mask, we first feed the input image to the shadow-independent structure extraction network (SE-Net) to get a structure map. Then, we perform shadow removal by inpainting the shadow regions with a diffusion model conditioned on the structure map. Finally, the shadow removal output is refined by a detail restoration diffusion m… view at source ↗
Figure 3
Figure 3. Training pipeline of SE-Net. Note, we adopt a trained PDG model for structure extraction from [46], while it is not in￾variant to illumination changes. 3.1. Shadow-Independent Structure Extraction As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Training pipeline of our diffusion models. Here we introduce how our structure-guided portrait inpainting diffusion model and gradient-guided detail restoration diffusion model are trained. 3.2. Structure-Guided Portrait Inpainting As shown in [PITH_FULL_IMAGE:figures…
Figure 5
Figure 5. Figure 5: Visual comparison with state-of-the-art methods on test images from the dataset provided in [53]. Method All Shadow region Non-shadow region SSIM↑ LPIPS↓ RMSE↓ SSIM↑ LPIPS↓ RMSE↓ SSIM↑ LPIPS↓ RMSE↓ ShadowDiffusion [11] 0.650 0.177 38.246 0.901 0.037 15.561 0.579 0.223 …
Figure 6
Figure 6. Figure 6: Visual comparison with state-of-the-art methods on test images from our collected dataset. nential Moving Average (EMA) in all our experiments. We use 2000 diffusion steps and a noise schedule βt linearly in￾creasing from 1 × 10−6 to 1 × 10−2 for training. During infer…
Figure 7
Figure 7. Figure 7: Visual comparison of structure maps extracted by different methods and the corresponding shadow removal results. (a) Input. (b)-(d) are structure maps generated by PDG [46], SE-Net trained on data synthesized by the method of PSM [53], and SE-Net trained on our synthes…
Figure 8
Figure 8. Figure 8: examine how structure map derived from the SE￾Net affects the performance of our diffusion-based shadow inpainting. As shown, omitting the structure map, the fa￾cial structure details behind shadows are inevitably modi￾fied, producing unpredictable portrait inconsisten…
Figure 9
Figure 9. Figure 9: Effect of our detail restoration [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Effect of inaccurate shadow masks. Our method is highly robust to inaccurate masks (see 3rd and 4th columns). Input SD + ControlNet Ours GT [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Effect of leveraging different diffusion models for shadow inpainting. Note, the same structure map as ours is used to produce the shadow inpainting result of “SD + ControlNet”, and our result here is free of detail restoration. Input UPSR BSR Ours [PITH_FULL_IMAGE:f…
Figure 12
Figure 12. Figure 12: Failure case. Our method may produce unnatural re￾sult for complex self-attached shadows (see the eyelid area). Method UPSR BSR ShadowDiffusion IC-light Ours Param. 135M 72M 108.79M - 185.12M Time 192s 11s 2.3s 32s 9.5s [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 48 canonical work pages

  1. [6]

    Robust shadow detection by exploring effective shadow contexts

    Xianyong Fang, Xiaohao He, Linbo Wang, and Jianbing Shen. Robust shadow detection by exploring effective shadow contexts. InACM MM, pages 2927–2935, 2021. 4

    work page 2021

  2. [1]

    Latentpaint: Image inpainting in latent space with diffusion models

    Ciprian Corneanu, Raghudeep Gadde, and Aleix M Mar- tinez. Latentpaint: Image inpainting in latent space with diffusion models. InWACV, pages 4334–4343, 2024. 2

    2024

  3. [2]

    Re- gion filling and object removal by exemplar-based image in- painting.IEEE Transactions on Image Processing, 13(9): 1200–1212, 2004

    Antonio Criminisi, Patrick P ´erez, and Kentaro Toyama. Re- gion filling and object removal by exemplar-based image in- painting.IEEE Transactions on Image Processing, 13(9): 1200–1212, 2004. 2

    work page 2004

  4. [3]

    Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan

    Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. InAAAI, pages 10680– 10687, 2020. 2

    work page 2020

  5. [4]

    Deep rank hashing net- work for cancellable face identification.Pattern Recognition, 131:108886, 2022

    Xingbo Dong, Sangrae Cho, Youngsam Kim, Soohyung Kim, and Andrew Beng Jin Teoh. Deep rank hashing net- work for cancellable face identification.Pattern Recognition, 131:108886, 2022. 1

    work page 2022

  6. [5]

    The watchlist imbalance effect in biometric face identifica- tion: Comparing theoretical estimates and empiric measure- ments

    Pawel Drozdowski, Christian Rathgeb, and Christoph Busch. The watchlist imbalance effect in biometric face identifica- tion: Comparing theoretical estimates and empiric measure- ments. InICCV, pages 3757–3765, 2021. 1

    work page 2021

  7. [7]

    Entropy minimization for shadow removal.International Journal of Computer Vision, 85(1):35–57, 2009

    Graham D Finlayson, Mark S Drew, and Cheng Lu. Entropy minimization for shadow removal.International Journal of Computer Vision, 85(1):35–57, 2009. 2

    work page 2009

  8. [8]

    Controllable light diffusion for portraits

    David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escolano, Brian Curless, Daniel S`ykora, and Ro- hit Pandey. Controllable light diffusion for portraits. In CVPR, pages 8412–8421, 2023. 2, 6

    work page 2023

Show all 57 references
  1. [9]

    Learn- ing to remove soft shadows.ACM Transactions on Graphics (TOG), 34(5):1–15, 2015

    Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learn- ing to remove soft shadows.ACM Transactions on Graphics (TOG), 34(5):1–15, 2015. 2

    2015

  2. [10]

    Shadowformer: Global context helps shadow removal

    Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. InAAAI, pages 710–718, 2023. 2, 5, 6

    2023

  3. [11]

    Shadowd- iffusion: When degradation prior meets diffusion model for shadow removal

    Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowd- iffusion: When degradation prior meets diffusion model for shadow removal. InCVPR, pages 14049–14058, 2023. 1, 2, 5, 6

    2023

  4. [12]

    Paired regions for shadow detection and removal.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 35(12):2956–2967,

    Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 35(12):2956–2967,

  5. [13]

    Delving deep into rectifiers: Surpassing human-level perfor- mance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level perfor- mance on imagenet classification. InICCV, pages 1026– 1034, 2015. 5

    2015

  6. [14]

    Unsupervised portrait shadow removal via generative priors

    Yingqing He, Yazhou Xing, Tianjia Zhang, and Qifeng Chen. Unsupervised portrait shadow removal via generative priors. InACM MM, pages 236–244, 2021. 2, 5, 6

    2021

  7. [15]

    Denoising diffu- sion probabilistic models.NeurIPS, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.NeurIPS, 33:6840–6851, 2020. 4, 5

    2020

  8. [16]

    Compose: Comprehensive portrait shadow editing

    Andrew Hou, Zhixin Shu, Xuaner Zhang, He Zhang, Yan- nick Hold-Geoffroy, Jae Shin Yoon, and Xiaoming Liu. Compose: Comprehensive portrait shadow editing. In ECCV, pages 356–373. Springer, 2024. 2, 6

    2024

  9. [17]

    Mask-shadowgan: Learning to remove shadows from unpaired data

    Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. InICCV, pages 2472–2481, 2019. 2

    2019

  10. [18]

    Globally and locally consistent image completion.ACM Transactions on Graphics (ToG), 36(4):1–14, 2017

    Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion.ACM Transactions on Graphics (ToG), 36(4):1–14, 2017. 2

    2017

  11. [19]

    A diffusion model with state estimation for degradation-blind inverse imaging

    Liya Ji, Zhefan Rao, Sinno Jialin Pan, Chenyang Lei, and Qifeng Chen. A diffusion model with state estimation for degradation-blind inverse imaging. InAAAI, pages 2471– 2479, 2024. 2

    2024

  12. [20]

    Talk-to-edit: Fine-grained facial editing via dialog

    Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. InICCV, pages 13799–13808, 2021. 1

    2021

  13. [21]

    Dc- shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network

    Yeying Jin, Aashish Sharma, and Robby T Tan. Dc- shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In ICCV, pages 5027–5036, 2021. 2

    2021

  14. [22]

    Des3: Adaptive attention-driven self and soft shadow removal using vit similarity

    Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. InAAAI, pages 2634–2642,

  15. [23]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. InCVPR, pages 8110–8119,

  16. [24]

    Shadow removal via shadow image decomposition

    Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. InICCV, pages 8578–8587, 2019. 2

    2019

  17. [25]

    Leveraging inpainting for single-image shadow removal

    Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. InICCV, pages 13055– 13064, 2023. 2, 5, 6

    2023

  18. [26]

    Blind removal of facial foreign shadows

    Yaojie Liu, Andrew Z Hou, Xinyu Huang, Liu Ren, and Xi- aoming Liu. Blind removal of facial foreign shadows. In BMVC, page 88, 2022. 1, 2, 5, 6

    2022

  19. [27]

    Recasting regional lighting for shadow removal

    Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. InAAAI, pages 3810–3818, 2024. 2

    2024

  20. [28]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. InICCV, pages 3730–3738, 2015. 3, 5

    2015

  21. [29]

    From shadow generation to shadow removal

    Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927–4936, 2021. 2

    2021

  22. [30]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InCVPR, pages 11461–11471, 2022. 2

    2022

  23. [31]

    Portrait eyeglasses and shadow removal by leveraging 3d synthetic data

    Junfeng Lyu, Zhibo Wang, and Feng Xu. Portrait eyeglasses and shadow removal by leveraging 3d synthetic data. In CVPR, pages 3429–3439, 2022. 2

    2022

  24. [32]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. InCVPR, pages 2536–2544, 2016. 2

    2016

  25. [33]

    Deshadownet: A multi-context embedding deep network for shadow removal

    Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. InCVPR, pages 4067–4075, 2017. 2

    2017

  26. [34]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, pages 10684– 10695, 2022. 8

    2022

  27. [35]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1–10,

    2022

  28. [36]

    Image super-resolution via iterative refinement.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713– 4726, 2022

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Sal- imans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713– 4726, 2022. 5

    2022

  29. [37]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 4

    2010 arXiv

  30. [38]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. InWACV, pages 2149–2159, 2022. 2

    2022

  31. [39]

    Representative forgery mining for fake face detection

    Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. InCVPR, pages 14923– 14932, 2021. 1

    2021

  32. [40]

    Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal

    Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. InCVPR, pages 1788–1797,

  33. [41]

    Deep portrait delighting

    Joshua Weir, Junhong Zhao, Andrew Chalmers, and Taehyun Rhee. Deep portrait delighting. InECCV, pages 423–439. Springer, 2022. 2, 6

    2022

  34. [42]

    Space- time completion of video.IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3):463–476, 2007

    Yonatan Wexler, Eli Shechtman, and Michal Irani. Space- time completion of video.IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(3):463–476, 2007. 2

    2007

  35. [43]

    Homoformer: Homogenized trans- former for image shadow removal

    Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized trans- former for image shadow removal. InCVPR, pages 25617– 25626, 2024. 2, 5, 6

    2024

  36. [44]

    Smartbrush: Text and shape guided object inpainting with diffusion model

    Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. InCVPR, pages 22428–22437, 2023. 2

    2023

  37. [45]

    Transed- itor: Transformer-based dual-space gan for highly control- lable facial editing

    Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transed- itor: Transformer-based dual-space gan for highly control- lable facial editing. InCVPR, pages 7683–7692, 2022. 1

    2022

  38. [46]

    Unpaired portrait drawing generation via asymmetric cycle mapping

    Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. InCVPR, pages 8217–8225, 2020. 3, 7, 8

    2020

  39. [47]

    Generative portrait shadow removal.ACM Trans- actions on Graphics (TOG), 43(6):1–13, 2024

    Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal.ACM Trans- actions on Graphics (TOG), 43(6):1–13, 2024. 2, 6

    2024

  40. [48]

    Generative image inpainting with contex- tual attention

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contex- tual attention. InCVPR, pages 5505–5514, 2018. 2

    2018

  41. [49]

    Scaling in-the-wild training for diffusion-based illumination harmo- nization and editing by imposing consistent light transport

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmo- nization and editing by imposing consistent light transport. InICLR. 1, 2, 5, 6

  42. [50]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836–3847, 2023. 8

    2023

  43. [51]

    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. InCVPR, pages 586–595,

  44. [52]

    Re- fineface: Refinement neural network for high performance face detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4008–4020, 2020

    Shifeng Zhang, Cheng Chi, Zhen Lei, and Stan Z Li. Re- fineface: Refinement neural network for high performance face detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4008–4020, 2020. 1

    2020

  45. [53]

    Por- trait shadow manipulation.ACM Transactions on Graphics (TOG), 39(4):78–1, 2020

    Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Por- trait shadow manipulation.ACM Transactions on Graphics (TOG), 39(4):78–1, 2020. 2, 3, 5, 6, 7

    2020

  46. [54]

    Deep single-image portrait relighting

    Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Ja- cobs. Deep single-image portrait relighting. InICCV, pages 7194–7202, 2019. 3

    2019

  47. [55]

    Text image inpainting via global structure-guided diffusion models

    Shipeng Zhu, Pengfei Fang, Chenjie Zhu, Zuoyan Zhao, Qiang Xu, and Hui Xue. Text image inpainting via global structure-guided diffusion models. InAAAI, pages 7775– 7783, 2024. 2

    2024

  48. [56]

    Bijective mapping network for shadow removal

    Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. InCVPR, pages 5627–5636, 2022. 2

    2022

  49. [57]

    Efficient model-driven network for shadow removal

    Yurui Zhu, Zeyu Xiao, Yanchi Fang, Xueyang Fu, Zhiwei Xiong, and Zheng-Jun Zha. Efficient model-driven network for shadow removal. InAAAI, pages 3635–3643, 2022. 2

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.