Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Reference-Guided Identity Preserving Face Restoration

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a reference face is best exploited both as a fused multi-level context and as an extra hard-example identity target, and that several references can be combined at inference without retraining.

desk verdict A clean, well-ablated incremental improvement over RefLDM whose 'state-of-the-art' claim is undercut by circular ArcFace metrics and test-set hyperparameter tuning; the independent FaceNet evidence shows only a small effect. read the letter →

arxiv 2505.21905 v1 pith:MXAVHT2B submitted 2025-05-28 cs.CV cs.MM

Mo Zhou , Keren Ye , Viraj Shah , Kangfu Mei , Mauricio Delbracio , Peyman Milanfar , Vishal M. Patel , Hossein Talebi This is my paper
classification cs.CVcs.MM
keywords reference-basedfacerestorationdiffusionmodelidentitypreservationhardexamplelosscompositecontextmulti-referenceinferenceclassifier-freeguidance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to show that the reference face in reference-based face restoration is underused, and that using it in two ways—as a fused multi-level conditioning context and as an extra "hard" target in the identity loss—restores more of the person's identity. On the FFHQ-Ref and CelebA-Ref-Test benchmarks, the method reports higher identity-similarity scores than prior reference-based restorers, with the largest margins on heavy degradation. It also claims that multiple reference faces can be exploited at inference time without retraining. If the result holds, reference-guided restoration becomes a matter of representation and loss design rather than more training data or bigger models.

What carries the argument

The load-bearing object is Composite Context, \(c = \text{Concat}(W_H \varphi_H(x_{\text{REF}}), W_G \varphi_G(x_{\text{REF}})) + e_{\text{position}}\), a fixed-length sequence of 198 tokens built from a high-level identity embedding and a general face representation, injected through cross-attention in a latent diffusion UNet. The Hard Example Identity Loss \(L_{\text{HID}}\) is the second mechanism: it adds the reference face as a second cosine-similarity target, with the same identity embedding used for both context and loss, and balances the two targets by \(\$\lambda$\). This targets the documented plateau of ordinary identity loss early in training. The multi-reference extension is carried by classifier-free guidance: the noise prediction is a weighted combination of unconditioned, low-quality-only, and per-reference conditioned predictions, averaged over references.

What would settle it

A blind human side-by-side identity test on the FFHQ-Ref Severe set, or evaluation with a face-recognition model whose architecture and training data differ from the one used in the loss, would show whether the reported IDS gains correspond to genuine identity preservation. If human raters see no consistent identity advantage, the central claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that identity preservation in diffusion-based face restoration is limited not by what the reference face contains but by how it is encoded and how the identity signal is supervised. The paper conditions a latent diffusion UNet on a Composite Context that concatenates a high-level identity embedding from a pretrained face-recognition network with the full token sequence from a general face-representation network, projecting both into the cross-attention space. It also replaces the standard timestep-scaled identity loss with \(L_{\text{HID}}(x_{\text{HQ}}, x_{\text{REF}}, \hat{x}) = (1-\$\lambda$)L_{\text{ID}}(x_{\text{HQ}}, \hat{x}) + \$\lambda$ L_{\text{ID}}(x_{\text{REF}}, \hat{x})\), where \(\hat{x}\) is the direct denoising estimate and the reference face plays the role of a hard example. On the FFHQ-Ref benchmarks the reported identity similarity to the ground truth rises from 0.826 to 0.843 under moderate degradation and from 0.571 to 0.609 under severe degradation, while similarity to the reference face rises from 0.624 to 0.732. The same single-reference model can use several reference faces at inference through an averaged classifier-free-guidance rule, with reported identity gains that saturate near five references.

Load-bearing premise

The paper's main identity metric is the same cosine similarity that its loss maximizes, so the strongest reading of the reported gains assumes that this similarity tracks real perceived identity rather than fitting the metric.

Editorial extensions

If this is right

  • A single-reference training setup is enough; additional reference faces can be combined at inference and the reported identity similarity rises until about five references, with no extra training.
  • Conditioning on both identity-level and appearance-level reference information recovers details like skin texture and facial marks, not just global identity.
  • Adding the reference as a hard example keeps the identity loss from collapsing to a tiny value, so identity supervision remains active throughout training and helps most on severely degraded inputs.
  • The identity gains are not accompanied by a quality collapse: perceptual and no-reference quality indices remain competitive with prior methods.

Reading between the lines

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

  • My inference: the reported drop in IDS(REF) when more reference faces are added may be partly a metric artifact, since that score is measured against the first reference only and the averaged prediction is allowed to move toward common identity traits of the other references.
  • My inference: if the single-reference training plus reference-averaged inference works because the conditioning is additive, the same training-free multi-reference trick could transfer to other diffusion restorers whose reference conditioning is also separable; the paper does not test this.
  • My inference: the FaceNet numbers in the paper are consistent but smaller than the ArcFace gains, so a cautious reading is that the direction of improvement is real but the headline magnitude may shrink under an identity metric the model never saw.
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 / 6 minor

Summary. This paper proposes a reference-based face restoration method built on a latent diffusion model. The method has three components: Composite Context, which concatenates an ArcFace identity embedding with a full FaRL token sequence to condition the UNet; Hard Example Identity Loss, which adds an ArcFace identity loss against the reference face to the existing identity loss against the ground-truth face; and a training-free multi-reference inference scheme that averages classifier-free guidance outputs over several reference contexts. The method is evaluated on FFHQ-Ref Moderate, FFHQ-Ref Severe, and CelebA-Ref-Test against CodeFormer, DiffBIR, RefLDM, and RestorerID, reporting identity metrics IDS and IDS(REF) alongside FaceNet, LPIPS, MUSIQ, NIQE, and FID. The paper claims state-of-the-art identity-preserving restoration.

Significance. Identity preservation is a recognized weakness of diffusion-based face restoration, so a method that makes better use of reference faces would be practically valuable. The paper's strengths include a clear and simple formulation, careful use of the same training and test data as RefLDM for comparability, detailed ablations with per-image standard deviations, an independent FaceNet identity metric, a wrong-reference analysis, and a training-free multi-reference extension that is easy to adopt. The central claim, however, is weakened by the fact that the headline identity metrics are exactly the ArcFace objectives optimized during training; the independent FaceNet metric shows much smaller gains. The significance of the work therefore depends on whether the identity-preservation claim can be supported by metrics and comparisons that are not entangled with the training loss.

major comments (4)
  1. [§4 Evaluation; Eq. (3); Tables 1-3] The headline metrics IDS and IDS(REF) are cosine similarities in the ArcFace embedding space, and Eq. (3) explicitly minimizes ArcFace distance to both x_HQ and x_REF. Consequently, the large reported gains on IDS(REF) — for example 0.732 vs 0.624 on FFHQ-Ref Moderate and 0.712 vs 0.554 on FFHQ-Ref Severe — are expected from direct optimization of the evaluation metric and do not by themselves establish better identity preservation. The paper acknowledges this risk in the Evaluation paragraph and adds FaceNet, but the abstract and conclusion still base the state-of-the-art claim primarily on IDS and IDS(REF). Please re-center the identity-preservation claim on the independent FaceNet results or add additional independent validation such as a different face recognition model, verification TAR@FAR, or a human study, and state the FaceNet gains explicitly in the abstract and conclusion.
  2. [§4.2 Table 6; §4 Implementation Details; Table 13] The hyperparameters lambda and the classifier-free guidance scales s_i and s_c are selected using the same test benchmarks on which the final numbers are reported. Specifically, lambda=0.6 is justified by Table 6 evaluated on FFHQ-Ref Moderate and Severe test sets, and s_i=1.2, s_c=1.2 are selected using Table 13 on FFHQ-Ref Severe. This makes the reported comparisons partially in-sample, inflating the apparent advantage over baselines whose hyperparameters may not be tuned in the same way. Please add a held-out validation split or otherwise quantify the selection-induced optimism.
  3. [Tables 1, 7, 8] The abstract states that the method 'consistently outperforms previous work,' but on image-quality metrics the method is not consistently better. For example, on FFHQ-Ref Severe the method has LPIPS 0.3647 versus 0.3366 for RefLDM and 0.3113 for CodeFormer, and FID 38.3 versus 36.0 for RefLDM. The claim should be scoped to identity preservation, or the quality trade-offs should be discussed and the abstract tempered accordingly.
  4. [§3.3 Eq. (6); Table 2] The multi-reference extension is a plausible heuristic, but the evidence for its benefit is limited to identity metrics: with more references, IDS and FaceNet improve while IDS(REF) declines and LPIPS, MUSIQ, NIQE, and FID remain essentially flat. Since IDS(REF) is one of the terms optimized by the Hard Example Identity Loss, its decline with more references deserves a more careful treatment than the current explanation, and the claim that 'the identity preservation improves' should be scoped to the metrics that actually improve.
minor comments (6)
  1. [§3.1] The word 'dimention' should be 'dimension' in the description of the projection matrix W_G.
  2. [§4 Implementation Details] The phrase 'The classifier-guidance scales' should read 'The classifier-free guidance scales' for consistency with Eq. (5).
  3. [Tables 4-6 and text] The naming is inconsistent: 'Hard Example Identity Loss' in the text and Figure 2, but 'Hard Identity Loss' in the Table 6 caption and occasionally in Section 4.2. Please unify the terminology.
  4. [§3.2] Equation (2) uses the notation sqrt(alpha_bar_t) without defining it in this paper; adding a one-line definition or a reference to the DDPM convention would make the equation self-contained.
  5. [Figure 2] The loss curves in Figure 2 lack axis labels and a numerical scale, which makes the claimed plateau and the recovery effect hard to verify. Please add labeled axes and, ideally, the training hyperparameters used to produce the curves.
  6. [§4.1, Tables 7-9] The per-image standard deviations are useful, but the paper does not report significance tests for the differences against RefLDM. A paired test or confidence interval on the FaceNet differences would strengthen the claim that the independent metric improvement is meaningful.

Circularity Check

2 steps flagged · score 4.0 of 10

ArcFace IDS and IDS(REF) are the exact objectives of the Hard Example Identity Loss, so the largest identity-preservation gains are partially circular; independent FaceNet gains are small, and the central SOTA claim is not firmly established.

  1. self definitional [Eq. (3) with Eq. (2); Section 4 'Evaluation']
    "LID(xHQ, ˆx) = √¯αt · (1−cos⟨ϕH(xHQ), ϕH( ˆx)⟩) ... LHID(xHQ,xREF, ˆx) = (1−λ)LID(xHQ, ˆx) + λLID(xREF, ˆx). ... IDS (i.e., the cosine similarity of ArcFace [16] embedding) for person identity preservation. ... Since we optimize the identity loss using the ArcFace [16] model during training, using IDS alone may not properly reflect generalization performance due to potential overfitting."

    Minimizing Eq. (3) maximizes the ArcFace cosine similarity between the restored output and both xHQ and xREF. The paper's headline metrics are exactly those quantities: IDS is the ArcFace cosine similarity between output and HQ, and IDS(REF) is the same similarity with the first reference face. Thus the largest reported gains (especially IDS(REF), +0.108 Moderate and +0.158 Severe over RefLDM) are the training objective re-measured on the test set, not independent evidence. The paper itself concedes the overfitting risk. The only independent identity metric, FaceNet, shows much smaller gains (0.850 vs 0.837 Moderate, 0.743 vs 0.733 Severe, 0.827 vs 0.821 CelebA), so the SOTA identity claim leans on a partially circular metric.

  2. fitted input called prediction [Section 4.2 'Hard Example Identity Loss Ablation' (Table 6); Section 4 'Implementation Details' (Table 13)]
    "Hence, we empirically set the λ parameter as 0.6 by default, by considering all the three identity preservation metrics. ... The classifier-guidance scales are selected as si = 1.2 and sc = 1.2 for inference."

    Table 6 sweeps λ on the FFHQ-Ref Moderate and Severe test sets and selects 0.6 by looking at IDS, FaceNet and IDS(REF) on those same test sets; Table 13 selects si=1.2 and sc=1.2 using FFHQ-Ref Severe test metrics. The same test benchmarks are then reported as the method's state-of-the-art comparison. This is a benchmark-selection loop: the reported numbers are partly fit to the evaluation set rather than a clean out-of-sample prediction, making the headline identity gains additionally forced.

full rationale

The method is not a tautology: Composite Context and the Hard Example Identity Loss are concrete architectural and loss contributions, and the model is not simply re-emitting its inputs. However, the signature identity metrics IDS and IDS(REF) are the exact ArcFace cosine terms minimized in Eq. (3), so gains on those metrics are expected by construction; the paper's own Evaluation section flags this overfitting risk. The only independent identity metric, FaceNet, shows small improvements relative to per-image standard deviations, so the central 'state-of-the-art identity preserving restoration' claim rests substantially on the circular metric. In addition, λ and the classifier-free guidance scales were selected on the same test benchmarks, further weakening the out-of-sample interpretation. Independent quality metrics (LPIPS, MUSIQ, NIQE, FID) and qualitative results still show consistent, if modest, improvements, so the circularity is partial rather than total. No load-bearing self-citation was found; the authors' own metric-learning references [40, 41] are peripheral related work.

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

The paper's central claim depends on the assumption that pre-trained ArcFace and FaRL features are sufficient and complementary for extracting identity and appearance from the reference, that the synthetic degradation pipeline matches real-world degradation, and that ArcFace similarity is a trustworthy identity metric even when optimized as the training loss. The hand-chosen hyperparameters λ, wHID, si, and sc were selected partly on test sets, which adds further assumptions about generalization.

free parameters (4)
  • lambda (λ) = 0.6
    Balances HQ versus reference in Hard Example Identity Loss (Eq. 3); chosen empirically by evaluating all three identity metrics on test sets (Table 6).
  • w_HID = 0.1
    Balancing weight for Hard Example Identity Loss in the total loss (Eq. 4); adopted from RefLDM [12], not independently tuned.
  • CFG scale s_i = 1.2
    Guidance scale for the LQ condition in Eq. (5); selected via ablation on FFHQ-Ref Severe test set (Appendix Table 13).
  • CFG scale s_c = 1.2
    Guidance scale for Composite Context in Eq. (5); selected via ablation on FFHQ-Ref Severe test set (Appendix Table 13).
assumptions (5)
  • domain assumption Pre-trained ArcFace and FaRL embeddings encode complementary identity and appearance information useful for cross-attention conditioning.
    Section 3.1: Composite Context assumes these external representations fully capture high-level identity and low-level appearance; no ablation against other feature extractors is provided.
  • standard math The DDPM/LDM formulation with concatenated LQ latent and cross-attention conditioning is a valid image restoration backbone.
    Section 3: adopts LDM [9] and the noise-prediction objective (Eq. 15 of DDPM) without modification.
  • domain assumption The second-order degradation pipeline of Real-ESRGAN simulates real-world face degradations.
    Section 4: training uses the [44] degradation simulation; if real test degradations differ, performance may drop, as acknowledged in the Limitations section.
  • ad hoc to paper Averaging classifier-free guidance outputs over multiple reference contexts in Eq. (6) is a valid multi-reference conditioning mechanism.
    Section 3.3: introduced as a training-free extension without derivation; results in Table 2 support it empirically.
  • domain assumption ArcFace cosine similarity is a reliable proxy for identity preservation even when optimized as a training loss.
    Section 4 Evaluation: IDS is used as the primary identity metric; the paper acknowledges potential overfitting but relies on it for state-of-the-art claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reference-Guided Identity Preserving Face Restoration." pith.science (2026). https://pith.science/paper/MXAVHT2B

@misc{pith2026250521905,
  author       = {Pith},
  title        = {Pith review of: Reference-Guided Identity Preserving Face Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXAVHT2B}},
  note         = {Machine review of arXiv:2505.21905}
}
read the original abstract

Preserving face identity is a critical yet persistent challenge in diffusion-based image restoration. While reference faces offer a path forward, existing reference-based methods often fail to fully exploit their potential. This paper introduces a novel approach that maximizes reference face utility for improved face restoration and identity preservation. Our method makes three key contributions: 1) Composite Context, a comprehensive representation that fuses multi-level (high- and low-level) information from the reference face, offering richer guidance than prior singular representations. 2) Hard Example Identity Loss, a novel loss function that leverages the reference face to address the identity learning inefficiencies found in the existing identity loss. 3) A training-free method to adapt the model to multi-reference inputs during inference. The proposed method demonstrably restores high-quality faces and achieves state-of-the-art identity preserving restoration on benchmarks such as FFHQ-Ref and CelebA-Ref-Test, consistently outperforming previous work.

Figures

Figures reproduced from arXiv: 2505.21905 by the authors.

Figure 1
Figure 1. Overview of our proposed method. The Composite Context and Hard Example Identity [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Loss curves of LID and LHID during the training process. The curves are truncated to the be￾ginning part of the training process. During experiments, we observe that the identity loss in Eq. (2) decreases quickly and plateaus at a very small magnitude, as shown by the blue curve in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison with other state-of-the-art face restoration methods on FFHQ-Ref [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with other state-of-the-art face restoration methods on FFHQ-Ref [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: According to our observation, when the in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Additional qualitative comparison with other state-of-the-art face restoration methods on [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Additional qualitative comparison with other state-of-the-art face restoration methods on [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Additional qualitative comparison with other state-of-the-art face restoration methods on [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Visualization of multi-reference face restoration on FFHQ-Ref Severe. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Additional visualizations with wrong reference face. This table is a continuation of the [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Additional visualizations with wrong reference face. This table is a continuation of the [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Additional visualizations with wrong reference face. This table is a continuation of the [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior

    cs.CV 2025-12 conditional novelty 6.0 of 10

    Pre-trained restoration models can be adapted to unlabeled out-of-distribution images by distilling quality-gated pseudo-targets from a frozen text-to-image diffusion oracle via mixed-supervision fine-tuning.

Reference graph

Works this paper leans on

52 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Esrgan: Enhanced super-resolution generative adversarial networks, 2018

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Chen Change Loy, Yu Qiao, and Xiaoou Tang. Esrgan: Enhanced super-resolution generative adversarial networks, 2018

    work page 2018

  2. [2]

    Chan, and Chen Change Loy

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024

    2024

  3. [3]

    Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild, 2024

    Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild, 2024

    2024

  4. [4]

    Diffbir: Towards blind image restoration with generative diffusion prior, 2024

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior, 2024

    work page 2024

  5. [5]

    Seesr: Towards semantics-aware real-world image super-resolution

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456–25467, 2024

    2024

  6. [6]

    Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization

    Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, , and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. InThe European Conference on Computer Vision (ECCV) 2024, 2023

    work page 2024

  7. [7]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

    2020

  8. [8]

    Denoising diffusion implicit models, 2022

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022

    2022

Show all 52 references
  1. [9]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

    2022

  2. [10]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    2023

  3. [11]

    Chan, Chongyi Li, and Chen Change Loy

    Shangchen Zhou, Kelvin C.K. Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. InNeurIPS, 2022

    2022

  4. [12]

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

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

    2024

  5. [13]

    Restorerid: Towards tuning-free face restoration with id preservation, 2024

    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, 2024

    2024

  6. [14]

    Instantrestore: Single-step personalized face restoration with shared-image attention, 2024

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

    2024

  7. [15]

    Learning dual memory dictionaries for blind face restoration, 2022

    Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration, 2022

    2022

  8. [16]

    Arcface: Additive angular margin loss for deep face recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962–5979, October 2022

    Jiankang Deng, Jia Guo, Jing Yang, Niannan Xue, Irene Kotsia, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5962–5979, October 2022

    2022

  9. [17]

    General facial representation learning in a visual-linguistic manner

    Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021

    2021 arXiv

  10. [18]

    Osdface: One-step diffusion model for face restoration, 2025

    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 restoration, 2025. 10

    2025

  11. [19]

    Facenet: A unified embedding for face recognition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 815–823. IEEE, June 2015

    2015

  12. [20]

    A metric learning reality check, 2020

    Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check, 2020

    2020

  13. [21]

    Revisiting training strategies and generalization performance in deep metric learning, 2020

    Karsten Roth, Timo Milbich, Samarth Sinha, Prateek Gupta, Björn Ommer, and Joseph Paul Cohen. Revisiting training strategies and generalization performance in deep metric learning, 2020

    2020

  14. [22]

    Classifier-free diffusion guidance, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022

    2022

  15. [23]

    Consistency models, 2023

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models, 2023

    2023

  16. [24]

    Diffusion models beat gans on image synthesis, 2021

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021

    2021

  17. [25]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021

    2021

  18. [26]

    The power of context: How multimodality improves image super-resolution

    Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Patel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  19. [27]

    Fsrnet: End-to-end learning face super-resolution with facial priors, 2017

    Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. Fsrnet: End-to-end learning face super-resolution with facial priors, 2017

    2017

  20. [28]

    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. InCVPR, 2020

    2020

  21. [29]

    Towards real-world blind face restoration with generative facial prior, 2021

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior, 2021

    2021

  22. [30]

    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. InNeurIPS, 2023

    2023

  23. [31]

    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. InECCV, 2020

    2020

  24. [32]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023

    2023

  25. [33]

    Pdgrad: Guiding diffusion model for reference-based blind face restoration with pivot direction gradient guidance.Sensors, 24(22), 2024

    Geon Min, Tae Bok Lee, and Yong Seok Heo. Pdgrad: Guiding diffusion model for reference-based blind face restoration with pivot direction gradient guidance.Sensors, 24(22), 2024

    2024

  26. [34]

    Pfstorer: Personal- ized face restoration and super-resolution, 2024

    Tuomas Varanka, Tapani Toivonen, Soumya Tripathy, Guoying Zhao, and Erman Acar. Pfstorer: Personal- ized face restoration and super-resolution, 2024

    2024

  27. [35]

    Faceme: Robust blind face restoration with personal identification, 2025

    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, 2025

    2025

  28. [36]

    Llv-fsr: Exploiting large language-vision prior for face super-resolution, 2024

    Chenyang Wang, Wenjie An, Kui Jiang, Xianming Liu, and Junjun Jiang. Llv-fsr: Exploiting large language-vision prior for face super-resolution, 2024

    2024

  29. [37]

    U-net: Convolutional networks for biomedical image segmentation, 2015

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015

    2015

  30. [38]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015

    2015

  31. [39]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  32. [40]

    Mo Zhou and Vishal M. Patel. Enhancing adversarial robustness for deep metric learning, 2022

    2022

  33. [41]

    Adversarial attack and defense in deep ranking.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5306– 5324, 2024

    Mo Zhou, Le Wang, Zhenxing Niu, Qilin Zhang, Nanning Zheng, and Gang Hua. Adversarial attack and defense in deep ranking.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(8):5306– 5324, 2024

    2024

  34. [42]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 11

    2023

  35. [43]

    A style-based generator architecture for generative adversarial networks, 2019

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019

    2019

  36. [44]

    Real-esrgan: Training real-world blind super- resolution with pure synthetic data, 2021

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super- resolution with pure synthetic data, 2021

    2021

  37. [45]

    Pali: A jointly-scaled multilingual language-image model, 2023

    Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James B...

    2023

  38. [46]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018

    2018

  39. [47]

    Musiq: Multi-scale image quality transformer, 2021

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer, 2021

    2021

  40. [48]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012

    2012

  41. [49]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

    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 equilibrium, 2018

    2018

  42. [50]

    Webface260m: A benchmark unveiling the power of million-scale deep face recognition, 2021

    Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, and Jie Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition, 2021

    2021

  43. [51]

    Deepfake generation and detection: A benchmark and survey, 2024

    Gan Pei, Jiangning Zhang, Menghan Hu, Zhenyu Zhang, Chengjie Wang, Yunsheng Wu, Guangtao Zhai, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey, 2024

    2024

  44. [52]

    identity blending

    Pierre Fernandez, Hady Elsahar, I. Zeki Yalniz, and Alexandre Mourachko. Video seal: Open and efficient video watermarking.arXiv preprint arXiv:2412.09492, 2024. A Detailed Quantitative Results and More Visualizations The detailed results and comparison with state-of-the-art m...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.