Pith. sign in

REVIEW 3 major objections 6 minor 84 references

LAFR: Efficient Diffusion-based Blind Face Restoration via Latent Codebook Alignment Adapter

T0 review · 3 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 A codebook-based latent alignment adapter lets a frozen Stable Diffusion VAE restore degraded faces from just 600 training images, matching state-of-the-art face restoration with 7.5M trainable parameters.

desk verdict The 600-image efficiency claim is not credible from the paper's own tables; the method idea is salvageable but this draft overclaims. read the letter →

arxiv 2505.23462 v1 pith:AWTF2OR2 submitted 2025-05-29 cs.CV

Runyi Li , Bin Chen , Jian Zhang , Radu Timofte This is my paper
classification cs.CV
keywords blindfacerestorationdiffusionmodelslatentcodebookalignmentefficientfine-tuningLoRAidentitypreservationsmall-dataadaptation
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

LAFR is a blind face restoration method built on a pre-trained Stable Diffusion model. The paper's central claim is that a lightweight codebook-based latent-alignment adapter, combined with LoRA fine-tuning of the denoising UNet's convolutional layers, closes the gap between low-quality and high-quality latents so well that only 600 training images (0.9% of FFHQ) and 7.5M trainable parameters are needed to match state-of-the-art face restoration methods, while cutting training time by 70%. The authors argue that faces are unusually regular in layout, so a small sample suffices, and they validate this with experiments on synthetic and real-world benchmarks. They also note in the limitations section that the codebook design rests on this facial regularity and is not expected to transfer to diverse natural images.

What carries the argument

The central object is the Latent Codebook Alignment Adapter, a three-part module placed between the frozen VAE encoder and the denoising UNet: a shallow feature extractor, a codebook lookup that selects the nearest HQ anchor for each LQ latent, and a mapping network that turns the selected anchor into an aligned latent. It is trained with an $L_1$ alignment loss against the HQ latent. Around this adapter, the paper adds a multi-level restoration loss — $\mathcal{L}_{\text{res}}$, a CLIP-based identity loss, and a facial-structure cosine loss — and uses LoRA on the UNet's convolutional layers, with text and timestep embeddings replaced by fixed tensors, to keep training and inference cheap.

What would settle it

Train LAFR with codebooks of 128, 256, 512, and 1024 entries on the same 600-image set and measure identity-preservation metrics (Deg., LMD) and FID on CelebA-Test; if quality collapses as the codebook shrinks, the bottleneck — not the LoRA tuning — is carrying the restored identity, and any identity detail beyond 1024 anchors is already lost.

Watch

Extended reading notes

Core claim

The discovery is that the semantic misalignment between low-quality (LQ) and high-quality (HQ) latents in a VAE trained only on HQ images can be repaired by a small adapter without retraining the VAE. The adapter extracts a feature $f$ from the LQ latent, finds the nearest anchor $c^* = \arg\min_{c_i\in\mathcal{C}}\|f-c_i\|_2$ in a learned codebook of 1024 entries, and produces an aligned latent $z_{\text{aligned}} = M(c^*)$ through a small mapping network; a simple $L_1$ loss against the ground-truth HQ latent trains it. Restoring from $z_{\text{aligned}}$ instead of the raw LQ latent substantially improves quality, and when this adapter is combined with LoRA fine-tuning of convolution layers and a multi-level loss (image appearance, CLIP identity embeddings, and facial structure), the paper reports performance comparable to state-of-the-art diffusion methods on CelebA-Test and real-world benchmarks, using only 600 FFHQ images and 7.5M trainable parameters.

Load-bearing premise

The load-bearing premise is that every degraded face can be faithfully carried through a nearest-neighbor lookup against just 1024 learned anchor codes; the paper itself concedes in Section E that this codebook design depends on faces sharing highly similar layouts and would not transfer to diverse natural images.

Editorial extensions

If this is right

  • If the central claim holds, blind face restoration no longer requires full VAE retraining or a 70K-image dataset; the same pre-trained diffusion prior can be adapted from a few hundred examples.
  • The two-stage design decouples latent alignment from restoration, so the alignment adapter can be retrained or swapped without touching the denoising UNet.
  • Pruning the text and timestep embedding modules and fixing them as precomputed tensors cuts inference cost and parameter count with, per the ablation, negligible quality loss.
  • Because the multi-level loss supervises identity embeddings and facial structure directly against the HQ ground truth, the method avoids relying on unreliable LQ-conditioned guidance during sampling.
  • Training time falls by about 70% relative to large-scale fine-tuning, making single-GPU, small-data deployment realistic.

Reading between the lines

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

  • A test the paper does not run: shrink the codebook from 1024 anchors toward 128 while monitoring identity metrics; the point at which quality collapses would reveal how much identity information the bottleneck itself carries versus the mapping network and LoRA layers.
  • The same adapter idea could be probed on other structured image domains (e.g., document, medical, or plate images) that also form compact layout distributions; the paper's own limitation note predicts it would fail on unconstrained natural images, so such domains are the natural testbed.
  • Since the identity loss uses a CLIP vision encoder, which is not trained specifically for face recognition, an ArcFace-style embedding in the loss might shift the fidelity-identity trade-off; the paper uses ArcFace only for evaluation, not supervision.
  • The t-SNE compactness argument for faces is qualitative; a formal measure across many identity and pose groups would sharpen the claim about why 600 images suffice.
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 proposes LAFR, a blind face restoration method that adapts a pre-trained Stable Diffusion model using a lightweight codebook-based latent alignment adapter and LoRA fine-tuning of the denoising UNet's convolutional layers. The authors claim that training on only 600 FFHQ images (0.9% of FFHQ) with 7.5M trainable parameters achieves performance comparable to state-of-the-art methods while reducing training time by 70%. Experiments are reported on CelebA-Test and real-world benchmarks (LFW-Test, WebPhoto-Test, Wider-Test) across a comprehensive set of full-reference and no-reference metrics.

Significance. If the central efficiency claim were solidly supported, LAFR would be a practically valuable contribution: it would demonstrate that a tiny data budget can suffice for face-restoration domain adaptation, with a compact adapter and low training cost. The paper also has a plausible architectural idea (latent-space codebook alignment without retraining the VAE) and an honest discussion of scope in Sec. E. However, the quantitative evidence is internally inconsistent, and the central claim is therefore not verifiable from the manuscript as written.

major comments (3)
  1. [Sec. 4.2 (Table 1) vs. Sec. B (Tables 8, 9)] The same LAFR configuration at 600 training images is reported as PSNR 26.26 in Table 1 (and Table 5, row #5) but 25.91 in Table 8. The same OSEDiff retrained on 70K images is 25.88 PSNR in Table 1 but 26.85 in Table 9. These discrepancies directly affect the headline comparison against state-of-the-art methods and the 'comparable to SOTA' claim. No error bars, random-seed information, or code/checkpoints are provided to establish which numbers are correct.
  2. [Sec. 4.3 (Fig. 6) and Sec. B (Table 8)] Table 8 shows LAFR PSNR rising to 26.64 at 800 images and 26.43 at 7,000 images, then collapsing to 24.77 at 21,000 and 25.16 at 70,000 images, which is below the 600-image value of 25.91. This non-monotonic scaling is not discussed and contradicts the text's statement that 'Beyond 600 training images, the improvements become marginal.' The data as presented suggest the 600-image choice is post hoc and that the model becomes unstable or is differently configured at larger scales.
  3. [Sec. 3.3] The training objective for the codebook is underspecified. The forward pass selects c* = argmin ||f - c_i|| and computes zaligned = M(c*), with Lalign = ||zaligned - zHQ||_1. Because argmin is non-differentiable, this loss cannot update the codebook entries c_i unless a straight-through estimator or commitment loss is used; none is described. As written, the mechanism by which the 1024 anchors are 'learned from HQ training data' is not reproducible.
minor comments (6)
  1. [Sec. 4.4] The sentence 'Conditioning the model on degraded image could lead to superior restoration quality' appears to be the opposite of the authors' argument (which is that LQ conditioning is unreliable). This is likely a typo and should be corrected.
  2. [Sec. 3.1] The t-SNE/ResNet-50 evidence uses CelebA versus ImageNet, while the method is trained on FFHQ; the connection between the two is not made. In addition, a silhouette score of 0.18 is conventionally interpreted as weak cluster structure, so the claim of 'a high degree of structural coherence' is overstated.
  3. [Sec. B] Tables 8 and 9 do not report the full set of metrics from Table 1 (e.g., LPIPS, DISTS, FID), making it impossible to compare the scaling behavior with the main results.
  4. [Abstract and Sec. 1] The abstract and Sec. 1 call 600 images '0.9% of FFHQ'; the exact fraction is 600/70,000 ≈ 0.857%. Please state the precise number or define the rounding.
  5. [Fig. 9] Fig. 9 does not appear to match the numbers in Tables 8 and 9; for example, the plotted PSNR values are inconsistent with the tabulated values. Please ensure the figure and tables correspond to the same runs.
  6. [Sec. E] Sec. E appropriately limits the method to faces, but this limitation is not mentioned in the abstract or introduction; consider stating the scope upfront.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed derivation chain is an empirical training/evaluation pipeline with external benchmarks, and no load-bearing step reduces by construction to its own inputs.

full rationale

The paper's central claims are (1) a codebook-based latent alignment adapter corrects LQ/HQ VAE latent mismatch, and (2) 600 FFHQ images suffice for competitive face restoration. Neither claim is circular by the paper's own equations. The adapter maps an LQ latent feature f to a nearest codebook anchor c* and then to an aligned latent zaligned = M(c*), trained with Lalign = ||zaligned - zHQ||1 against the ground-truth HQ latent code (Sec. 3.3). This is direct supervised regression, not a quantity that is defined in terms of the result it is used to predict. The multi-level loss (Sec. 3.4) uses external encoders (CLIP, D3DFR) and pixel/feature losses; again, these are training objectives, not re-labeled predictions. The claim that facial images are structurally regular is supported by a t-SNE/silhouette analysis (Sec. 3.1), which is motivation rather than a derivation of the efficiency result. The data-efficiency conclusion is an empirical claim tested on CelebA-Test and real-world benchmarks, with external metrics (PSNR, SSIM, LPIPS, FID, NIQE, Deg, LMD); no fitted parameter is presented as a prediction. The paper does cite related work including work by co-author Jian Zhang (Ref. [59], DDNM), but that citation is background information and is not load-bearing. The most notable concern is internal inconsistency among reported PSNR values across Tables 1, 8, and 9, and the non-monotonic scaling curve in Table 8 (e.g., LAFR at 600 images PSNR 25.91 vs. 25.16 at 70K images). That is a reproducibility/accuracy problem that could undermine the empirical support for the efficiency claim, but it is not a circularity: the 600-image result is not constructed from the conclusion that 600 images suffice. No step in the paper equates an output with an input by definition, imports a unique solution from a self-citation, or renames a known empirical pattern as a derivation. Therefore the appropriate finding is no significant circularity, with score 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central claim rests on a small number of domain assumptions about facial regularity and the validity of the HQ VAE latent space as a target, plus hand-set hyperparameters. The most fragile input is the codebook bottleneck: the paper's own limitation section concedes the design depends on facial layout regularity. No new physical entities are introduced; the codebook is a learned component, listed here for completeness.

free parameters (6)
  • codebook vocabulary size = 1024
    Set by hand; capacity of the alignment bottleneck. No sensitivity study is reported (Sec. 4.1).
  • codebook hidden dimension = 256
    Dimension of the feature extractor output and codebook entries; chosen without ablation.
  • LoRA rank = 4
    Rank for low-rank adaptation of conv layers only; no rank sweep is shown.
  • loss weights = lambda_res = lambda_id = lambda_fs = 1; LPIPS lambda = 2
    Equal weighting in Eq. 3, chosen by hand; only on/off ablations are reported, no weight sensitivity.
  • training schedule = 100 epochs (stage 1), 17,000 steps (stage 2)
    Fixed schedule; directly determines the 70% training-time saving claim.
  • training subset size = 600 FFHQ images (0.9%)
    Headline data budget. Table 8 shows 800 images gives 26.64 PSNR versus 25.91 at 600, so 600 is not an obvious plateau and appears selected post hoc.
assumptions (6)
  • domain assumption Faces share a highly regular layout and distribution across identities
    Used to justify both small-data fine-tuning and codebook alignment; stated in Sec. 3.1 and in Limitation E, where the authors concede it may not hold for natural images.
  • domain assumption The frozen Stable Diffusion VAE latent space is a valid target space for aligning LQ images
    The method maps LQ latents to HQ latents with an L1 loss and never retrains or adjusts the VAE; if the HQ latent space is not semantically smooth, the alignment objective is ill-posed (Sec. 3.3).
  • ad hoc to paper Nearest-neighbor lookup into 1024 learned anchors is a sufficient bottleneck for latent alignment
    This is the paper's specific design bet; no comparison to other bottleneck sizes beyond one swap with VQFR's module in Table 7.
  • domain assumption CLIP (IP-Adapter) features and D3DFR structure features are valid measures of identity and facial structure
    Used to define the multi-level loss (Eqs. 1 and 2); if these embeddings do not track perceived identity, the loss cannot preserve it.
  • domain assumption Removing text and timestep embeddings and fixing the prompt to 'face, high quality' preserves face restoration quality
    Only validated for the face domain by ablation in Table 5 (#3 versus #4); may not transfer.
  • domain assumption The pre-trained Stable Diffusion v2-1 prior contains a useful face distribution
    The paper relies on this prior for generating face details after light LoRA tuning (Sec. 3.4).
invented entities (1)
  • Latent codebook of 1024 HQ semantic anchors
    purpose: Acts as a discrete dictionary for mapping LQ VAE latents to HQ-aligned latents before diffusion sampling (Sec. 3.3).
    The anchors are learned from 600 FFHQ images and have no falsifiable handle outside this paper; their benefit is shown only through internal ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LAFR: Efficient Diffusion-based Blind Face Restoration via Latent Codebook Alignment Adapter." pith.science (2026). https://pith.science/paper/AWTF2OR2

@misc{pith2026250523462,
  author       = {Pith},
  title        = {Pith review of: LAFR: Efficient Diffusion-based Blind Face Restoration via Latent Codebook Alignment Adapter},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AWTF2OR2}},
  note         = {Machine review of arXiv:2505.23462}
}
read the original abstract

Blind face restoration from low-quality (LQ) images is a challenging task that requires not only high-fidelity image reconstruction but also the preservation of facial identity. While diffusion models like Stable Diffusion have shown promise in generating high-quality (HQ) images, their VAE modules are typically trained only on HQ data, resulting in semantic misalignment when encoding LQ inputs. This mismatch significantly weakens the effectiveness of LQ conditions during the denoising process. Existing approaches often tackle this issue by retraining the VAE encoder, which is computationally expensive and memory-intensive. To address this limitation efficiently, we propose LAFR (Latent Alignment for Face Restoration), a novel codebook-based latent space adapter that aligns the latent distribution of LQ images with that of HQ counterparts, enabling semantically consistent diffusion sampling without altering the original VAE. To further enhance identity preservation, we introduce a multi-level restoration loss that combines constraints from identity embeddings and facial structural priors. Additionally, by leveraging the inherent structural regularity of facial images, we show that lightweight finetuning of diffusion prior on just 0.9% of FFHQ dataset is sufficient to achieve results comparable to state-of-the-art methods, reduce training time by 70%. Extensive experiments on both synthetic and real-world face restoration benchmarks demonstrate the effectiveness and efficiency of LAFR, achieving high-quality, identity-preserving face reconstruction from severely degraded inputs.

Figures

Figures reproduced from arXiv: 2505.23462 by the authors.

Figure 1
Figure 1. Comparison of our proposed LAFR with state-of-the-art face restoration methods. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of alignment strategies of LQ-HQ code in restoration. (a) Our proposed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overall pipeline of our proposed LAFR. We keep VAE encoder and decoder frozen, only [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualized results on CelebA-Test dataset. RF++ is short for RestoreFormer++ [63]. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualized results of real-world face restoration. RF++ is RestoreFormer++ [ [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Performance trend on CelebA-Test as the amount of the training images increases. The [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: T-SNE results show that compared to natural [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: Performance trend of OSEDiff and our LAFR on CelebA-Test, as the amount of training [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Results on animation images with real-world degradations. Our method yields high-quality [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

84 extracted references · 62 canonical work pages

  1. [1]

    Identity-preserving diffusion for face restoration

    Xiaying Bai, Yuhao Yang, Wenming Yang, Rui Zhu, and Jing-Hao Xue. Identity-preserving diffusion for face restoration. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2025

    2025

  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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11896–11905, 2021

    2021

  3. [3]

    Faithdiff: Unleashing diffusion priors for faithful image super-resolution

    Junyang Chen, Jinshan Pan, and Jiangxin Dong. Faithdiff: Unleashing diffusion priors for faithful image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  4. [4]

    Towards real-world blind face restoration with generative diffusion prior

    Xiaoxu Chen, Jingfan Tan, Tao Wang, Kaihao Zhang, Wenhan Luo, and Xiaochun Cao. Towards real-world blind face restoration with generative diffusion prior. IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  5. [5]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009

    work page 2009

  6. [6]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2019

    2019

  7. [7]

    Accurate 3d face reconstruc- tion with weakly-supervised learning: From single image to image set

    Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruc- tion with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2019

    work page 2019

  8. [8]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In Proceedings of the Advances in Neural Information Processing Sysstems (NeurIPS), 2021

    work page 2021

Show all 84 references
  1. [9]

    Image quality assessment: Unifying structure and texture similarity

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020

    2020

  2. [10]

    Tsd-sr: One-step diffusion with target score distillation for real-world image super- resolution

    Linwei Dong, Qingnan Fan, Yihong Guo, Zhonghao Wang, Qi Zhang, Jinwei Chen, Yawei Luo, and Changqing Zou. Tsd-sr: One-step diffusion with target score distillation for real-world image super- resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern ...

    2025

  3. [11]

    Generative diffusion prior for unified image restoration and enhancement

    Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  4. [12]

    Infobfr: Real-world blind face restoration via information bottleneck

    Nan Gao, Jia Li, Huaibo Huang, Ke Shang, and Ran He. Infobfr: Real-world blind face restoration via information bottleneck. arXiv preprint arXiv:2501.15443, 2025

    2025 arXiv

  5. [13]

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

    Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, and Ming-Ming Cheng. Vqfr: Blind face restoration with vector-quantized dictionary and parallel decoder. In Proceedings of the European Conference on Computer Vision (ECCV), 2022

    2022

  6. [14]

    Deep residual learning for image recognition

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

    2016

  7. [15]

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

    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. In Advances in neural information processing systems (NeurIPS), volume 30, 2017

    2017

  8. [16]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  9. [17]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  10. [18]

    Face restoration via plug-and-play 3d facial priors

    Xiaobin Hu, Wenqi Ren, Jiaolong Yang, Xiaochun Cao, David Wipf, Bjoern Menze, Xin Tong, and Hongbin Zha. Face restoration via plug-and-play 3d facial priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8910–8926, 2021. 13

    2021

  11. [19]

    Labeled faces in the wild: A database forstudying face recognition in unconstrained environments

    Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database forstudying face recognition in unconstrained environments. In Workshop on faces in’Real- Life’Images: detection, alignment, and recognition, 2008

    2008

  12. [20]

    Decomposed diffusion sampler for accelerating large-scale inverse problems

    Chung Hyungjin, Lee Suhyeon, and Ye Jong Chul. Decomposed diffusion sampler for accelerating large-scale inverse problems. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  13. [21]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019

    2019

  14. [22]

    Denoising diffusion restoration models

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  15. [23]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148–5157, 2021

    2021

  16. [24]

    Auto-encoding variational bayes

    Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In Proceedings of the Interna- tional Conference on Learning Representations (ICLR), 2014

    2014

  17. [25]

    Dual prompting image restoration with diffusion transformers

    Dehong Kong, Fan Li, Zhixin Wang, Jiaqi Xu, Renjing Pei, Wenbo Li, and WenQi Ren. Dual prompting image restoration with diffusion transformers. arXiv preprint arXiv:2504.17825, 2025

    2025 arXiv

  18. [26]

    Enhancing variational autoencoders with smooth robust latent encoding

    Hyomin Lee, Minseon Kim, Sangwon Jang, Jongheon Jeong, and Sung Ju Hwang. Enhancing variational autoencoders with smooth robust latent encoding. arXiv preprint arXiv:2504.17219, 2025

    2025 arXiv

  19. [27]

    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 Proceedings of the European Conference on Computer Vision (ECCV), pages 399–415, 2020

    2020

  20. [28]

    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 Proceedings of the European Conference on Computer Vision (ECCV), September 2018

    2018

  21. [29]

    Authface: Towards au- thentic blind face restoration with face-oriented generative diffusion prior.arXiv preprint arXiv:2410.09864, 2024

    Guoqiang Liang, Qingnan Fan, Bingtao Fu, Jinwei Chen, Hong Gu, and Lin Wang. Authface: Towards au- thentic blind face restoration with face-oriented generative diffusion prior.arXiv preprint arXiv:2410.09864, 2024

    2024

  22. [30]

    Diffusion restoration adapter for real-world image restoration

    Hanbang Liang, Zhen Wang, and Weihui Deng. Diffusion restoration adapter for real-world image restoration. arXiv preprint arXiv:2502.20679, 2025

    2025 arXiv

  23. [31]

    Catch missing details: Image reconstruction with frequency augmented variational autoencoder

    Xinmiao Lin, Yikang Li, Jenhao Hsiao, Chiuman Ho, and Yu Kong. Catch missing details: Image reconstruction with frequency augmented variational autoencoder. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1736–1745, 2023

    2023

  24. [32]

    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

    2025 arXiv

  25. [33]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730–3738, 2015

    2015

  26. [34]

    3d priors-guided diffusion for blind face restoration

    Xiaobin Lu, Xiaobin Hu, Jun Luo, Ben Zhu, Yaping Ruan, and Wenqi Ren. 3d priors-guided diffusion for blind face restoration. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1829–1838, 2024

    2024

  27. [35]

    Pulse: Self-supervised photo upsampling via latent space exploration of generative models

    Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  28. [36]

    Diffusion autoencoders: Toward a meaningful and decodable representation

    Konpat Preechakul, Nattanat Chatthee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10619–10629, June 2022

    2022

  29. [37]

    Diffbfr: Bootstrapping diffusion model for blind face restoration

    Xinmin Qiu, Congying Han, ZiCheng Zhang, Bonan Li, Tiande Guo, and Xuecheng Nie. Diffbfr: Bootstrapping diffusion model for blind face restoration. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7785–7795, 2023. 14

    2023

  30. [38]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on M...

    2021

  31. [39]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  32. [40]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis

    Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53–65, 1987

    1987

  33. [41]

    Image super-resolution via iterative refinement

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022

    2022

  34. [42]

    Cluster quality analysis using silhouette score

    Ketan Rajshekhar Shahapure and Charles Nicholas. Cluster quality analysis using silhouette score. In 2020 IEEE 7th International Conference on Data Science and Advanced Analytics (DSAA) , pages 747–748, 2020

    2020

  35. [43]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  36. [44]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  37. [45]

    Clr-face: Conditional latent refinement for blind face restoration using score-based diffusion models

    Maitreya Suin and Rama Chellappa. Clr-face: Conditional latent refinement for blind face restoration using score-based diffusion models. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2024

    2024

  38. [46]

    Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration

    Maitreya Suin, Nithin Gopalakrishnan Nair, Chun Pong Lau, Vishal M Patel, and Rama Chellappa. Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vis...

    2024

  39. [47]

    Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach

    Lingchen Sun, Rongyuan Wu, Zhiyuan Ma, Shuaizheng Liu, Qiaosi Yi, and Lei Zhang. Pixel-level and semantic-level adjustable super-resolution: A dual-lora approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  40. [48]

    Dual associated encoder for face restoration

    Yu-Ju Tsai, Yu-Lun Liu, Lu Qi, Kelvin CK Chan, and Ming-Hsuan Yang. Dual associated encoder for face restoration. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  41. [49]

    Joint face image restoration and frontalization for recognition

    Xiaoguang Tu, Jian Zhao, Qiankun Liu, Wenjie Ai, Guodong Guo, Zhifeng Li, Wei Liu, and Jiashi Feng. Joint face image restoration and frontalization for recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1285–1298, 2022

    2022

  42. [50]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  43. [51]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008

    2008

  44. [52]

    Pfstorer: Person- alized face restoration and super-resolution

    Tuomas Varanka, Tapani Toivonen, Soumya Tripathy, Guoying Zhao, and Erman Acar. Pfstorer: Person- alized face restoration and super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2372–2381, June 2024

    2024

  45. [53]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, AidanN. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  46. [54]

    Old photo restoration via deep latent space translation

    Ziyu Wan, Bo Zhang, Dong Chen, Pan Zhang, Dong Chen, Fang Wen, and Jing Liao. Old photo restoration via deep latent space translation. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(2):2071–2087, 2023

    2023

  47. [55]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 15

    2023

  48. [56]

    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. International Journal of Computer Vision (IJCV), 2024

    2024

  49. [57]

    Osdface: One-step diffusion model for face restoration

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  50. [58]

    Towards real-world blind face restoration with generative facial prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  51. [59]

    Zero-shot image restoration using denoising diffusion null-space model

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR), 2023

    2023

  52. [60]

    Sinsr: diffusion-based image super-resolution in a single step

    Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image super-resolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (C...

    2024

  53. [61]

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

    Zhixin Wang, Ziying Zhang, Xiaoyun Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1704–1713, 2023

    2023

  54. [62]

    Image quality assessment: from error visibility to structural similarity

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

    2004

  55. [63]

    Restoreformer++: Towards real-world blind face restoration from undegraded key-value pairs

    Zhouxia Wang, Jiawei Zhang, Tianshui Chen, Wenping Wang, and Ping Luo. Restoreformer++: Towards real-world blind face restoration from undegraded key-value pairs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15462–15476, 2023

    2023

  56. [64]

    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. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  57. [65]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  58. [66]

    Diffbir: Towards blind image restoration with generative diffusion prior

    Lin Xinqi, He Jingwen, Chen Ziyan, Lyu Zhaoyang, Dai Bo, Yu Fanghua, Ouyang Wanli, Qiao Yu, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior. In Proceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  59. [67]

    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. In Advances in Neural Information Processing Systems, 2023

    2023

  60. [68]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment

    Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  61. [69]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 672–681, 2021

    2021

  62. [70]

    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. arXiv preprint arxiv:2308.06721, 2023

    2023 arXiv

  63. [71]

    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

    2024 arXiv

  64. [72]

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

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

    2023

  65. [73]

    Adaptbir: Adaptive blind image restoration with latent diffusion prior for higher fidelity

    Liu Yingqi, He Jingwen, Liu Yihao, Lin Xinqi, Yu Fanghua, Hu Jinfan, Qiao Yu, and Dong Chao. Adaptbir: Adaptive blind image restoration with latent diffusion prior for higher fidelity. Pattern Recognition, 155:110659, 2024. 16

    2024

  66. [74]

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

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2024

  67. [75]

    Arbitrary-steps image super-resolution via diffusion inversion

    Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2025

    2025

  68. [76]

    Difface: Blind face restoration with diffused error contraction

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

    2024

  69. [77]

    Resshift: Efficient diffusion model for image super-resolution by residual shifting

    Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  70. [78]

    A feature-enriched completely blind image quality evaluator

    Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579–2591, 2015

    2015

  71. [79]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE International Conference on Computer Vision (ICCV), 2023

    2023

  72. [80]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  73. [81]

    Towards authentic face restoration with iterative diffusion models and beyond

    Yang Zhao, Tingbo Hou, Yu-Chuan Su, Xuhui Jia, Yandong Li, and Matthias Grundmann. Towards authentic face restoration with iterative diffusion models and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7312–7322, October 2023

    2023

  74. [82]

    Rethinking deep face restoration

    Yang Zhao, Yu-Chuan Su, Chun-Te Chu, Yandong Li, Marius Renn, Yukun Zhu, Changyou Chen, and Xuhui Jia. Rethinking deep face restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652–7661, 2022

    2022

  75. [83]

    Towards robust blind face restoration with codebook lookup transformer

    Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In Advances in Neural Information Processing Systems, 2022

    2022

  76. [84]

    Oftsr: One-step flow for image super-resolution with tunable fidelity-realism trade-offs

    Yuanzhi Zhu, Ruiqing Wang, Shilin Lu, Junnan Li, Hanshu Yan, and Kai Zhang. Oftsr: One-step flow for image super-resolution with tunable fidelity-realism trade-offs. arXiv preprint arXiv:2412.09465, 2024. 17

    2024

Pith tools

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