Pith. sign in

REVIEW 4 major objections 6 minor 77 references

HonestFace: Towards Honest Face Restoration with One-Step Diffusion Model

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 HonestFace claims state-of-the-art face restoration via a one-step diffusion model that preserves identity and realistic texture.

desk verdict A useful incremental improvement over OSDFace, but the paper's own A-LD metric is used both to select references and as the headline identity metric, making the SOTA claim conditional until A-LD is independently validated. read the letter →

arxiv 2505.18469 v1 pith:DV43WZLH submitted 2025-05-24 cs.CV

Jingkai Wang , Wu Miao , Jue Gong , Zheng Chen , Xing Liu , Hong Gu , Yutong Liu , Yulun Zhang This is my paper · ORCID
classification cs.CV
keywords facerestorationone-stepdiffusionidentitypreservationreference-basedaffinelandmarkdistancemaskedalignmentperceptuallosstexturerealism
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

HonestFace is a face-restoration model built on a one-step diffusion backbone that takes a low-quality face plus several high-quality reference photos of the same person and outputs a restored face in a single step. The paper's central claim is that this model is 'honest': it keeps the person's identity stable, including eye color, wrinkles, skin texture, and overall facial structure, while avoiding the over-smoothed, plastic, or repetitive textures that many generative restorers produce. To get there, the authors add an identity embedder that pulls fine-grained features from the best-matching reference and global identity cues from all references, and a masked face alignment loss that concentrates restoration effort on perceptually important regions such as eyes and mouth. They also propose a new evaluation metric, the affine landmark distance, which measures landmark error after removing global offset, rotation, and scale, arguing it reflects identity geometry better than raw L2 landmark distance. If the claims hold, HonestFace is a practical single-step restorer that beats prior reference-based and no-reference methods on the reported quality and identity metrics.

What carries the argument

The central object is the one-step latent diffusion generator $G_\psi(x_L, p)$ with prompt embedding $p$ formed by three parallel embedders: the visual representation embedder (VRE) that tokenizes the low-quality input, and the proposed identity embedder (IDE) composed of a facial feature extractor (VGG16 features of eye patches from the best-matching reference) and a face identity encoder (attention aggregation of ArcFace embeddings of all references). Training is driven by adversarial distillation plus three losses: an identity loss (cosine distance on an ensemble of face recognizers), a global perceptual loss (edge-aware DISTS), and the proposed masked face alignment loss (LPIPS on heatmap-masked blended images). The proposed affine landmark distance, $d_{\text{A-LD}}(L,H;W)$, solves a weighted least-squares problem for the optimal 2D affine transform mapping restored landmarks to ground-truth landmarks and takes the residual; it is used both to pick the best reference image and as the paper's identity metric.

What would settle it

A blind user study would settle it: show human raters pairs of restored images from HonestFace and each baseline, ask which better preserves identity and texture, then compute how often A-LD and the reported metrics agree with the human choice. If A-LD, or the other identity metrics, systematically disagree with human judgment, the central claim of superior identity honesty collapses.

Watch

Extended reading notes

Core claim

The paper claims that HonestFace delivers state-of-the-art face restoration by combining three components: an identity embedder (a facial feature extractor for fine details like eye color and skin texture plus a face identity encoder aggregating multiple references), a masked face alignment loss focusing perceptual fidelity on landmarks such as eyes, mouth, and brows, and a new affine landmark distance metric that evaluates identity consistency after factoring out global pose. On the CelebHQRef-Test and Reface-Test sets, HonestFace reports the best scores on every perceptual, no-reference, and identity metric, including LPIPS, DISTS, CLIP-IQA, MANIQA, MUSIQ, FID, degradation angle, L2 landmark distance, and the proposed affine landmark distance. The authors also state that HonestFace avoids the over-smoothing, repetitive hair textures, and color shifts typical of diffusion- and transformer-based restorers, while keeping the one-step inference speed of its baseline.

Load-bearing premise

The load-bearing premise is that the proposed affine landmark distance, which both selects the reference image and serves as the paper's main identity metric, is a trustworthy measure of identity consistency: if it disagrees with human perception or with face-recognition embeddings, the reported identity improvements and the reference-selection strategy would lose their support.

Editorial extensions

If this is right

  • HonestFace outperforms all compared reference-based methods (ASFFNet, DMDNet, PGDiff, FaceMe, MGFR) on every reported metric on both CelebHQRef-Test and Reface-Test, including perceptual, no-reference, and identity metrics.
  • It also beats no-reference restorers like CodeFormer, DAEFR, OSEDiff, and its own baseline OSDFace on perceptual and pixel metrics, while performing one-step inference at roughly 0.13 seconds per 512x512 image on an A6000 GPU.
  • The identity embedder and masked face alignment loss each contribute measurable gains in identity consistency and texture naturalness, as shown by the paper's ablation studies.
  • The affine landmark distance is introduced as a more pose-invariant alternative to raw L2 landmark distance, with potential use both as a reference-selection criterion and as an evaluation metric for identity consistency.
  • HonestFace preserves color fidelity better than the compared methods, as measured by the color-aware MS-SWD metric, addressing a known failure mode of diffusion-based restorers.

Reading between the lines

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

  • Because the affine landmark distance is closed-form, cheap to compute, and invariant to global affine transforms, it could be reused as a training loss or as a general face-alignment evaluation tool beyond restoration, though the paper does not explore these uses.
  • A-LD may reward methods that globally align the face while still altering internal features; an independent check against ArcFace similarity or human identity ratings would clarify whether the metric truly tracks identity preservation.
  • The paper leaves open how HonestFace behaves when the provided reference images are of a different identity, are poorly aligned, or are themselves degraded, since all experiments assume reliable same-identity references.
  • The reported 0.13-second inference time is on a specific GPU; a direct wall-clock comparison across devices and frameworks would be needed to substantiate the practical speed advantage beyond the paper's setting.
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. The paper proposes HonestFace, a one-step latent diffusion model for reference-based face restoration. The method introduces an identity embedder (IDE) that combines a facial feature extractor and a face identity encoder, a masked face alignment (MFA) loss that uses landmark heatmaps to focus perceptual losses on salient facial regions, and a new affine landmark distance (A-LD) metric for identity evaluation. The model is fine-tuned from OSDFace on Reface-HQ and CelebRef-HQ and evaluated on CelebHQRef-Test and Reface-Test. The authors report improvements over reference-based baselines (ASFFNet, DMDNet, PGDiff, FaceMe, MGFR) and over several no-reference methods on perceptual and identity metrics, and conclude that HonestFace achieves state-of-the-art results. Code and pre-trained models are promised but not yet available.

Significance. If the claims hold, the paper contributes a practical one-step diffusion architecture with multi-reference support, a plausible mechanism for preserving local identity features through the identity embedder, and an evaluation metric that is more robust to affine misalignment than raw landmark L2 distance. The ablations in Table 3 give some evidence that each proposed component helps. However, the central SOTA claim is currently weakened by the dual role of the proposed A-LD metric (used both for reference selection and as a headline identity metric), the absence of error bars or significance tests, and underspecified comparison and implementation details. The paper is likely of interest to the face-restoration community if these issues are resolved, but the quantitative evidence needs strengthening before the broad superiority claim can be accepted.

major comments (4)
  1. [Sec. 3.4, Eq. (6), Table 1] The proposed affine landmark distance (A-LD) is used in two roles: Eq. (6) selects the optimal reference r0 by minimizing A-LD between the LQ input and each candidate reference, and Table 1 reports A-LD as evidence that HonestFace best preserves identity. Because the restored image is conditioned on r0, HonestFace is explicitly steered at inference toward low A-LD, while the compared methods (ASFFNet, DMDNet, PGDiff, FaceMe, MGFR) do not use this selection rule. The paper asserts in Sec. 3.4 that A-LD “reflects landmark geometry in a way that matches human face perception,” but no correlation with human perceptual judgments or with an independent identity embedding is provided. Thus, the identity-consistency advantage in Table 1 is conditional on A-LD being a valid and unbiased metric. I ask the authors to (i) validate A-LD against human ratings and against a held-out face-recognition embedding not used in training or in the loss, and (ii) specify the reference protocol used for each baseline, including the number of references, the selection rule, and whether the same reference set is provided to all methods.
  2. [Tables 1 and 2] No error bars, confidence intervals, or significance tests are reported for any quantitative comparison. Several reported advantages are small in absolute value, for example C-IQA 0.6673 vs. 0.6606 and MUSIQ 71.996 vs. 71.598 on CelebHQRef-Test. Without repeated evaluation runs or per-image statistics, it is unclear whether the improvements are stable or within run-to-run noise. Please report standard deviations over at least three evaluation runs, or bootstrap confidence intervals, together with the number of test images used for each metric.
  3. [Sec. 3.3, Eq. (12)] The masked face alignment loss is one of the paper's core contributions, but the “neutral background B” used to construct xMFA,H and ˆxMFA,H is never defined. B could be a constant gray image, a blurred version of the image, or a learned background, and the LPIPS loss in Eq. (13) depends on this choice. The paper also does not identify the pre-trained heatmap detection model XHM used in Eq. (10). These details are needed to reproduce the method and to interpret the MFA ablation in Table 3b. Please specify B and XHM, and report sensitivity to the choice of B.
  4. [Sec. 4.2, Table 2] The abstract and conclusion claim that HonestFace surpasses existing state-of-the-art methods, but Table 2 shows that CodeFormer achieves higher PSNR than HonestFace on both CelebHQRef-Test (25.49 vs. 25.17) and Reface-Test (25.65 vs. 25.25). The text in Sec. 4.2 carefully limits the pixel-level claim to “other diffusion-based approaches,” which is accurate, but the broader wording in the abstract and conclusion overstates the evidence. Please either qualify the SOTA claim to the metrics and method families for which it is supported, or add the missing comparisons that would support the broader statement.
minor comments (6)
  1. [Sec. 4.1, Eq. (3)] The value of the fixed timestep TL used in Eq. (3) is not reported, although it is a central parameter of the one-step formulation; please state the value and, ideally, provide a sensitivity analysis.
  2. [Sec. 3.4, Eq. (14)] The landmark weights W in Eq. (14) are never specified; please state whether they are uniform or hand-chosen and describe how the weights affect the A-LD values reported in Table 1.
  3. [Sec. 3.3, Eq. (10)] The heatmap detection model XHM is not identified by name or architecture; please provide this information for reproducibility.
  4. [Sec. 4.1] The training details do not state whether the VAE encoder and decoder are frozen during fine-tuning and which parameters of the UNet are trainable beyond the LoRA adapters; please clarify.
  5. [Sec. 1, Contributions] “Identity embedded” appears to be a typo for “identity embedder”; please correct it.
  6. [Table 1, Sec. 4.1] The tables do not report the number of reference images N used for HonestFace and for each baseline; please add this information, since the multi-reference setting is central to the method.

Circularity Check

3 steps flagged · score 6.0 of 10

Several headline metrics double as training objectives or as the reference-selection criterion, so parts of the SOTA claim restate the model's own optimization rather than independent evidence.

  1. fitted input called prediction [Sec. 3.3 (Eq. 8), Sec. 4.1 Metrics, Table 1 (Deg.)]
    "The identity loss is then computed as the sum of cosine distances between the embeddings of the ground truth image xH and the restored image ˆxH: LID = ... . For identity consistency, as it is crucial for honest restoration, we assess ArcFace [8] embedding angular differences “Deg.”"

    ArcFace is one of the identity extractors in the ensemble F used to compute L_ID, so the model is explicitly trained to minimize ArcFace cosine distance between restored and ground-truth faces. The headline evaluation metric Deg. is exactly the ArcFace angular difference. Reporting HonestFace's improvement on Deg. is therefore partially a re-statement of its own training objective rather than independent validation. The other identity metrics (L2-LD, A-LD) are not directly optimized in training, which limits the scope of the circularity.

  2. fitted input called prediction [Sec. 3.3 (Eqs. 9 and 13), Table 1 and Table 2 (LPIPS, DISTS)]
    "To ensure that the overall structural and textural attributes of the restored image ˆxH align with those of the ground truth xH from a human perceptual standpoint, we incorporate a global perceptual loss LPer. We utilize the DISTS metric ... The local perceptual loss, LMFA, is then computed using the LPIPS metric ... LMFA = Llpips(xMFA,H, ˆxMFA,H)."

    The paper reports LPIPS and DISTS as top-row perceptual metrics in Table 1 and Table 2, yet both are used directly as differentiable training losses: DISTS appears in L_Per and LPIPS appears in L_MFA. A model trained to minimize a metric is not an unbiased test of that metric against methods trained under different objectives. The reported gains on LPIPS and DISTS are therefore partly forced by the chosen losses rather than being independent confirmation of perceptual quality, though FID, MUSIQ, MANIQA, and CLIP-IQA remain untrained metrics.

1 more flagged steps
  1. other [Sec. 3.2 Eq. (6), Sec. 3.4 Eq. (14), Table 1 (ALD-e, ALD-m)]
    "The optimal reference image r0 could be identified by r0 = arg min_{ri∈R} d(ri, xL), where d(·, ·) denotes the affine landmark distance, which is formally defined in Sec. 3.4."

    A-LD is used twice: at inference it is the criterion for selecting the reference image r0 in Eq. (6), and in evaluation it is reported as evidence of identity preservation through ALD-e and ALD-m in Table 1. The restoration is conditioned on a reference chosen to minimize this same A-LD function, while the compared methods are not given the same A-LD-based reference selection and their reference protocols are not specified. This creates an evaluator-method coupling rather than an independent test of A-LD or of identity fidelity. Training losses never directly optimize A-LD, so this step is partial rather than fully by construction.

full rationale

The dominant circularity is the overlap between the paper's evaluation metrics and its own training objectives. The identity metric Deg. is the ArcFace cosine distance that L_ID explicitly minimizes; LPIPS and DISTS are likewise used as losses and then reported as headline perceptual metrics. In addition, the proposed A-LD metric is used both to select the reference image at inference and as a primary identity-evaluation metric, making the reported ALD advantage partly contingent on the method's own selection rule. These couplings do not make the paper vacuous: Table 1 also reports untrained metrics such as FID, MUSIQ, MANIQA, and CLIP-IQA, and the qualitative comparisons plus L2-LD provide some independent support. The self-citation to OSDFace is not load-bearing circularity because OSDFace is used as a baseline and as a component source, and the claimed advances are attributable to new modules and are compared against it. Overall, the SOTA claim is partially forced by the authors' choice of losses and the dual role of A-LD, so a score of 6 reflects partial, rather than total, circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a handful of standard assumptions about face recognition embeddings, landmark robustness, and perceptual metrics. There are a few free hyperparameters (k, token counts, LoRA rank) that are chosen without sensitivity analysis, and the paper does not specify the timestep TL, the A-LD weights W, or the MFA background B. No new physical entities are introduced.

free parameters (5)
  • MFA sharpening exponent k = 10
    Controls the contrast of the attention mask in Eq. 11; chosen by hand, no sensitivity analysis.
  • Prompt token counts (n1, n2, n3) = 4, 7, 77
    Capacity of the identity, identity-encoder, and VRE embeddings; chosen empirically without ablation on token counts.
  • LoRA rank and alpha = 16
    Low-rank adaptation hyperparameters for fine-tuning the UNet; standard but arbitrary.
  • Fixed timestep TL = not specified
    The one-step diffusion timestep (Eq. 3) is said to be predetermined but its value is never given in the paper.
  • A-LD landmark weights W = not specified
    Weights in Eq. 14 are defined as positive but no values are provided; likely equal weights but unspecified.
assumptions (6)
  • standard math Affine least-squares has a unique closed-form solution when landmarks are non-collinear
    Used to derive Eq. 15 for the A-LD metric.
  • domain assumption Pre-trained face recognition models (ArcFace, AdaFace) produce embeddings that align with human identity perception
    Used in identity loss (Eq. 8) and identity encoder; if these embeddings are unreliable, identity preservation training is compromised.
  • domain assumption Heatmap-based landmark detection is robust to severe image degradation
    Assumed for MFA (Sec. 3.3) and A-LD reference selection/evaluation (Sec. 3.4); no validation on the degraded inputs used in the paper.
  • domain assumption The VQFR degradation pipeline used to synthesize low-quality training data is representative of real-world degradation
    Training and evaluation are conducted on synthetic degradations; real-world performance may differ.
  • domain assumption DISTS and LPIPS metrics correlate with human perception
    Used as global and local perceptual losses and evaluation; standard in the field.
  • ad hoc to paper The fixed-timestep one-step diffusion formulation from OSDFace can be fine-tuned without losing restoration capacity
    HonestFace inherits this formulation (Eq. 3) from the baseline without re-deriving or validating it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HonestFace: Towards Honest Face Restoration with One-Step Diffusion Model." pith.science (2026). https://pith.science/paper/DV43WZLH

@misc{pith2026250518469,
  author       = {Pith},
  title        = {Pith review of: HonestFace: Towards Honest Face Restoration with One-Step Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DV43WZLH}},
  note         = {Machine review of arXiv:2505.18469}
}
read the original abstract

Face restoration has achieved remarkable advancements through the years of development. However, ensuring that restored facial images exhibit high fidelity, preserve authentic features, and avoid introducing artifacts or biases remains a significant challenge. This highlights the need for models that are more "honest" in their reconstruction from low-quality inputs, accurately reflecting original characteristics. In this work, we propose HonestFace, a novel approach designed to restore faces with a strong emphasis on such honesty, particularly concerning identity consistency and texture realism. To achieve this, HonestFace incorporates several key components. First, we propose an identity embedder to effectively capture and preserve crucial identity features from both the low-quality input and multiple reference faces. Second, a masked face alignment method is presented to enhance fine-grained details and textural authenticity, thereby preventing the generation of patterned or overly synthetic textures and improving overall clarity. Furthermore, we present a new landmark-based evaluation metric. Based on affine transformation principles, this metric improves the accuracy compared to conventional L2 distance calculations for facial feature alignment. Leveraging these contributions within a one-step diffusion model framework, HonestFace delivers exceptional restoration results in terms of facial fidelity and realism. Extensive experiments demonstrate that our approach surpasses existing state-of-the-art methods, achieving superior performance in both visual quality and quantitative assessments. The code and pre-trained models will be made publicly available at https://github.com/jkwang28/HonestFace .

Figures

Figures reproduced from arXiv: 2505.18469 by the authors.

Figure 1
Figure 1. Overall training pipeline of HonestFace. First, the LQ input [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Facial feature extractor. Facial Feature Extractor. Key facial characteristics, such as the eyes, nose, and mouth, are important for human identity perception. Variations in eye color, skin texture, and edge structure significantly affect an individual’s distinct appearance. Given that the input LQ image often suffers from considerable degradation, extracting such detailed identity solely from xL is unreliable. Cons… view at source ↗
Figure 3
Figure 3. Illustration of face identity encoder. Face Identity Encoder. Extracting robust global identity information is a central challenge in face recognition. Many face recognition models work hard to make embeddings from the same identity clustered closely, while keeping embeddings from different identities apart. We process the set of aligned reference images R = {r1, r2, . . . , rN } using such a pre-trained face recogn… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Masked face alignment. Local Perceptual Loss via Masked Face Alignment. Human attention focuses mainly on distinctive facial details, such as moles, scars, or wrinkles near the eyes and mouth. While other areas, like the cheeks or hair, may only need natural-looking te…
Figure 5
Figure 5. Figure 5: Visual comparison of CelebHQRef-Test. Please zoom in for a better view. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Landmark comparisons for Reface-Test 159. Landmark for ground truth is colored by green. Although FaceMe has a lower L2- LD for the right eyebrow, Honest￾Face achieves better fidelity, with a smaller A-LD calculated on eyes. Let the landmark of the restored image set b…
Figure 7
Figure 7. Figure 7: Visual comparison of Reface-Test. Please zoom in for a better view. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 72 canonical work pages

  1. [1]

    Super-fan: Integrated facial landmark localization and super- resolution of real-world low resolution faces in arbitrary poses with GANs

    Adrian Bulat and Georgios Tzimiropoulos. Super-fan: Integrated facial landmark localization and super- resolution of real-world low resolution faces in arbitrary poses with GANs. In CVPR, 2018. 6

    work page 2018

  2. [2]

    GLEAN: Generative latent bank for large-factor image super-resolution

    Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. GLEAN: Generative latent bank for large-factor image super-resolution. In CVPR, 2021. 1

    work page 2021

  3. [3]

    Chaofeng Chen, Xiaoming Li, Yang Lingbo, Xianhui Lin, Lei Zhang, and Kwan-Yee K. Wong. Progressive semantic-aware style transformation for blind face restoration. In CVPR, 2021. 1

    work page 2021

  4. [4]

    Pixart-$\alpha$: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-$\alpha$: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 3

    work page 2024

  5. [5]

    Towards real-world blind face restoration with generative diffusion prior

    Xiaoxu Chen, Jingfan Tan, Tao Wang, Kaihao Zhang, Wenhan Luo, and Xiaocun Cao. Towards real-world blind face restoration with generative diffusion prior. arXiv preprint arXiv:2312.15736, 2023. 1, 6

    arXiv 2023

  6. [6]

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

    Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. FSRNet: End-to-end learning face super-resolution with facial priors. In CVPR, 2018. 2

    work page 2018

  7. [7]

    Image super-resolution with text prompt diffusion

    Zheng Chen, Yulun Zhang, Jinjin Gu, Xin Yuan, Linghe Kong, Guihai Chen, and Xiaokang Yang. Image super-resolution with text prompt diffusion. arXiv preprint arXiv:2303.06373, 2023. 4

    arXiv 2023

  8. [8]

    ArcFace: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5, 8

    work page 2019

Show all 77 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 TPAMI, 2020. 5, 8

    2020

  2. [10]

    Restoration by generation with constrained priors

    Zheng Ding, Xuaner Zhang, Zhuowen Tu, and Zhihao Xia. Restoration by generation with constrained priors. In CVPR, 2024. 3

    2024

  3. [11]

    Exemplar guided face image super-resolution without facial landmarks

    Berk Dogan, Shuhang Gu, and Radu Timofte. Exemplar guided face image super-resolution without facial landmarks. In CVPRW, 2019. 2

    2019

  4. [12]

    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 CVPR, 2025. 3

    2025

  5. [13]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In NeurIPS, 2015. 5

    2015

  6. [14]

    Image processing using multi-code gan prior

    Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code gan prior. In CVPR, 2020. 2

    2020

  7. [15]

    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 ECCV, 2022. 6, 7

    2022

  8. [16]

    Multiscale sliced Wasserstein distances as perceptual color difference measures

    Jiaqi He, Zhihua Wang, Leon Wang, Tsein-I Liu, Yuming Fang, Qilin Sun, and Kede Ma. Multiscale sliced Wasserstein distances as perceptual color difference measures. In ECCV, 2024. 9

    2024

  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, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 8

    2022

  10. [18]

    MUSIQ: Multi-scale Image Quality Transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale Image Quality Transformer . In ICCV, 2021. 8

    2021

  11. [19]

    Progressive face super-resolution via attention to facial landmark

    Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. In BMVC, 2019. 2

    2019

  12. [20]

    AdaFace: Quality adaptive margin for face recognition

    Minchul Kim, Anil K Jain, and Xiaoming Liu. AdaFace: Quality adaptive margin for face recognition. In CVPR, 2022. 5, 8

    2022

  13. [21]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 3

    2024

  14. [22]

    Revisting quantization error in face alignment

    Xing Lan, Qinghao Hu, and Jian Cheng. Revisting quantization error in face alignment. In ICCVW, 2021. 6

    2021

  15. [23]

    Unleashing the power of one-step diffusion based image super-resolution via a large-scale diffusion discriminator

    Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. Unleashing the power of one-step diffusion based image super-resolution via a large-scale diffusion discriminator. arXiv preprint arXiv:2410.04224, 2024. 4, 6 10

    arXiv

  16. [24]

    One diffusion step to real-world super-resolution via flow trajectory distillation

    Jianze Li, Jiezhang Cao, Yong Guo, Wenbo Li, and Yulun Zhang. One diffusion step to real-world super-resolution via flow trajectory distillation. In ICML, 2025. 3

    2025

  17. [26]

    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 ECCV, 2018. 2

    2018

  18. [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 ECCV, 2020. 2

    2020

  19. [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. In CVPR, 2020. 2, 8

    2020

  20. [29]

    Learning dual memory dictionaries for blind face restoration

    Xiaoming Li, Shiguang Zhang, Shangchen Zhou, Lei Zhang, and Wangmeng Zuo. Learning dual memory dictionaries for blind face restoration. IEEE TPAMI, 2022. 3, 7, 8

    2022

  21. [30]

    DiffBIR: Towards blind image restoration with generative diffusion prior

    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. In ECCV, 2024. 1

    2024

  22. [31]

    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. In AAAI, 2025. 2, 3, 7, 8, 9

    2025

  23. [32]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 8

    2019

  24. [33]

    Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models

    Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, 2023. 4

    2023

  25. [34]

    Understanding deep image representations by inverting them

    Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them . In CVPR, 2015. 5

    2015

  26. [35]

    Contour-hugging heatmaps for landmark detection

    James McCouat and Irina V oiculescu. Contour-hugging heatmaps for landmark detection. In CVPR, 2022. 6

    2022

  27. [36]

    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 CVPR, 2020. 2

    2020

  28. [37]

    WaveFace: Authentic face restoration with efficient frequency recovery

    Yunqi Miao, Jiankang Deng, and Jungong Han. WaveFace: Authentic face restoration with efficient frequency recovery. In CVPR, 2024. 1

    2024

  29. [38]

    Mystyle: A personalized generative prior

    Yotam Nitzan, Kfir Aberman, Qiurui He, Orly Liba, Michal Yarom, Yossi Gandelsman, Inbar Mosseri, Yael Pritch, and Daniel Cohen-Or. Mystyle: A personalized generative prior. ACM TOG, 2022. 3

    2022

  30. [39]

    Regressing heatmaps for multiple landmark localization using CNNs

    Christian Payer, Darko Štern, Horst Bischof, and Martin Urschler. Regressing heatmaps for multiple landmark localization using CNNs. In MICCAI, 2016. 6

    2016

  31. [40]

    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 ACM MM, 2023. 1

    2023

  32. [41]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3

    2021

  33. [42]

    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 CVPR, 2022. 3

    2022

  34. [43]

    Fast high-resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. In ACM SIGGRAPH Asia,

  35. [44]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 4

    2024

  36. [45]

    Deep semantic face deblurring

    Ziyi Shen, Wei-Sheng Lai, Tingfa Xu, Jan Kautz, and Ming-Hsuan Yang. Deep semantic face deblurring. In CVPR, 2018. 2

    2018

  37. [46]

    Very deep convolutional networks for large-scale image recognition

    K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5 11

    2015

  38. [47]

    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 IJCAI, 2024. 1

    2024

  39. [48]

    Deep high-resolution representation learning for human pose estimation

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 6

    2019

  40. [49]

    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 CVPR, 2025. 3, 4

    2025

  41. [50]

    Lafs: Landmark-based facial self-supervised learning for face recognition

    Zhonglin Sun, Chen Feng, Ioannis Patras, and Georgios Tzimiropoulos. Lafs: Landmark-based facial self-supervised learning for face recognition. In CVPR, 2024. 6

    2024

  42. [51]

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

    Keda Tao, Jinjin Gu, Yulun Zhang, Xiucheng Wang, and Nan Cheng. Overcoming false illusions in real-world face restoration with multi-modal guided diffusion model. In ICLR, 2025. 2, 3, 4, 7, 8, 9

    2025

  43. [52]

    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 ICLR, 2024. 1, 2, 6, 8, 9

    2024

  44. [53]

    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 CVPR, 2024. 2, 3

    2024

  45. [54]

    Bringing old photos back to life

    Ziyu Wan, Bo Zhang, Dongdong Chen, Pan Zhang, Dong Chen, Jing Liao, and Fang Wen. Bringing old photos back to life. In CVPR, 2020. 2

    2020

  46. [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 AAAI, 2023. 8

    2023

  47. [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. IJCV, 2024. 3

    2024

  48. [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 CVPR, 2025. 1, 3, 4, 5, 6, 7, 8, 9

    2025

  49. [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 CVPR, 2021. 1, 2

    2021

  50. [59]

    Adaptive wing loss for robust face alignment via heatmap regression

    Xinyao Wang, Liefeng Bo, and Li Fuxin. Adaptive wing loss for robust face alignment via heatmap regression. In ICCV, 2019. 6

    2019

  51. [60]

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 4

    2023

  52. [61]

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

    Zhixin Wang, Xiaoyun Zhang, Ziying Zhang, Huangjie Zheng, Mingyuan Zhou, Ya Zhang, and Yanfeng Wang. Dr2: Diffusion-based robust degradation remover for blind face restoration. In CVPR, 2023. 1

    2023

  53. [62]

    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 TPAMI, 2023. 1

    2023

  54. [63]

    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 NeurIPS, 2024. 1, 2, 4, 8, 9

    2024

  55. [64]

    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 CVPR, 2024. 3, 4

    2024

  56. [65]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformer

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformer. arXiv preprint arXiv:2410.10629, 2024. 3

    arXiv

  57. [66]

    Learning degradation-unaware representation with prior-based latent transformations for blind face restoration

    Lianxin Xie, Csbingbing Zheng, Wen Xue, Le Jiang, Cheng Liu, Si Wu, and Hau San Wong. Learning degradation-unaware representation with prior-based latent transformations for blind face restoration. In CVPR, 2024. 1

    2024

  58. [67]

    Hifacegan: Face renovation via collaborative suppression and replenishment

    Lingbo Yang, Shanshe Wang, Siwei Ma, Wen Gao, Chang Liu, Pan Wang, and Peiran Ren. Hifacegan: Face renovation via collaborative suppression and replenishment. In ACM MM, 2020. 2

    2020

  59. [68]

    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 NeurIPS, 2023. 3, 7, 8, 9 12

    2023

  60. [69]

    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 CVPRW, 2022. 8

    2022

  61. [70]

    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 CVPR, 2021. 1

    2021

  62. [71]

    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 CVPR, 2024. 3

    2024

  63. [72]

    Super-resolving very low-resolution face images with supplementary attributes

    Xin Yu, Basura Fernando, Richard Hartley, and Fatih Porikli. Super-resolving very low-resolution face images with supplementary attributes. In CVPR, 2018. 2

    2018

  64. [73]

    DifFace: Blind Face Restoration with Diffused Error Contraction

    Zongsheng Yue and Chen Change Loy. DifFace: Blind Face Restoration with Diffused Error Contraction . IEEE TPAMI, 2024. 1, 6

    2024

  65. [74]

    Mystyle++: A controllable personalized generative prior

    Libing Zeng, Lele Chen, Yi Xu, and Nima Khademi Kalantari. Mystyle++: A controllable personalized generative prior. In ACM SIGGRAPH Asia, 2023. 3

    2023

  66. [75]

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

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

    arXiv

  67. [76]

    Joint face detection and alignment using multitask cascaded convolutional networks

    Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 2016. 6

    2016

  68. [77]

    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. In CVPR, 2018. 6, 8

    2018

  69. [78]

    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. In NeurIPS, 2022. 1, 2, 8, 9 13

    2022

Pith tools

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