Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Visual Style Prompt Learning Using Diffusion Models for Blind Face Restoration

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that blind face restoration is best driven by denoised StyleGAN latent codes used as visual style prompts, combined with a style-modulated multi-dilation layer that captures both local details and global context.

desk verdict Solid latent-space diffusion method for face restoration, but the claimed first/second ranking needs error bars; still deserves a serious referee. read the letter →

arxiv 2412.21042 v1 pith:TSR2ZKHU submitted 2024-12-30 cs.CV cs.MM

Wanglong Lu , Jikai Wang , Tao Wang , Kaihao Zhang , Xianta Jiang , Hanli Zhao This is my paper · ORCID
classification cs.CVcs.MM
keywords blindfacerestorationvisualstylepromptsdiffusionprobabilisticmodelsGANlatentspaceW+codescodediffuserstyle-modulatedaggregationtransformationgenerativefacialpriors
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

Blind face restoration is the task of recovering a recognizable, high-quality face from an image damaged by unknown blur, noise, downsampling, or compression. The paper proposes that the right way to guide such restoration is to first convert the degraded face into a visual style prompt: a denoised latent code inside a pretrained StyleGAN, produced by a diffusion model that refines a rough initial code. The denoised code is then used to generate candidate facial features and to modulate a restoration network through a new style-modulated aggregation transformation layer. The authors report that this prompt-based scheme improves perceptual quality on both synthetic and real-world face datasets and that it also sharpens downstream tasks such as facial landmark detection and emotion recognition.

What carries the argument

The load-bearing object is the diffusion-based style prompt module operating on StyleGAN's W+ latent codes (512-dimensional style vectors per layer). A style encoder produces initial codes; a "code diffuser" made of four temporal-aware code-to-code blocks predicts the noise added at each diffusion step, conditioned on the initial code and step index, and repeated denoising yields the final visual prompt. The second mechanism is the SMART layer: it takes style vectors from the denoised code, applies StyleGAN-style modulation and demodulation to 3x3 convolution kernels, runs the reweighted kernels at four dilation rates, and aggregates the resulting feature maps to capture both local and distant context. Together these convert the problem of "guess what the clean face looks like" into "produce a latent code whose StyleGAN rendering is the clean face," then use that rendering's features to guide restoration.

What would settle it

Take a set of faces with strong makeup, extreme expressions, or occlusions; compute the LPIPS distance between each ground-truth face and its own StyleGAN inversion using the paper's encoder and generator. If the restored output is not consistently closer to the ground truth than this self-reconstruction baseline, the claim that the StyleGAN feature bank supplies faithful guidance is falsified: the method's quality would be bounded exactly by the face manifold the generator can express.

Watch

Extended reading notes

Core claim

The paper's central claim is that a degradation-ridden face image can be converted into a clean visual prompt by running a small diffusion model directly in the W+ style latent space of a pretrained StyleGAN, and that this prompt, together with StyleGAN's internal feature maps, is enough to drive a restoration auto-encoder to a faithful face. The style encoder first maps the degraded face to an initial W+ code; the code diffuser then removes noise from a randomly sampled code over T steps, conditioned on that initial code. Training the diffuser with diffusion loss plus LPIPS and identity losses on the StyleGAN rendering of the denoised code makes the final prompt align perceptually and semantically with the ground truth. A style-modulated aggregation transformation layer injects the prompt into the restoration network by rescaling 3x3 convolution kernels and fusing feature maps at dilation rates 1, 2, 4, and 8, letting the network recover both fine details and global context.

Load-bearing premise

The whole pipeline assumes the fixed pretrained StyleGAN generator can represent every face it is asked to restore, so faces outside that generator's learned range—unusual identities, strong makeup, extreme expressions—cannot be faithfully recovered no matter how accurate the prompt diffuser is.

Editorial extensions

If this is right

  • If the central claim is right, blind face restoration can be done by diffusion in a compact latent space rather than slow pixel-space denoising, keeping inference fast enough for practical use.
  • The visual prompts are renderable: denoised codes can be fed to StyleGAN to show exactly what facial attributes the restoration is being guided toward, making the guidance interpretable.
  • The SMART layer gives one network both local detail and long-range context, suggesting the same layer could benefit other image-to-image restoration tasks that need context reasoning.
  • Because the prompts live in a latent space that supports editing, restoration and attribute manipulation could be driven by the same style codes, opening a route to text-guided restoration.
  • The method's applications to landmark detection and emotion recognition suggest that restoration can serve as a preprocessing step for face-related analysis tasks.

Reading between the lines

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

  • A testable extension of the latent-diffusion idea: apply the same code diffuser to other domains with pretrained generators, such as full-body or animal synthesis, by swapping StyleGAN for the matching generator; the framework makes no face-specific assumption beyond a W-like latent space.
  • The fixed dilation rates {1, 2, 4, 8} are a heuristic; a natural variant predicts dilation rates per spatial location, which could sharpen the local/global trade-off on textured regions like hair.
  • Because the method's ceiling is the StyleGAN manifold, identity fidelity on uncommon faces could be improved by fine-tuning or growing the feature bank on those identities, rather than by improving the prompt diffuser.
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 / 4 minor

Summary. The paper proposes VSPBFR, a blind face restoration framework that uses a pretrained StyleGAN generator as a facial feature bank and a diffusion-based code diffuser to transform e4e-style initial latent codes into denoised W+ style prompts. A restoration auto-encoder with a style-modulated aggregation transformation (SMART) layer integrates these prompts, random styles, and StyleGAN features. Training combines diffusion, LPIPS, identity, and adversarial losses. Evaluation is on the synthetic CelebA-Test and three real-world benchmarks, with ablations and two downstream applications. The central claim is that the method ranks first on real-world datasets (mean rank 3.67) and second on synthetic CelebA-Test (mean rank 3.75).

Significance. If the reported ranking is reproducible, the work is a solid contribution: it moves diffusion-based denoising into the W+ latent space, achieving much faster inference than pixel-space diffusion methods (about 81 ms per image), and the SMART layer provides a plausible mechanism for combining global context with local detail. The use of standard supervised losses and external benchmarks avoids circularity, and the paper includes ablations, efficiency analysis, downstream applications, and a source-code link. The main limitation is statistical: the state-of-the-art claims rest on single-run comparisons with very small metric margins and no uncertainty quantification, so the current evidence is moderate rather than conclusive.

major comments (4)
  1. [Section IV-B, Tables I and II] The central claim of first/second ranking rests on mean ranks computed from single evaluations without error bars, standard deviations, random seeds, or significance tests. The differences between the top methods are smaller than typical run-to-run variability of FID/NIQE on small sets: for example, WebPhoto-Test FID is 74.25 (Ours) vs 75.06 (RestoreFormer++), CelebChild-Test NIQE is 4.12 vs 4.15, and LFW-Test FID is 46.48 vs 46.31 (DifFace). With no uncertainty quantification, the reported first/second rank is not established. Please report bootstrap or multi-seed variability and state whether the rank differences are statistically meaningful; otherwise the claim should be softened to 'competitive with state of the art.'
  2. [Section IV-A vs Section IV-D and Fig. 8] The test protocol is inconsistent. The implementation states T=4 'for both training and testing,' but Fig. 7 and Fig. 8 display and evaluate denoising steps up to 10, and the text says the FID 'saturates at step 7.' Clarify how steps beyond 4 are sampled at inference time, and use the same number of steps in the reported quantitative evaluation; if the model can run with an arbitrary number of steps, explain why training used T=4 and why the main tables do not use the apparently better step count.
  3. [Section IV-F, Table V and text] The paper labels the application improvements as 'relative improvements' when the reported numbers are absolute percentage-point differences. NME decreases from 6.08% to 2.43% (a 3.65-point absolute drop, about 60% relative), and emotion recognition accuracy rises from 79.73% to 86.03% (a 6.30-point absolute gain, about 7.9% relative). Please correct the terminology so the application results are reported accurately.
  4. [Table IV vs Tables I and II] The 'Ours' row in the ablation study is trained for 400,000 iterations, while the final model in Tables I and II is trained for 500,000 iterations, and the metric values differ accordingly (e.g., WebPhoto-Test FID 74.80 vs 74.25, NIQE 4.33 vs 4.28). The internal ablation comparisons are valid because all variants share the same training budget, but the text should explicitly state that the 'Ours' row in Table IV is not the final model and should not be directly compared with the final results in Table II.
minor comments (4)
  1. [Throughout] The text alternates between 'code diffuser' and 'coder diffuser' (e.g., Section IV-C and Table IV); please unify to 'code diffuser.'
  2. [Equation (9)] The adversarial loss has a missing closing parenthesis in `log(1 − D(Iout)]`; it should be `log(1 − D(Iout))`.
  3. [Section IV-D] Calling the e4e-encoded latent codes of ground-truth images 'ground-truth latent codes' is imprecise; they are encoder estimates, and the comparison in Fig. 8 is between images rendered from two encoder/denoiser outputs. Please rephrase to avoid implying a canonical ground-truth latent code.
  4. [Section IV-A, Table I caption] Define FID† and unpaired FID in the caption or directly in the metrics paragraph, because the current distinction is stated only in the running text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is a supervised training pipeline evaluated against external benchmarks, with no prediction that reduces by construction to its own inputs.

full rationale

The paper's central claim is that the proposed visual style prompt learning framework achieves high-quality blind face restoration, supported by quantitative comparisons on CelebA-Test and three real-world datasets against 14 external methods. The code diffuser is trained with a diffusion loss plus LPIPS and identity losses between the StyleGAN-rendered inverted image and the ground-truth image, and the restoration auto-encoder is trained with LPIPS, identity, and adversarial losses. These are standard supervised objectives tied to ground-truth images, not to the evaluation metrics being reported. The ground-truth latent codes in Sec. IV-D are explicitly obtained by feeding ground-truth images into the style encoder, which is a baseline inversion choice rather than a circular definition of success; the paper even reports that denoised codes surpass this baseline's FID, an independent empirical result. Self-citations such as PromptRR [19] appear only in related-work context and are not load-bearing for the derivation. The lack of error bars and code verification is a legitimate robustness concern, but it does not constitute circularity. No step in the paper reduces by construction to its own inputs or renames a fitted parameter as a prediction, so the circularity score is 0.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several untested background assumptions typical for deep learning face restoration: StyleGAN's latent space covers the face distribution, e4e provides reliable inversions, the synthetic degradation model transfers to real-world data, and pre-trained VGG/ArcFace losses are effective proxies for perceptual and identity fidelity. The hand-set hyperparameters (loss weights, diffusion steps, dilation rates) are free parameters chosen by the authors; the network weights themselves are learned from data and are not counted here. The paper proposes no new physical or ontological entities; the 'visual style prompt' is a conceptual reuse of existing StyleGAN latent codes, validated only by in-domain experiments.

free parameters (9)
  • Number of diffusion steps T = 4
    Set to 4 for training and testing; analysis in Sec. IV-D shows FID saturates around step 7, so T=4 is a hand-chosen trade-off between quality and speed.
  • TACC block count = 4
    The code diffuser uses four temporal-aware code-to-code blocks; this number is set by the authors without a derived optimality criterion.
  • Variance schedule beta_1, beta_T = beta_1=0.1, beta_T=0.99 (linear)
    Chosen for the code diffuser's forward process; standard DDPM schedule, not derived from face restoration data.
  • Code diffuser loss weight lambda_lpips = 0.1
    Hand-set in total loss O(theta_p) in Sec. III-D.
  • Code diffuser loss weight lambda_id = 0.1
    Hand-set in total loss O(theta_p) in Sec. III-D.
  • Restoration loss weight lambda_hat_id = 0.1
    Hand-set in total objective O(theta_g, theta_f, theta_d) in Sec. III-D.
  • Restoration loss weight lambda_hat_lpips = 0.5
    Hand-set in total objective O(theta_g, theta_f, theta_d) in Sec. III-D.
  • R1 regularization strength gamma = 10
    Set in the adversarial loss in Eq. (9), common for StyleGAN training.
  • Dilation rates in SMART layer = [1, 2, 4, 8]
    Chosen to capture local to global context; not optimized or derived.
assumptions (5)
  • domain assumption The pre-trained StyleGAN generator (S_theta_s) provides a well-disentangled latent space that can faithfully represent the faces to be restored.
    The entire pipeline is built around StyleGAN W+ latent codes; Sec. III-A defines the facial feature bank using S_theta_s.
  • domain assumption The e4e style encoder provides a reasonable mapping from images to W+ latent codes for both clean and degraded faces.
    The code diffuser is conditioned on initial codes from e4e and evaluated against e4e inversions of ground-truth images (Sec. III-D, IV-D).
  • domain assumption The synthetic degradation model (Gaussian blur, downsampling, noise, JPEG) with sampled parameters approximates real-world face degradation.
    Training data is generated with this model (Sec. IV-A); generalization to real-world datasets is assumed to follow.
  • ad hoc to paper Four temporal-aware code-to-code blocks (TACC) suffice to model the denoising transition in W+ space.
    The code diffuser architecture uses 4 TACC blocks and T=4 steps; no proof of sufficiency is given, but experiments support it.
  • domain assumption LPIPS and ArcFace identity losses steer the predicted latent codes toward perceptually and identity-preserving solutions.
    Used in code diffuser and restoration training (Sec. III-D), relying on pre-trained VGG and ArcFace.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual Style Prompt Learning Using Diffusion Models for Blind Face Restoration." pith.science (2026). https://pith.science/paper/TSR2ZKHU

@misc{pith2026241221042,
  author       = {Pith},
  title        = {Pith review of: Visual Style Prompt Learning Using Diffusion Models for Blind Face Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSR2ZKHU}},
  note         = {Machine review of arXiv:2412.21042}
}
read the original abstract

Blind face restoration aims to recover high-quality facial images from various unidentified sources of degradation, posing significant challenges due to the minimal information retrievable from the degraded images. Prior knowledge-based methods, leveraging geometric priors and facial features, have led to advancements in face restoration but often fall short of capturing fine details. To address this, we introduce a visual style prompt learning framework that utilizes diffusion probabilistic models to explicitly generate visual prompts within the latent space of pre-trained generative models. These prompts are designed to guide the restoration process. To fully utilize the visual prompts and enhance the extraction of informative and rich patterns, we introduce a style-modulated aggregation transformation layer. Extensive experiments and applications demonstrate the superiority of our method in achieving high-quality blind face restoration. The source code is available at \href{https://github.com/LonglongaaaGo/VSPBFR}{https://github.com/LonglongaaaGo/VSPBFR}.

Figures

Figures reproduced from arXiv: 2412.21042 by the authors.

Figure 1
Figure 1. The overall pipeline of our framework: the degraded image is processed through a diffusion-based style prompt module (a) to get denoised codes [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The detailed diffusion (←) and denoising (→) processes in the style latent space. We also show the corresponding inverted images of latent codes in steps. where ˆϵ t = Pθp (wt , wˆ , t); variance σ 2 t = 1−α¯t−1 1−α¯t βt and noise ϵ ∼ N (0, I). Code diffuser. Our code diffuser is with four temporal￾aware code-to-code blocks (TACC). Each block is an extension of FFCLIP’s semantic modulation block [36]. It sets Γ t i−… view at source ↗
Figure 3
Figure 3. Illustration of the style-modulated aggregation transformation (SMART). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Visual comparisons of our method and the SOTA facial restoration methods. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visual comparisons of our method and the SOTA facial restoration methods. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visual examples from the ablation study with (a) removing the coder diffuser, (b) replacing our SMART layers with style layers, (c) removing the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visualization of denoised latent codes for each step and restored images. Through the denoising process, the codes incrementally reveal more clear [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: FID comparisons between denoised latent codes and ground-truth latent codes. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visualization of feature maps from SMART layers at resolutions from [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Visualization of facial landmark detection on degraded images, restored images, and ground-truth images. Green points correspond to visualized [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Visualization of face emotion recognition using HSEmotion and our enhanced. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

  1. [1]

    T. Wang, K. Zhang, X. Chen, W. Luo, J. Deng, T. Lu, X. Cao, W. Liu, H. Li, S. Zafeiriou, A survey of deep face restoration: Denoise, super- resolution, deblur, artifact removal, arXiv preprint arXiv:2211.02831 (2022)

    arXiv 2022

  2. [2]

    Hsieh, P.-C

    P.-W. Hsieh, P.-C. Shao, Blind image deblurring based on the sparsity of patch minimum information, PR 109 (2021) 107597

    work page 2021

  3. [3]

    Shen, Z.-Q

    H. Shen, Z.-Q. Zhao, W. Liao, W. Tian, D.-S. Huang, Joint operation and attention block search for lightweight image restoration, PR 132 (2022) 108909

    work page 2022

  4. [4]

    Y . Wang, X. Bai, Versatile recurrent neural network for wide types of video restoration, PR 138 (2023) 109360

    work page 2023

  5. [5]

    Zhao, S.-K

    S. Zhao, S.-K. Oh, J.-Y . Kim, Z. Fu, W. Pedrycz, Motion-blurred image restoration framework based on parameter estimation and fuzzy radial basis function neural networks, PR 132 (2022) 108983

    work page 2022

  6. [6]

    T. Wang, G. Tao, W. Lu, K. Zhang, W. Luo, X. Zhang, T. Lu, Restoring vision in hazy weather with hierarchical contrastive learning, PR 145 (2024) 109956

    work page 2024

  7. [7]

    M. Wang, W. Lu, J. Lyu, K. Shi, H. Zhao, Generative image inpainting with enhanced gated convolution and transformers, Displays 75 (2022) 102321

    work page 2022

  8. [8]

    W. Lu, X. Jiang, X. Jin, Y .-L. Yang, M. Gong, T. Wang, K. Shi, H. Zhao, Grig: Few-shot generative residual image inpainting (2023). arXiv: 2304.12035

    work page Pith review arXiv 2023

Show all 46 references
  1. [9]

    Zheng, W

    J. Zheng, W. Lu, W. Wang, Y . Cao, X. Zhang, X. Jiang, Handling the non-smooth challenge in tensor svd: A multi-objective tensor recovery framework, in: ECCV , 2025, pp. 449–464

    2025

  2. [10]

    C. Chen, X. Li, L. Yang, X. Lin, L. Zhang, K.-Y . K. Wong, Progressive semantic-aware style transformation for blind face restoration, in: CVPR, 2021, pp. 11896–11905

    2021

  3. [11]

    Z. Wang, J. Zhang, R. Chen, W. Wang, P. Luo, Restoreformer: High- quality blind face restoration from undegraded key-value pairs, in: CVPR, 2022, pp. 17512–17521

    2022

  4. [12]

    X. Wang, Y . Li, H. Zhang, Y . Shan, Towards real-world blind face restoration with generative facial prior, in: CVPR, 2021, pp. 9168–9178

    2021

  5. [13]

    Menon, A

    S. Menon, A. Damian, S. Hu, N. Ravi, C. Rudin, Pulse: Self-supervised photo upsampling via latent space exploration of generative models, in: CVPR, 2020, pp. 2437–2445

    2020

  6. [14]

    Z. Wang, J. Zhang, T. Chen, W. Wang, P. Luo, Restoreformer++: Towards real-world blind face restoration from undegraded key-value paris, TPAMI 45 (12) (2023) 15462–15476

    2023

  7. [15]

    Karras, S

    T. Karras, S. Laine, T. Aila, A style-based generator architecture for generative adversarial networks, TPAMI 43 (12) (2021) 4217–4228

    2021

  8. [16]

    van den Oord, O

    A. van den Oord, O. Vinyals, k. kavukcuoglu, Neural discrete represen- tation learning, in: NeurIPS, V ol. 30, 2017, p. 6309–6318

    2017

  9. [17]

    Potlapalli, S

    V . Potlapalli, S. W. Zamir, S. Khan, F. Khan, PromptIR: Prompting for all-in-one image restoration, in: NeurIPS, 2023, p. 19

    2023

  10. [18]

    B. Xia, Y . Zhang, S. Wang, Y . Wang, X. Wu, Y . Tian, W. Yang, L. Van Gool, Diffir: Efficient diffusion model for image restoration, in: ICCV , 2023, pp. 13095–13105

    2023

  11. [19]

    T. Wang, W. Lu, K. Zhang, W. Luo, T.-K. Kim, T. Lu, H. Li, M.-H. Yang, Promptrr: Diffusion models as prompt generators for single image reflection removal, arXiv preprint arXiv:2402.02374 (2024)

    2024 arXiv

  12. [20]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, in: NeurIPS, V ol. 33, 2020, pp. 6840–6851

    2020

  13. [21]

    Roich, R

    D. Roich, R. Mokady, A. H. Bermano, D. Cohen-Or, Pivotal tuning for latent-based editing of real images, TOG 42 (1) (2022) 1–13

    2022

  14. [22]

    O. Tov, Y . Alaluf, Y . Nitzan, O. Patashnik, D. Cohen-Or, Designing an encoder for stylegan image manipulation, TOG 40 (4) (2021) 1–14

    2021

  15. [23]

    Y . Gu, X. Wang, L. Xie, C. Dong, G. Li, Y . Shan, M.-M. Cheng, Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder, in: ECCV , 2022, pp. 126–143

    2022

  16. [24]

    Karras, S

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, T. Aila, Analyzing and improving the image quality of stylegan, in: CVPR, 2020, pp. 8107–8116

    2020

  17. [25]

    Zhang, K

    P. Zhang, K. Zhang, W. Luo, C. Li, G. Wang, Blind face restoration: Benchmark datasets and a baseline model, Neurocomputing 574 (2024) 127271

    2024

  18. [26]

    Z. Chen, J. He, X. Lin, Y . Qiao, C. Dong, Towards real-world video face restoration: A new benchmark, in: CVPRW, 2024, pp. 5929–5939

    2024

  19. [27]

    Z. Wan, B. Zhang, D. Chen, P. Zhang, D. Chen, J. Liao, F. Wen, Bringing old photos back to life, in: CVPR, 2020, pp. 2747–2757

    2020

  20. [28]

    T. Yang, P. Ren, X. Xie, L. Zhang, Gan prior embedded network for blind face restoration in the wild, in: CVPR, 2021, pp. 672–681

    2021

  21. [29]

    F. Zhu, J. Zhu, W. Chu, X. Zhang, X. Ji, C. Wang, Y . Tai, Blind face restoration via integrating face shape and generative priors, in: CVPR, 2022, pp. 7662–7671

    2022

  22. [30]

    Y . Chen, Y . Tai, X. Liu, C. Shen, J. Yang, Fsrnet: End-to-end learning face super-resolution with facial priors, in: CVPR, 2018, pp. 2492–2501

    2018

  23. [31]

    P. Yang, S. Zhou, Q. Tao, C. C. Loy, PGDiff: Guiding diffusion models for versatile face restoration via partial guidance, in: NeurIPS, 2023, pp. 1–13

    2023

  24. [32]

    W. Lu, H. Zhao, X. Jiang, X. Jin, Y .-L. Yang, K. Shi, Do inpainting yourself: Generative facial inpainting guided by exemplars, Neurocom- puting 617 (2025) 128996

    2025

  25. [33]

    W. Lu, J. Wang, X. Jin, X. Jiang, H. Zhao, Facemug: A multimodal generative and fusion framework for local facial editing, TVCG (2024) 1–15

    2024

  26. [34]

    X. Qiu, C. Han, Z. Zhang, B. Li, T. Guo, X. Nie, Diffbfr: Bootstrapping diffusion model for blind face restoration, ACM MM (2023) 7785–7795

    2023

  27. [35]

    Z. Yue, C. C. Loy, Difface: Blind face restoration with diffused error contraction, TPAMI (2024) 1–15

    2024

  28. [36]

    Y . Zhu, H. Liu, Y . Song, Z. Yuan, X. Han, C. Yuan, Q. Chen, J. Wang, One model to edit them all: Free-form text-driven image manipulation with semantic modulations, in: NeurIPS, V ol. 35, 2022, pp. 25146– 25159

    2022

  29. [37]

    J. Fu, J. Liu, H. Tian, Y . Li, Y . Bao, Z. Fang, H. Lu, Dual attention network for scene segmentation, in: CVPR, 2019, pp. 3146–3154

    2019

  30. [38]

    S. Zhao, J. Cui, Y . Sheng, Y . Dong, X. Liang, E. I. Chang, Y . Xu, Large scale image completion via co-modulated generative adversarial networks, in: ICLR, 2021, p. 25

    2021

  31. [39]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, O. Wang, The unrea- sonable effectiveness of deep features as a perceptual metric, in: CVPR, 2018, pp. 586–595

    2018

  32. [40]

    S. Zhou, K. C. Chan, C. Li, C. C. Loy, Towards robust blind face restoration with codebook lookup transformer, in: A. H. Oh, A. Agarwal, D. Belgrave, K. Cho (Eds.), NeurIPS, 2022, pp. 30599–30611

    2022

  33. [41]

    Z. Wang, Z. Zhang, X. Zhang, H. Zheng, M. Zhou, Y . Zhang, Y . Wang, Dr2: Diffusion-based robust degradation remover for blind face restora- tion, in: CVPR, 2023, pp. 1704–1713

    2023

  34. [42]

    Z. Liu, P. Luo, X. Wang, X. Tang, Deep learning face attributes in the wild, in: ICCV , 2015, pp. 3730–3738

    2015

  35. [43]

    G. B. Huang, M. Ramesh, T. Berg, E. Learned-Miller, Labeled faces in the wild: A database for studying face recognition in unconstrained environments, Tech. Rep. 07-49, University of Massachusetts (2007)

    2007

  36. [44]

    Karras, T

    T. Karras, T. Aila, S. Laine, J. Lehtinen, Progressive growing of GANs for improved quality, stability, and variation, in: ICLR, 2018, p. 26

    2018

  37. [45]

    Bulat, G

    A. Bulat, G. Tzimiropoulos, How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks), in: ICCV , 2017, pp. 1021–1030

    2017

  38. [46]

    Savchenko, Facial expression recognition with adaptive frame rate based on multiple testing correction, in: ICML, V ol

    A. Savchenko, Facial expression recognition with adaptive frame rate based on multiple testing correction, in: ICML, V ol. 202, 2023, pp. 30119–30129

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.