Pith. sign in

REVIEW 4 major objections 5 minor 43 references

MoFRR: Mixture of Diffusion Models for Face Retouching Restoration

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read MoFRR undoes face retouching, beating baselines by up to 6.44 dB in PSNR.

desk verdict A useful new task and architecture, but the headline PSNR gains may be inflated by train/test identity leakage given an image-level split over retouched variants of the same FFHQ originals. read the letter →

arxiv 2507.19770 v1 pith:T4OC5BX4 submitted 2025-07-26 cs.CV

classification cs.CV
keywords faceretouchingrestorationmixtureofexpertsdiffusionmodelswavelettransformimageforensicsFFHQ++low-frequency
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

Face retouching restoration (FRR) is introduced as a new task: given a retouched face image, reconstruct the original face without any reference. The paper argues that existing makeup-removal and image-restoration methods are ill-suited because retouching changes facial structure, not just texture, so restoration must focus on low-frequency content. MoFRR is proposed as a mixture-of-experts framework in which each retouching operation (whitening, smoothing, face lifting, eye enlargement) has its own diffusion-based expert and a shared expert handles universal traces. On the new RetouchingFFHQ++ dataset, MoFRR outperforms all baselines, with particularly large margins on images that have undergone multiple retouching operations.

What carries the argument

The load-bearing machinery is MoFRR, a mixture-of-experts model with four specialized WaveFRR experts (one per retouching type), a shared DDIM-based expert that captures universal retouching traces, a router (ResNet-MAM) that predicts which operations were applied, and a lightweight Combine Module (U-Net) that merges the expert outputs. Each WaveFRR expert decomposes the input via discrete wavelet transform, restores the low-frequency sub-band with a conditional DDIM guided by a degree estimator and the Iterative Distortion Evaluation Module (IDEM), refines the high-frequency sub-bands with a Cross-Attention High-Frequency module (HFCAM), and recombines via the inverse transform. The key identity is that retouching mainly modifies low-frequency content, so the diffusion process works on a half-resolution LL sub-band conditioned on a pixel-wise distortion map derived from the predicted degree.

What would settle it

Generate a test set where two or more retouching operations are applied by a novel engine in a deliberately non-additive order (e.g., smooth after eye enlargement, or lift after whitening) and measure whether MoFRR's composed experts still restore PSNR close to the multi-operation baseline. If the per-expert outputs fail to compose and the Combine Module cannot compensate, the independence premise is falsified.

Watch

Extended reading notes

Core claim

The central claim is that MoFRR accurately restores original faces from retouched ones, and that it outperforms existing methods by a substantial margin on multi-operation retouching: it surpasses the second-best method by 5.49 dB PSNR and achieves the highest improvement of 6.44 dB on the multi-operated subset. The paper further claims that decomposition into operation-specific experts plus a shared expert is what enables this gain, and that the dual-branch wavelet design—DDIM-based low-frequency restoration guided by a distortion estimate and cross-attention high-frequency refinement—is effective for single-operation restoration. As a direct corollary, the paper asserts that FRR should be evaluated not only by pixel fidelity but by feature-space identity similarity, where MoFRR also shows a sharp peak near perfect alignment.

Load-bearing premise

The load-bearing premise is that the impact of different face retouching operations is independent in the face images, so a router can activate a set of single-operation experts whose outputs compose into a correct multi-operation restoration.

Editorial extensions

If this is right

  • If MoFRR works as claimed, retouched images on social platforms can be automatically reverted to their pre-edit state, giving regulators and platforms a tool to enforce disclosure rules.
  • Because the router only selects experts, new retouching types can be handled by plugging in an additional expert without retraining the whole model, a cost saving the paper notes.
  • The focus on low-frequency restoration suggests that FRR is fundamentally different from super-resolution or deblurring, and that frequency decomposition should be a default component of future FRR systems.
  • The evaluation protocol—combining PSNR/SSIM with ArcFace and AdaFace cosine similarity—sets a template for forensic-grade restoration benchmarks.

Reading between the lines

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

  • The independence assumption behind the router composition is only tested on API-generated mixtures from the same pipelines; a stress test on adversarial combinations (e.g., smoothing applied after eye enlargement in a way that changes geometry) would clarify whether the experts truly compose or the Combine Module is absorbing the interaction.
  • The degree redefinition via PSNR distribution is data-driven but API-specific; a cross-API calibration study could test whether degree labels transfer to unseen retouching engines.
  • If FRR reaches the accuracy claimed, it also raises a privacy tension: the same restoration that exposes deceptive advertising can be used to strip consented stylistic edits from portraits, so deployment norms may need to distinguish retouching-as-deception from retouching-as-expression.
  • The 'shared expert' inspired by DeepSeek's expert isolation suggests that sparse activation in mixtures of experts can be repurposed from language modeling to low-level vision, potentially opening a general design pattern for restoration tasks with multiple degradation types.
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

4 major / 5 minor

Summary. The paper introduces Face Retouching Restoration (FRR), a task of recovering original faces from retouched images, and proposes MoFRR, a mixture-of-experts framework with four operation-specific WaveFRR experts (whitening, smoothing, face lifting, eye enlarging), a shared DDIM-based expert, a router, and a Combine Module. It also presents RetouchingFFHQ++, an extension of RetouchingFFHQ with over one million retouched images from four commercial retouching APIs and PSNR-based degree labels. Experiments report substantial PSNR/SSIM gains over retrained baselines, especially 34.47 dB on multi-operation images, claimed to be 5.49 dB above the second-best method, with ablations and cross-API tests supporting the design.

Significance. If the empirical claims hold, MoFRR would be a useful first system for face retouching restoration, and RetouchingFFHQ++ is a substantial new resource for the community. The design is thoughtful: sparse activation of specialized experts, a shared expert inspired by DeepSeek, a wavelet-domain DDIM with IDEM and HFCAM, and a dual evaluation protocol combining PSNR/SSIM with biometric feature similarity. The ablation study and cross-API experiments are systematic in structure, and the authors are careful to retrain all baselines on their dataset. The main caveats are experimental validity: the reported split appears to allow identity leakage between training and test, no statistical significance or variance is reported, and the operation-independence premise is not directly tested. These issues affect the central quantitative claims and require a major revision.

major comments (4)
  1. [§3.1, §4.1, Table 2] The evaluation protocol does not rule out identity leakage. RetouchingFFHQ++ is generated from 57,910 FFHQ originals by applying 0–4 retouching operations via four APIs, yielding over 1.07 million retouched images (Table 1), so each original has many retouched versions. Section 4.1 states only that the dataset is divided 8:1:1 as training/evaluation/test, with no identity-disjoint split. Under a random image-level split the same original face almost certainly appears in both training and test, letting the model memorize the exact target and inflate PSNR/SSIM; this directly affects the headline Multi gain of 5.49 dB in Table 2. The cross-API test in Table 3 has the same issue because all API subsets share the same FFHQ originals. The authors should re-split by original identity and report all tables under that protocol, or provide evidence that the results are unchanged.
  2. [§4.2, Tables 2–4] No error bars, significance tests, or test-set sizes are reported for any quantitative comparison. The claimed superiority rests on point estimates whose spread is unknown; for example, the per-operation improvements in Table 2 range from about 0.5 to 4.0 dB, and the Multi gain is a single number. I ask for mean ± std over at least three training runs or bootstrap confidence intervals, together with the number of evaluation images per column. Without this, it is difficult to assess whether the reported improvements are stable or within noise.
  3. [§3.4, Table 4] The ablation narrative is contradicted by the table in the Multi column. The text says removing any one module degrades performance compared to the full model, but Table 4 shows w/o IDEM (32.94), w/o Degree (32.95), and w/o HFCAM (33.04) all outperform the full WaveFRR (31.09) on Multi before the Combine Module is applied. This suggests either the full WaveFRR was not trained to convergence for multi-operation inputs, or the ablation procedure differs from what the text describes. The claim as written is not supported by the table; please clarify, and if the Combine Module is what produces the final gain, state that explicitly.
  4. [§1, §3.4] The central modeling premise that 'the impact of different face retouching operations is independent in the face images' is not directly tested. Specialized experts are trained only on single-operation images, and multi-operation inputs are left to the router and Combine Module. Although the Combine Module is trained on multi-operation data and Table 4 shows that removing the router costs only 0.28 dB on Multi, the manuscript should validate composition on independently generated mixtures, for example by applying operations from different APIs to held-out originals or testing unseen operation combinations, and should report router precision/recall. This would separate the validity of the independence assumption from the learned Combine Module's capacity to compensate for nonlinear interactions.
minor comments (5)
  1. [§4.3] The claimed cross-API improvements of 1.2 dB and 3.87 dB do not match Table 3, where MoFRR exceeds the best baseline by 1.72 dB on Single and 4.49 dB on Multi; please correct the numbers or clarify which baseline is used for comparison.
  2. [§3.4] The shared expert is said to be trained on 'subset B', but subset B is never defined; please specify its composition, size, and how it relates to the overall RetouchingFFHQ++ split.
  3. [§3.1] The degree re-categorization by PSNR distribution with proportions 15/25/25/25/10 is presented without validation; please provide evidence that these labels correlate with perceptual or API-provided degrees, or discuss the limitation of this heuristic.
  4. [§3.4] The classification loss is written as a five-class cross-entropy, but the router is a multi-label binary classifier; please state the exact loss used for the router and the degree estimator separately.
  5. [References] Reference [1] is cited as 'Anthropic' for PortraitPro 24; the company name is Anthropics, not Anthropic, and the URL should be checked.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MoFRR is trained and evaluated against external FFHQ ground truth; no prediction reduces to a fitted input or self-citation chain.

full rationale

MoFRR's claimed derivation chain is empirical rather than analytic. The router, degree estimators, WaveFRR experts, shared expert, and Combine Module are all trained with losses defined against ground-truth FFHQ originals or ground-truth operation/degree labels; the final restored image is compared to the ground truth via L_hyb in Section 3.4. No equation defines an output in terms of the quantity it is claimed to predict. The degree labels are redefined from the PSNR distribution of the dataset, but the degree is only an auxiliary conditioning variable, and the degree estimator is trained and evaluated on the same labeling rule; this is a dataset-construction choice, not a circular reduction. The dataset extends the authors' prior RetouchingFFHQ, but the underlying originals are the external FFHQ dataset and all baselines are retrained on the same RetouchingFFHQ++ splits, so the self-citation does not alone force the reported improvements. The train/test identity-overlap concern raised by the skeptic is a potential benchmarking-validity risk, not a circularity: even if test identities overlap training, the model output is still compared with the held-out retouched-to-original pairs and does not equal the input by construction. No circular step can be quoted from the paper, so the score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities or theoretical objects are introduced; IDEM and HFCAM are module designs rather than entities. The free parameters are hand-set architecture and loss choices rather than physical constants. The central empirical claim does not depend on a closed-form derivation, so the circularity burden is mainly about dataset self-reference and degree labeling.

free parameters (4)
  • Degree group proportions for re-labeling = 15%, 25%, 25%, 25%, 10% of ascending PSNR
    Section 3.1 redefines retouching degree by binning PSNR of retouched images into five groups; the proportions determine the labels that supervise each degree estimator.
  • HFCAM loss weights lambda_1 and lambda_2 = 0.1 and 0.01
    Set by hand in Section 3.4 for the high-frequency loss; no sensitivity analysis is given.
  • Router activation threshold = 0.5
    Default threshold in Eq. (1) decides whether an expert is activated; no tuning or robustness test is reported.
  • Number of retouching types N = 4 (whitening, smoothing, face lifting, eye enlarging)
    Section 3.2 fixes the number of specialized experts and the router output dimension to four typical operations.
assumptions (5)
  • domain assumption Retouching operations act independently on face images.
    Section 1 states this to justify one expert per operation; Section 3.4 trains each specialized expert only on single-operated images, so multi-operation restoration relies on this independence.
  • standard math DDIM/DDPM Markov chain with learned noise predictor models the retouching removal distribution.
    The low-frequency branch follows the standard DDPM/DDIM framework from [18, 34] without modification; correctness depends on this established model class.
  • standard math DWT provides an invertible decomposition into low and high frequency sub-bands sufficient for FRR.
    Section 3.3 decomposes X into {x_LL, x_H} and reconstructs via IDWT, assuming frequency decomposition captures semantically relevant retouching effects.
  • domain assumption RetouchingFFHQ++ generated with commercial APIs is representative of real-world retouching.
    The dataset in Section 3.1 is constructed from FFHQ plus four APIs; cross-API generalization claims assume these APIs cover real retouching distributions.
  • ad hoc to paper PSNR-based degree re-categorization yields meaningful supervision for degree estimation.
    Section 3.1 redefines degrees by PSNR distribution rather than by actual API strength, an arbitrary labeling choice that supervises the degree estimators.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoFRR: Mixture of Diffusion Models for Face Retouching Restoration." pith.science (2026). https://pith.science/paper/T4OC5BX4

@misc{pith2026250719770,
  author       = {Pith},
  title        = {Pith review of: MoFRR: Mixture of Diffusion Models for Face Retouching Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4OC5BX4}},
  note         = {Machine review of arXiv:2507.19770}
}
read the original abstract

The widespread use of face retouching on social media platforms raises concerns about the authenticity of face images. While existing methods focus on detecting face retouching, how to accurately recover the original faces from the retouched ones has yet to be answered. This paper introduces Face Retouching Restoration (FRR), a novel computer vision task aimed at restoring original faces from their retouched counterparts. FRR differs from traditional image restoration tasks by addressing the complex retouching operations with various types and degrees, which focuses more on the restoration of the low-frequency information of the faces. To tackle this challenge, we propose MoFRR, Mixture of Diffusion Models for FRR. Inspired by DeepSeek's expert isolation strategy, the MoFRR uses sparse activation of specialized experts handling distinct retouching types and the engagement of a shared expert dealing with universal retouching traces. Each specialized expert follows a dual-branch structure with a DDIM-based low-frequency branch guided by an Iterative Distortion Evaluation Module (IDEM) and a Cross-Attention-based High-Frequency branch (HFCAM) for detail refinement. Extensive experiments on a newly constructed face retouching dataset, RetouchingFFHQ++, demonstrate the effectiveness of MoFRR for FRR.

Figures

Figures reproduced from arXiv: 2507.19770 by the authors.

Figure 1
Figure 1. Application scenario of the proposed scheme (MoFRR). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Examples from the RetouchingFFHQ++ dataset, where [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Left: Overview of MoFRR, where input image is processed by the router and selectively sent into the specialized experts for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of the restored faces using different methods. The images are retouched through multiple operations, and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of images generated by five specific ex [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Probability density distribution of facial feature similar [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of the ablation studies, which [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 31 canonical work pages

  1. [23]

    Pasd: A performance analysis approach through the statistical debugging of kernel events

    Mohammed Adib Khan, Morteza Noferesti, and Naser Ezzati-Jivan. Pasd: A performance analysis approach through the statistical debugging of kernel events. In2023 IEEE 23rd International Working Conference on Source Code Analysis and Manipulation (SCAM), pages 151–161. IEEE, 2023. 1

  2. [1]

    Portraitpro 24.https : / / www

    Anthropic. Portraitpro 24.https : / / www . anthropics.com/portraitpro/, 2024. 3

  3. [2]

    Khadijah Ateq, Mohammed Alhajji, and Noara Alhusseini. The association between use of social media and the devel- opment of body dysmorphic disorder and attitudes toward cosmetic surgeries: a national survey.Frontiers in Public Health, 12:1324092, 2024. 1

  4. [3]

    Detecting facial retouching using supervised deep learning.IEEE Transactions on Information Forensics and Security, 11(9):1903–1913, 2016

    Aparna Bharati, Richa Singh, Mayank Vatsa, and Kevin W Bowyer. Detecting facial retouching using supervised deep learning.IEEE Transactions on Information Forensics and Security, 11(9):1903–1913, 2016. 1

  5. [4]

    A survey on mixture of experts.arXiv preprint arXiv:2407.06204, 2024

    Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts.arXiv preprint arXiv:2407.06204, 2024. 2

  6. [5]

    An aug- mented lagrangian method for total variation video restora- tion.IEEE Transactions on Image Processing, 20(11):3097– 3111, 2011

    Stanley H Chan, Ramsin Khoshabeh, Kristofor B Gib- son, Philip E Gill, and Truong Q Nguyen. An aug- mented lagrangian method for total variation video restora- tion.IEEE Transactions on Image Processing, 20(11):3097– 3111, 2011. 5

  7. [6]

    Pairedcyclegan: Asymmetric style transfer for apply- ing and removing makeup

    Huiwen Chang, Jingwan Lu, Fisher Yu, and Adam Finkel- stein. Pairedcyclegan: Asymmetric style transfer for apply- ing and removing makeup. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 40–48, 2018. 2, 5, 6, 7

  8. [7]

    Auto- matic facial makeup detection with application in face recog- nition

    Cunjian Chen, Antitza Dantcheva, and Arun Ross. Auto- matic facial makeup detection with application in face recog- nition. In2013 international conference on biometrics (ICB), pages 1–8. IEEE, 2013. 1

Show all 43 references
  1. [8]

    Makeup- go: Blind reversion of portrait edit

    Ying-Cong Chen, Xiaoyong Shen, and Jiaya Jia. Makeup- go: Blind reversion of portrait edit. InProceedings of the IEEE International Conference on Computer Vision, pages 4501–4509, 2017. 2

  2. [9]

    Truth in advertising act of 2014, 2014.https://www.congress.gov/bill/ 113th-congress/house-bill/4341

    United States Congress. Truth in advertising act of 2014, 2014.https://www.congress.gov/bill/ 113th-congress/house-bill/4341. 1

  3. [10]

    Attentional feature fusion

    Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. InProceed- ings of the IEEE/CVF winter conference on applications of computer vision, pages 3560–3569, 2021. 4

  4. [11]

    Antitza Dantcheva, Cunjian Chen, and Arun Ross. Can facial cosmetics affect the matching accuracy of face recognition systems? In2012 IEEE Fifth international conference on biometrics: theory, applications and systems (BTAS), pages 391–398. IEEE, 2012. 1

  5. [12]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 4690–4699, 2019. 3, 8

  6. [13]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint ar...

  7. [14]

    The marketing control act.https: //www.forbrukertilsynet.no/english/the- marketing-control-act, 2021

    Forbrukertilsynet. The marketing control act.https: //www.forbrukertilsynet.no/english/the- marketing-control-act, 2021. 1

  8. [15]

    Ladn: Local adversarial disentangling net- work for facial makeup and de-makeup

    Qiao Gu, Guanzhi Wang, Mang Tik Chiu, Yu-Wing Tai, and Chi-Keung Tang. Ladn: Local adversarial disentangling net- work for facial makeup and de-makeup. InProceedings of the IEEE/CVF International conference on computer vision, pages 10481–10490, 2019. 2

  9. [16]

    Vqfr: Blind face restoration with vector-quantized dictionary and parallel de- coder

    Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel de- coder. InEuropean Conference on Computer Vision, pages 126–143. Springer, 2022. 2

  10. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 4

  11. [18]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 4

  12. [19]

    Image-to-image translation with conditional adver- sarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adver- sarial networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1125–1134,

  13. [20]

    Detecting gans and retouching based digital alter- ations via dad-hcnn

    Anubhav Jain, Puspita Majumdar, Richa Singh, and Mayank Vatsa. Detecting gans and retouching based digital alter- ations via dad-hcnn. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition Work- shops, pages 672–673, 2020. 1

  14. [21]

    Hierarchical mixtures of experts and the em algorithm.Neural computation, 6(2): 181–214, 1994

    Michael I Jordan and Robert A Jacobs. Hierarchical mixtures of experts and the em algorithm.Neural computation, 6(2): 181–214, 1994. 2

  15. [22]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 3

  16. [24]

    Adaface: Quality adaptive margin for face recognition

    Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750–18759, 2022. 3, 8

  17. [25]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.CoRR, abs/1412.6980, 2015. 5

  18. [26]

    Blind face restoration via deep multi-scale component dictionaries

    Xiaoming Li, Chaofeng Chen, Shangchen Zhou, Xianhui Lin, Wangmeng Zuo, and Lei Zhang. Blind face restoration via deep multi-scale component dictionaries. InEuropean conference on computer vision, pages 399–415. Springer,

  19. [27]

    Diff- bir: Toward blind image restoration with generative diffusion prior

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff- bir: Toward blind image restoration with generative diffusion prior. InEuropean conference on computer vision, pages 430–448. Springer, 2024. 1

  20. [28]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 2

  21. [29]

    Psgan++: Robust detail-preserving makeup transfer and removal.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 44(11):8538–8551,

    Si Liu, Wentao Jiang, Chen Gao, Ran He, Jiashi Feng, Bo Li, and Shuicheng Yan. Psgan++: Robust detail-preserving makeup transfer and removal.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 44(11):8538–8551,

  22. [30]

    Supermodels without photoshop: Israel’s ‘pho- toshop law’ puts focus on digitally altered images, 2015

    IBT News. Supermodels without photoshop: Israel’s ‘pho- toshop law’ puts focus on digitally altered images, 2015. http://goo.gl/3PvaEf. 1

  23. [31]

    Prnu-based detection of facial retouching.IET Bio- metrics, 9(4):154–164, 2020

    Christian Rathgeb, Angelika Botaljov, Fabian Stockhardt, Sergey Isadskiy, Luca Debiasi, Andreas Uhl, and Christoph Busch. Prnu-based detection of facial retouching.IET Bio- metrics, 9(4):154–164, 2020. 1

  24. [32]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, par...

  25. [33]

    Resdiff: Combining cnn and diffusion model for image super-resolution

    Shuyao Shang, Zhengyang Shan, Guangxing Liu, LunQian Wang, XingHua Wang, Zekai Zhang, and Jinglin Zhang. Resdiff: Combining cnn and diffusion model for image super-resolution. InProceedings of the AAAI Conference on Artificial Intelligence, pages 8975–8983, 2024. 2, 5, 6, 8

  26. [34]

    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. 2, 4, 5

  27. [35]

    Ssat: A symmetric semantic-aware transformer network for makeup transfer and removal

    Zhaoyang Sun, Yaxiong Chen, and Shengwu Xiong. Ssat: A symmetric semantic-aware transformer network for makeup transfer and removal. InProceedings of the AAAI Conference on artificial intelligence, pages 2325–2334, 2022. 2

  28. [36]

    Content-style decoupling for unsupervised makeup transfer without generating pseudo ground truth

    Zhaoyang Sun, Shengwu Xiong, Yaxiong Chen, and Yi Rong. Content-style decoupling for unsupervised makeup transfer without generating pseudo ground truth. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7601–7610, 2024. 2

  29. [37]

    Face behind makeup

    Shuyang Wang and Yun Fu. Face behind makeup. InPro- ceedings of the AAAI Conference on Artificial Intelligence,

  30. [38]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 3, 5

  31. [39]

    Restoreformer++: Towards real- world blind face restoration from undegraded key-value pairs.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    Zhouxia Wang, Jiawei Zhang, Tianshui Chen, Wenping Wang, and Ping Luo. Restoreformer++: Towards real- world blind face restoration from undegraded key-value pairs.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2

  32. [40]

    Dr2: Diffusion-based robust degradation remover for blind face restoration

    Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1704– 171...

  33. [41]

    One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Process- ing Systems, 37:92529–92553, 2024

    Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Process- ing Systems, 37:92529–92553, 2024. 1

  34. [42]

    Retouchingffhq: A large-scale dataset for fine-grained face retouching detection

    Qichao Ying, Jiaxin Liu, Sheng Li, Haisheng Xu, Zhenxing Qian, and Xinpeng Zhang. Retouchingffhq: A large-scale dataset for fine-grained face retouching detection. InPro- ceedings of the 31st ACM International Conference on Mul- timedia, pages 737–746, 2023. 1, 2, 3

  35. [43]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728–5739,

Pith tools

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