Pith. sign in

REVIEW 3 major objections 6 minor 47 references

RefSTAR: Blind Facial Image Restoration with Reference Selection, Transfer, and Reconstruction

T0 review · 3 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 By learning which facial regions in a reference photo can be safely copied to a degraded input, RefSTAR transfers only those textures and achieves state-of-the-art identity preservation in blind face restoration.

desk verdict A solid, well-ablated reference-based face restoration method with a valuable new dataset, but the headline identity-preservation number needs a check on train/test identity overlap. read the letter →

arxiv 2507.10470 v1 pith:5PUTHPQS submitted 2025-07-14 cs.CV

Zhicun Yin , Junjie Chen , Ming Liu , Zhixin Wang , Fan Li , Renjing Pei , Xiaoming Li , Rynson W.H. Lau
show 1 more author
Wangmeng Zuo
This is my paper
classification cs.CV
keywords blindfacerestorationreference-basedidentitypreservationreferenceselectioncross-attentioncycle-consistencylossone-stepdiffusionRefSel-HQdataset
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

RefSTAR claims that reference-based face restoration fails not because reference images lack useful identity information, but because models cannot tell which parts of a reference are safe to copy when pose, expression, or age differ from the degraded input. The paper builds a three-stage pipeline: a RefSel module that predicts a binary mask of texture-consistent regions, a dual-stream cross-attention that forces selected reference features into the restoration model, and a mask-compatible cycle-consistency loss that verifies those features actually appear in the output. On synthetic Celeb-Ref-Test, this reaches an ID-GT identity similarity of 82.76 against 80.32 for the next-best method, and ID-Ref of 64.53 against 53.31, while improving FID and LPIPS as well. The result matters because blind face restoration is used on old photos, surveillance stills, and social-media images, where a wrong texture copy is immediately visible to the human eye.

What carries the argument

The load-bearing machinery is a set of three components that act on the selected reference features. (1) RefSel: a U-Net-shaped binary segmentation module, initialized from the U-Net used to build the dataset, that takes the degraded input and reference image and outputs a mask $M$ of texture-consistent regions, trained with OHEM-CE loss on 10,000 manually annotated triplets. (2) Dual-Stream Cross-Attention (DSCA), the identity $\sigma(Q_{\text{in}}K_{\text{in}}^{\top}/\sqrt{d_k})V_{\text{in}} + \sigma(Q_{\text{in}}K_{\text{ref}}^{\top}/\sqrt{d_k})V_{\text{ref}}$, which replaces vanilla cross-attention's concatenated $[K_{\text{in}},K_{\text{ref}}]$ form; in the vanilla form the model can learn to send $Q_{\text{in}}K_{\text{ref}}^{\top}$ to zero and ignore the reference, whereas the dual-stream sum forces reference features into the output. (3) Mask-Compatible Cycle-Consistency Loss, $\mathcal{L}_{\text{cycle}} = \sum_i \lambda_i \ell_i(M \odot I_{\text{ref}}^{\text{out}}, M \odot I_{\text{ref}})$, which degrades the reference image and requires the restored output, used as the new reference, to reconstruct the masked reference; this is what guarantees the selected textures are present rather than merely optional. The framework is built on an Arc2Face one-step diffusion base with LoRA adapters, and the reference stream is a ReferenceNet with DSCA inserted.

What would settle it

Split the 10,000 RefSel-HQ triplets into training and held-out test sets, retrain RefSel on only the training split, and recompute mask accuracy on the held-out 250 pairs; if accuracy falls well below the reported 88% (or near chance on the conflict categories), the claimed benefit of explicit selection is not established. A complementary test: apply the trained RefSel masks to a frozen existing reference-based method and check whether ID-GT rises, which would confirm the masks, not the new transfer or cycle losses, carry the improvement.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the bottleneck in reference-guided face restoration is not the reference itself but the mechanism for deciding what to take from it. Standard training, supervised only by the ground-truth image, is ambiguous: when the reference differs from the ground truth (open versus closed mouth, different expression wrinkles, added glasses), the network cannot tell whether a texture belongs in the output, so it either copies irrelevant detail or ignores the reference entirely. RefSTAR removes this ambiguity by predicting, for every pixel, whether the reference texture is consistent with the degraded input, and then hard-wiring the selected features into the generation path. The selection is trained on RefSel-HQ, a new dataset of 10,000 ground-truth/reference pairs with manually annotated consistency masks; the transfer uses dual-stream cross-attention so the reference branch cannot be softly ignored; and the reconstruction step adds a cycle loss that degrades the reference, reconstructs it from the output, and checks the output contains the chosen texture. The paper reports that this explicit, mask-guided design is what pushes identity similarity from 80.32 to 82.76 (ID-GT) and from 53.31 to 64.53 (ID-Ref) on Celeb-Ref-Test, with FID improving from 22.90 to 21.01.

Load-bearing premise

The central benefit of explicit reference selection rests on the 250-pair RefSel test set being disjoint from the 10,000 training triplets, so the 88% accuracy measures generalization rather than memorization.

Editorial extensions

If this is right

  • Reference-guided face restoration can be made to beat non-reference methods on identity preservation by explicitly deciding which regions to copy, rather than relying on global supervision to discover that decision.
  • The RefSel-HQ dataset of 10,000 annotated consistency masks is itself a reusable resource: it provides supervised labels for training and benchmarking future reference-selection modules.
  • The dual-stream attention correction applies to any cross-attention-based reference injection; the ablations show that without it the model effectively ignores the reference even when the cycle loss is present.
  • The mask-compatible cycle loss makes reference use auditable: the output must contain the selected textures, so improvements in ID-Ref are not merely coming from the global ArcFace identity embedding.
  • On the real-world RealRef60 pairs, the method improves FID and ID-Ref over prior reference-based methods, indicating the selection and transfer mechanism generalizes to uncontrolled photos without ground truth.

Reading between the lines

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

  • The selection mask turns a previously implicit optimization choice into an inspectable artifact: a human operator could review which regions the model decided to copy, which is a step toward auditable restoration in forensic or archival settings.
  • A direct extension the paper does not run: feeding RefSel's masks to existing reference-based methods (e.g., RefLDM or RestoreID) at inference should raise their ID-GT scores if the masks capture the true cause of improvement; that experiment would isolate the selection module's contribution from the new transfer and cycle losses.
  • The cycle-reconstruction idea is invertible supervision: instead of asking 'does the output look like the reference?', it asks 'can the reference be rebuilt from the output?', a self-check that could transfer to other image-to-image tasks such as style transfer or relighting, where a paired ground truth is unavailable.
  • Because the mask is predicted from the degraded input and reference at inference, the same framework could be applied to video, where per-frame masks would need temporal smoothing to avoid flicker; the paper does not address this.
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 / 6 minor

Summary. The paper presents RefSTAR, a reference-guided blind face restoration method with three components: a RefSel module that predicts which regions of a reference image are texture-consistent with the low-quality input, a dual-stream cross-attention (DSCA) module that forces reference features into the restoration network, and a mask-compatible cycle-consistency loss that encourages the output to contain selected reference textures. The restoration backbone is a one-step diffusion model built on Arc2Face. The authors construct a new RefSel-HQ dataset of 10,000 triplets with consistency masks and evaluate on a synthetic CelebA-based test set and a 60-pair real-world set (RealRef60). They report state-of-the-art identity preservation, e.g., ID-GT 82.76 vs. 80.32 for the runner-up and ID-Ref 64.53 vs. 53.31 on the synthetic test (Table 1).

Significance. If the empirical results are sound, RefSTAR is a useful contribution: it provides a concrete way to explicitly select transferable reference regions, forces their infusion via DSCA, and releases a new annotated dataset together with code and pretrained models, which supports reproducibility. The central claim—that explicit mask-based selection plus forced feature infusion improves identity preservation in reference-guided face restoration—is meaningful and well-motivated. However, the headline identity-preservation gains rest on dataset-split assumptions that the paper does not state, so the significance is conditional on those assumptions being clarified and, if necessary, the experiments being re-run on a properly disjoint split.

major comments (3)
  1. [§4 Testing Data / Table 1] The synthetic test set is described as 1,000 identities from CelebA (§4, Testing Data), while the restoration model is trained on CelebRef-HQ, which contains 1,005 identities (§4, Training Data), and the backbone is pretrained Arc2Face. The paper does not state that the 1,000 test identities are disjoint from the 1,005 training identities or from Arc2Face pretraining identities. Because ID-GT in Table 1 is an ArcFace similarity to the ground-truth image, any identity overlap can inflate ID-GT through the generative prior's memory of that identity, independent of the proposed reference mechanism. The claimed margin (82.76 vs. 80.32) is modest in absolute terms, so the authors should either report the identity-level disjointness of the test set or re-evaluate on a held-out identity set.
  2. [§4 Testing Data / Table 2] For the RefSel evaluation, §4 states that 'we use 250 pairs from the RefSel-HQ dataset as the test set,' but it does not state that these 250 pairs are disjoint from the 10,000 training triplets or from the 800 manually annotated pairs used to initialize the U-Net. If the test pairs contributed to training or to the mask-filtering process, the 88% average accuracy in Table 2 would be an optimism-biased estimate. Please provide the exact split and confirm the test pairs were not used in any part of the mask-generation or RefSel training pipeline.
  3. [§4.1 Quantitative Comparison / Table 1] Table 1 reports no variance or statistical significance for any metric, and the real-world test set has only 60 pairs. Without confidence intervals or per-pair analysis, the 'significantly outperforming' claim for ID-GT and ID-Ref is hard to verify; in particular, the 0.9-point ID-GT difference between the selected mask and the all-ones mask in Table 3 may be within run-to-run noise. Please report standard deviations over multiple training runs or a bootstrap over test images, and discuss effect sizes.
minor comments (6)
  1. [§1 Introduction / §3.1] The abstract and introduction say that 'we manually annotated 10,000 pairs,' but §3.1 states that only 800 pairs were manually annotated and the rest were generated by a U-Net with manual filtering and adjustment; please reconcile these statements.
  2. [§4.1 / Table 2] Table 2 reports 'accuracy' for RefSel but does not define the metric (e.g., pixel accuracy, IoU, or region-level accuracy); please specify.
  3. [§3.1 / Eq. (2)] The OHEM threshold τ in Eq. (2) and the criterion for replacing the mask with an all-one mask under severe degradation (§3.1) are not specified; these are free parameters that should be stated for reproducibility.
  4. [§3.3 / Eq. (5)] In Eq. (5), IRef_out is used before it is defined; please clarify that it is the output of the cycle reconstruction (degraded reference as input, original output as reference).
  5. [§4 Testing Data] The RealRef60 dataset description is thin: the paper should state how the 60 celebrity pairs were collected, whether the reference and degraded images are of the same identity, and how identity was verified.
  6. [Throughout] There are several typos: 'F or' in the abstract, 'cross-attnetion' in §3.2, 'seleted' in Table 3, and 'alsocontains' in §3.3.

Circularity Check

1 steps flagged · score 2.0 of 10

RefSTAR's restoration results are independent and externally benchmarked; the only mild circularity is the RefSel 88% accuracy, whose test set is drawn from the same RefSel-HQ pool used to train the module without a stated split.

  1. fitted input called prediction [Sec. 3.1 (Data Engine; Reference Selection Module) and Sec. 4 (Testing Data; Table 2)]
    "For training the RefSel module, we construct a RefSel-HQ dataset through a mask generation pipeline, which contains annotating masks for 10,000 ground truth-reference pairs ... we can train our RefSel module, which is initialized with the U-Net in the data engine ... To assess the performance of RefSel, we use 250 pairs from the RefSel-HQ dataset as the test set."

    RefSel is trained on {ILQ, IRef, M} triplets from the 10,000-pair RefSel-HQ set, and it is initialized with the same U-Net that the data engine used to produce those masks. The paper's only quantitative validation of the selector is the 88% accuracy measured on, in its words, '250 pairs from the RefSel-HQ dataset as the test set,' with no statement that those pairs are disjoint from the training pool. If the 250 pairs lie inside the 10,000 training triplets, the reported accuracy is a fit to the training labels rather than a prediction on unseen data, and the Table 2 claim reduces to a re-measurement of the training distribution. Because the paper does not document the split, the claim is not demonstrably independent of its own training input.

full rationale

RefSTAR is an empirical system rather than a derivation; there is no equation chain whose conclusion is its own input. The headline ID-GT (82.76) and ID-Ref (64.53) gains in Table 1 are measured against external baselines (GFPGAN, CodeFormer, DiffBIR, RefLDM, RestoreID, etc.) on externally constructed test protocols, and each proposed component is individually ablated (Tables 3-5, Figs. 6-8), so the central claim does not reduce to a fit or to a self-citation. Alignment between the ArcFace ID loss in Eq. (6) and the ArcFace ID-GT/ID-Ref evaluation metrics, and the fact that the Arc2Face base model consumes a reference-ID embedding at inference, are training/evaluation alignments shared by all reference-based baselines; they are not circular derivations. Per the reviewing rule, two documentation gaps are flagged and weighed as benchmark-validity risks rather than circularity. First, the paper never states that the 1,000 CelebA test identities are disjoint from the 1,005 CelebRef-HQ training identities or from Arc2Face pretraining; since Table 5 shows the base model contributes roughly 10 ID-GT points (82.76 vs 72.38 with SD1.5), identity overlap could inflate the headline margin. This is a data-hygiene concern, not a circular self-reduction. Second, the RefSel 88% accuracy evaluation uses 250 pairs drawn from the same RefSel-HQ dataset on which the module was trained, without a stated train/test split, and with mask labels that originate from the same U-Net lineage that initializes RefSel; this is the one mild circularity risk, scored as a 2. There is no load-bearing self-citation (citations to [16]-[19] supply a public dataset and baselines only), no imported uniqueness theorem, and no renaming of a known result.

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

The central empirical claim relies on standard domain assumptions about degradation realism and metric validity, plus the correctness of the mask labels. No theoretical parameters are fitted beyond standard loss weights and thresholds.

free parameters (2)
  • OHEM threshold tau = not reported
    Threshold for selecting hard examples in Eq. (2); manually chosen, affects RefSel training.
  • severe degradation mask replacement rule = not quantified
    The decision to replace the annotated mask with an all-1 mask when degradation is severe is a hand-set rule that changes training labels for RefSel.
assumptions (3)
  • domain assumption Real-ESRGAN degradation model with motion and defocus deblurring approximates real-world face degradations.
    Used to generate ILQ from I for training and testing; if unrepresentative, results on RealRef60 may not generalize.
  • domain assumption ArcFace identity similarity (ID-GT, ID-Ref) is a valid proxy for identity preservation.
    Used to measure the central claim; ArcFace is also used in the training loss (Lid), so the metric is aligned with the objective.
  • domain assumption The manual mask annotations (and the U-Net labels derived from them) correctly identify texture-consistent regions.
    Ground truth for RefSel training and evaluation; errors here propagate to selection and cycle loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RefSTAR: Blind Facial Image Restoration with Reference Selection, Transfer, and Reconstruction." pith.science (2026). https://pith.science/paper/5PUTHPQS

@misc{pith2026250710470,
  author       = {Pith},
  title        = {Pith review of: RefSTAR: Blind Facial Image Restoration with Reference Selection, Transfer, and Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PUTHPQS}},
  note         = {Machine review of arXiv:2507.10470}
}
read the original abstract

Blind facial image restoration is highly challenging due to unknown complex degradations and the sensitivity of humans to faces. Although existing methods introduce auxiliary information from generative priors or high-quality reference images, they still struggle with identity preservation problems, mainly due to improper feature introduction on detailed textures. In this paper, we focus on effectively incorporating appropriate features from high-quality reference images, presenting a novel blind facial image restoration method that considers reference selection, transfer, and reconstruction (RefSTAR). In terms of selection, we construct a reference selection (RefSel) module. For training the RefSel module, we construct a RefSel-HQ dataset through a mask generation pipeline, which contains annotating masks for 10,000 ground truth-reference pairs. As for the transfer, due to the trivial solution in vanilla cross-attention operations, a feature fusion paradigm is designed to force the features from the reference to be integrated. Finally, we propose a reference image reconstruction mechanism that further ensures the presence of reference image features in the output image. The cycle consistency loss is also redesigned in conjunction with the mask. Extensive experiments on various backbone models demonstrate superior performance, showing better identity preservation ability and reference feature transfer quality. Source code, dataset, and pre-trained models are available at https://github.com/yinzhicun/RefSTAR.

Figures

Figures reproduced from arXiv: 2507.10470 by the authors.

Figure 1
Figure 1. Pipeline of our proposed RefSTAR framework, including reference selection, transfer, and reconstruction. With the RefSel [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Overview of the network structure of our RefSTAR. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Visual comparison against state-of-the-art blind face restoration methods and reference-guided face restoration methods on [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Visual comparison against state-of-the-art blind face restoration methods and reference-guided face restoration methods on real [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The visual comparison w/ or w/O RefSel. tattoos, RefSTAR excels in transferring features from the reference image. The framework also accurately replicates common skin textures, including wrinkles and age spots, with this precise texture transfer being crucial for supe…
Figure 8
Figure 8. Figure 8: Comparison w/ or w/o DSCA and Cycle-Consistency [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 7
Figure 7. Figure 7: The visualization of predicted texture-consistent regions. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 38 canonical work pages

  1. [1]

    Super-fan: In- tegrated facial landmark localization and super-resolution of real-world low resolution faces in arbitrary poses with gans

    Adrian Bulat and Georgios Tzimiropoulos. Super-fan: In- tegrated facial landmark localization and super-resolution of real-world low resolution faces in arbitrary poses with gans. In IEEE conference on computer vision and pattern recogni- tion, pages 109–117, 2018. 2

    work page 2018

  2. [2]

    Progressive semantic- aware style transformation for blind face restoration

    Chaofeng Chen, Xiaoming Li, Lingbo Yang, Xianhui Lin, Lei Zhang, and Kwan-Yee K Wong. Progressive semantic- aware style transformation for blind face restoration. In IEEE conference on computer vision and pattern recogni- tion, pages 11896–11905, 2021. 2

    work page 2021

  3. [3]

    Fsrnet: End-to-end learning face super-resolution with facial priors

    Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors. In IEEE conference on computer vision and pattern recognition, pages 2492–2501, 2018. 2

    work page 2018

  4. [4]

    Copy or not? reference-based face image restoration with fine details

    Min Jin Chong, Dejia Xu, Yi Zhang, Zhangyang Wang, David Forsyth, Gurunandan Krishnan, Yicheng Wu, and Jian Wang. Copy or not? reference-based face image restoration with fine details. In Winter Conference on Applications of Computer Vision, pages 9642–9651, 2025. 3

    work page 2025

  5. [5]

    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. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2019. 5

    work page 2019

  6. [6]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition , pages 12873–12883, 2021. 3

    work page 2021

  7. [7]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commu- nications of the ACM, pages 139–144, 2020. 5

    work page 2020

  8. [8]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in Neural Information Processing Systems , pages 6626–6637, 2017. 5

    work page 2017

Show all 47 references
  1. [9]

    Ref-ldm: A latent diffusion model for reference-based face image restoration

    Chi-Wei Hsiao, Yu-Lun Liu, Cheng-Kun Yang, Sheng-Po Kuo, Kevin Jou, and Chia-Ping Chen. Ref-ldm: A latent diffusion model for reference-based face image restoration. Advances in Neural Information Processing Systems , 37: 74840–74867, 2024. 1, 3, 5

    2024

  2. [10]

    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 IEEE Con- ference on Computer Vision and Pattern Recognition , pages 8153–8163, 2024. 2, 4

    2024

  3. [11]

    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. In IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 4401–4410, 2019. 3

    2019

  4. [12]

    Analyzing and improv- ing the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improv- ing the image quality of stylegan. In IEEE Conference on Computer Vision and Pattern Recognition , pages 8110– 8119, 2020. 3

    2020

  5. [13]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In IEEE International Conference on Computer Vision , pages 5148–5157, 2021. 5

    2021

  6. [14]

    Progressive face super-resolution via attention to facial landmark

    Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. arXiv preprint arXiv:1908.08239, 2019. 2

    1908 arXiv

  7. [15]

    Survey on deep face restoration: From non- blind to blind and beyond

    Wenjie Li, Mei Wang, Kai Zhang, Juncheng Li, Xiaom- ing Li, Yuhang Zhang, Guangwei Gao, Weihong Deng, and Chia-Wen Lin. Survey on deep face restoration: From non- blind to blind and beyond. arXiv preprint arXiv:2309.15490,

    arXiv

  8. [16]

    Learning warped guidance for blind face restoration

    Xiaoming Li, Ming Liu, Yuting Ye, Wangmeng Zuo, Liang Lin, and Ruigang Yang. Learning warped guidance for blind face restoration. In European Conference on Computer Vi- sion, pages 272–289, 2018. 1, 3

    2018

  9. [17]

    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. In European Conference on Computer Vision, pages 399–415, 2020. 1

    2020

  10. [18]

    Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion

    Xiaoming Li, Wenyu Li, Dongwei Ren, Hongzhi Zhang, Meng Wang, and Wangmeng Zuo. Enhanced blind face restoration with multi-exemplar images and adaptive spatial feature fusion. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2706–2715, 2020. 1, 3, 5

    2020

  11. [19]

    Learning dual memory dictionaries for blind face restoration

    Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration. IEEE Transactions on Pattern Analy- sis and Machine Intelligence, pages 1–13, 2022. 1, 3, 5

    2022

  12. [20]

    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. In European Conference on Computer Vision , pages 430–448. Springer, 2024. 1, 3, 5

    2024

  13. [21]

    Faceme: Robust blind face restoration with personal identification

    Siyu Liu, Zheng-Peng Duan, Jia OuYang, Jiayi Fu, Hyunhee Park, Zikun Liu, Chun-Le Guo, and Chongyi Li. Faceme: Robust blind face restoration with personal identification. arXiv preprint arXiv:2501.05177, 2025. 1, 3, 5

    2025 arXiv

  14. [22]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In IEEE Inter- national Conference on Computer Vision, pages 3730–3738,

  15. [23]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5

    2017 arXiv

  16. [24]

    Arc2face: A foundation model for id-consistent human faces

    Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In European Conference on Computer Vision , pages 241–261. Springer, 2024. 4, 5

    2024

  17. [25]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition , pages 10684– 10695, 2022. 2, 3

    2022

  18. [26]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Inter- vention (MICCAI), pages 234–241. Springer, 2015. 3

    2015

  19. [27]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page...

    2023

  20. [28]

    Blind face restoration survey

    Sait Sharipov, Bulat Nutfullin, and Narek Maloyan. Blind face restoration survey. International Journal of Open Infor- mation Technologies, 11(6):11–28, 2023. 2

    2023

  21. [29]

    Deep semantic face deblurring

    Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming- Hsuan Yang. Deep semantic face deblurring. In IEEE con- ference on computer vision and pattern recognition , pages 8260–8269, 2018. 2

    2018

  22. [30]

    Overcoming false illusions in real-world face restoration with multi-modal guided diffusion model

    Keda Tao, Jinjin Gu, Yulun Zhang, Xiucheng Wang, and Nan Cheng. Overcoming false illusions in real-world face restoration with multi-modal guided diffusion model. arXiv preprint arXiv:2410.04161, 2024. 3

    2024 arXiv

  23. [31]

    Osdface: One-step diffusion model for face restora- tion

    Jingkai Wang, Jue Gong, Lin Zhang, Zheng Chen, Xing Liu, Hong Gu, Yutong Liu, Yulun Zhang, and Xiaokang Yang. Osdface: One-step diffusion model for face restora- tion. arXiv preprint arXiv:2411.17163, 2024. 1, 3

    2024 arXiv

  24. [32]

    Stableidentity: Insert- ing anybody into anywhere at first sight

    Qinghe Wang, Xu Jia, Xiaomin Li, Taiqing Li, Liqian Ma, Yunzhi Zhuge, and Huchuan Lu. Stableidentity: Insert- ing anybody into anywhere at first sight. arXiv preprint arXiv:2401.15975, 2024. 3

    2024 arXiv

  25. [33]

    To- wards real-world blind face restoration with generative facial prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. To- wards real-world blind face restoration with generative facial prior. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9168–9178, 2021. 1, 3, 5

    2021

  26. [34]

    Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In IEEE International Conference on Computer Vision, pages 1905–1914, 2021. 3, 5

    1905

  27. [35]

    RestoreFormer: High-quality blind face restoration from undegraded key-value pairs

    Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. RestoreFormer: High-quality blind face restoration from undegraded key-value pairs. In IEEE Conference on Computer Vision and Pattern Recognition , pages 17491–17500, 2022. 1

    2022

  28. [36]

    One-step effective diffusion network for real-world image super-resolution

    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. 5

    2024

  29. [37]

    Hifacegan: Face renovation via collaborative suppression and replenishment

    Lingbo Yang, Shanshe Wang, Siwei Ma, Wen Gao, Chang Liu, Pan Wang, and Peiran Ren. Hifacegan: Face renovation via collaborative suppression and replenishment. InProceed- ings of the 28th ACM international conference on multime- dia, pages 1551–1560, 2020. 2

    2020

  30. [38]

    Pgdiff: Guiding diffusion models for versatile face restoration via partial guidance

    Peiqing Yang, Shangchen Zhou, Qingyi Tao, and Chen Change Loy. Pgdiff: Guiding diffusion models for versatile face restoration via partial guidance. Ad- vances in Neural Information Processing Systems , 36: 32194–32214, 2023. 3

    2023

  31. [39]

    Gan prior embedded network for blind face restoration in the wild

    Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, pages 672–681, 2021. 1, 3, 5

    2021

  32. [40]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models. 2023. 2, 4

    2023

  33. [41]

    Restorerid: Towards tuning- free face restoration with id preservation

    Jiacheng Ying, Mushui Liu, Zhe Wu, Runming Zhang, Zhu Yu, Siming Fu, Si-Yuan Cao, Chao Wu, Yunlong Yu, and Hui-Liang Shen. Restorerid: Towards tuning- free face restoration with id preservation. arXiv preprint arXiv:2411.14125, 2024. 1, 3, 5

    2024 arXiv

  34. [42]

    Face super-resolution guided by facial component heatmaps

    Xin Yu, Basura Fernando, Bernard Ghanem, Fatih Porikli, and Richard Hartley. Face super-resolution guided by facial component heatmaps. In Proceedings of the European con- ference on computer vision (ECCV) , pages 217–233, 2018. 2

    2018

  35. [43]

    Inserting anybody in diffusion models via celeb ba- sis

    Ge Yuan, Xiaodong Cun, Yong Zhang, Maomao Li, Chenyang Qi, Xintao Wang, Ying Shan, and Huicheng Zheng. Inserting anybody in diffusion models via celeb ba- sis. In NeurIPS, 2023. 3

    2023

  36. [44]

    Difface: Blind face restoration with diffused error contraction

    Zongsheng Yue and Chen Change Loy. Difface: Blind face restoration with diffused error contraction. IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2024. 1, 3

    2024

  37. [45]

    Instantrestore: Single-step personalized face restoration with shared-image attention

    Howard Zhang, Yuval Alaluf, Sizhuo Ma, Achuta Kadambi, Jian Wang, and Kfir Aberman. Instantrestore: Single-step personalized face restoration with shared-image attention. arXiv preprint arXiv:2412.06753, 2024. 1, 3

    2024 arXiv

  38. [46]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition , pages 586–595,

  39. [47]

    Towards robust blind face restora- tion with codebook lookup transformer

    Shangchen Zhou, Kelvin CK Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restora- tion with codebook lookup transformer. arXiv preprint arXiv:2206.11253, 2022. 1, 3, 5 10

    2022 arXiv

Pith tools

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