Pith. sign in

REVIEW 2 major objections 4 minor 123 references

Robust Face Super-Resolution and Recognition Through Multi-Feature Aggregation in Diffusion Models

T0 review · 2 major / 4 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Fusing features from multiple low-resolution face images into a diffusion model yields super-resolved faces that preserve identity and raise recognition accuracy without soft attributes or gradient guidance.

desk verdict Solid incremental multi-frame face SR that actually moves recognition and quality metrics without soft attributes or gradient guidance. read the letter →

arxiv 2607.05702 v1 pith:TT5W22UG submitted 2026-07-06 cs.CV

Marcelo dos Santos , Rayson Laroca , João Carlos Raposo Neves , David Menotti This is my paper
classification cs.CV
keywords facesuper-resolutiondiffusionmodelsfeatureaggregationidentitypreservationrecognitionsurveillancescore-basedgenerative
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

Surveillance cameras often deliver faces that are too low-resolution and degraded for reliable recognition. Most super-resolution methods fill in missing detail but distort the person’s identity. This paper shows that a diffusion model can avoid that distortion when it is conditioned on both a single reference low-resolution image and one compact feature vector obtained by combining embeddings from several other low-resolution images of the same person. The combination is performed by a small neural network called a Feature Combiner that recovers high-frequency identity cues simple averaging misses. The resulting super-resolved images look more natural and simultaneously improve both standard image-quality scores and face-verification and identification rates on public benchmarks, including real surveillance footage. The method therefore supplies a practical pre-processing step that turns multi-frame low-quality video into higher-quality probes for recognition systems.

What carries the argument

The Feature Combiner ensemble: pairs of 512-dimensional face embeddings are averaged and refined by a small MLP, then the refined vectors are averaged across all pairs to form a single higher-signal feature that is injected, together with the reference low-resolution image and the diffusion timestep, into every residual block of the reverse SDE.

What would settle it

On a held-out multi-frame surveillance sequence with known high-resolution ground truth, replace the Feature Combiner ensemble by plain averaging or reduce the number of auxiliary frames to one and check whether Rank-1 accuracy and LPIPS of the super-resolved output fall below the previous second-best baseline; if the advantage disappears, the multi-feature claim fails.

Watch

Extended reading notes

Core claim

FASR++ conditions a score-based diffusion model on a reference low-resolution face together with a merged identity descriptor produced by an ensemble of Feature Combiner modules applied to multiple auxiliary low-resolution images of the same person; the generated super-resolved faces achieve state-of-the-art verification AUC, Rank-1 and Rank-5 accuracy, PSNR, SSIM and LPIPS on CelebA and Quis-Campi without requiring explicit soft-attribute labels or reverse-process gradient guidance.

Load-bearing premise

Features taken from several low-resolution images of one person, once fused by the trained combiner, still carry enough clean identity signal for the diffusion model to reconstruct high-frequency facial detail that generalizes to real surveillance footage.

Editorial extensions

If this is right

  • Multi-frame surveillance video can be turned into identity-preserving probes that raise Rank-1 and Rank-5 recognition rates.
  • Soft-attribute classifiers and gradient-based guidance become unnecessary for diffusion face super-resolution.
  • Image-quality metrics and recognition metrics can be improved together rather than traded off.
  • Training the combiner on larger, more diverse face collections further lifts Rank-1 accuracy.

Reading between the lines

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

  • The same multi-observation feature aggregation could be applied to other identity-preserving inverse problems such as deblurring or inpainting when several degraded views of one person are available.
  • Because absolute cosine similarity after fusion remains modest, better visual encoders or uncertainty-weighted fusion may still yield further gains.
  • Running feature averages could be cached and updated online so that super-resolution keeps pace with a live video stream.
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

2 major / 4 minor

Summary. The paper introduces FASR++, a diffusion-model face super-resolution method (NCSN++ backbone, VE-SDE) that conditions reverse sampling on a reference LR image (channel-concatenated) plus a merged 512-d AdaFace feature vector F_M. F_M is produced by an ensemble of Feature Combiner modules, each of which fuses a pair of LR embeddings via their arithmetic mean plus a learned residual network δ (trained by triplet loss on CASIA-WebFace). Two independent SR samples are averaged to form the final output. The method is claimed to recover identity-consistent high-frequency detail without explicit soft attributes or gradient guidance at inference, yielding state-of-the-art AUC, Rank-1/Rank-5, PSNR, SSIM and LPIPS on CelebA (500 identities) and Quis-Campi (90 identities) relative to GFPGAN, SPARNET, SwinIR, SR3, SDE-SR, IDM, SRDG and the authors’ prior FASR.

Significance. If the reported gains hold, FASR++ supplies a practical multi-frame conditioning strategy that improves both perceptual quality and downstream recognition utility for surveillance face SR. Concrete strengths include publicly released code, systematic ablations on the residual weight η, the size of the δ training set, and dual-sample averaging, paired t-tests (p < 10^{-5}), CMC curves, and an explicit failure-case discussion. Avoiding attribute classifiers and classifier-gradient guidance at test time is a genuine practical advantage. The advance is incremental over the authors’ FASR/SDE-SR line, yet the multi-feature aggregation mechanism and the empirical package are solid and reproducible.

major comments (2)
  1. [Section 4.2, Figure 5, Eq. 5] Section 4.2 and Figure 5: after FC fusion the mean cosine similarity to the gallery embedding rises only from 0.162 to 0.350. While the relative gain is statistically significant and the ablations (Tables 3–4) show that enabling δ (η = 1) improves Rank-1/PSNR/SSIM/LPIPS, the absolute value remains modest. Because F_M is the sole high-level identity signal injected into the reverse SDE (Eq. 5), the manuscript should quantify how much of the observed recognition and LPIPS gains can be attributed to this signal (e.g., by controlled noise injection into F_M or by correlating fused similarity with per-identity Rank-1).
  2. [Section 4.1, Tables 1–2] Section 4.1 and Tables 1–2: Quis-Campi probes are further 8 imes down-sampled even though the source frames are already low-quality surveillance imagery, and only five auxiliary images are used. The central claim of robustness under real surveillance conditions would be more convincing with an explicit ablation on the number of auxiliaries N (already partially present for CelebA) and, if feasible, a native-resolution probe experiment without the additional synthetic down-sampling step.
minor comments (4)
  1. [Throughout] Several concatenated words appear in the supplied text (e.g., “variationsinillumination”, “posesachallenge”). These are almost certainly PDF-extraction artefacts, but the camera-ready version should be checked for residual spacing or hyphenation errors.
  2. [Figure 1, Section 3.1] Figure 1 caption and Section 3.1: the dual-sample averaging step (SR_a, SR_b → SR) is described only briefly; a short sentence on the observed variance reduction would help readers understand why two samples are sufficient.
  3. [Section 3.5] Section 3.5: the precise channel dimension C after the linear projections of the time and feature embeddings is never stated; adding the value (or noting that it matches the residual-block width) would improve reproducibility.
  4. [Table 2] Table 2: SRDG is listed as a competitor that uses soft attributes; a one-sentence reminder that those attributes are oracle-provided (and therefore unavailable in true LR surveillance) would make the comparison fairer.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: empirical SR method with external baselines, standard metrics, and independent ablations; self-citations to prior FASR/SDE-SR are incremental, not load-bearing.

full rationale

The paper presents an engineering contribution (FASR++): a Feature Combiner ensemble (Eq. 1-3, trained with triplet loss on CASIA-WebFace) that produces a merged AdaFace vector F_M, which conditions an NCSN++ reverse SDE (Eq. 5-6, trained on FFHQ) together with a reference LR image. Evaluation uses held-out CelebA/Quis-Campi identities, standard external metrics (AdaFace cosine for Rank-1/5/AUC; PSNR/SSIM/LPIPS), and independent SOTA baselines (GFPGAN, SPARNET, SwinIR, SR3, IDM) plus the authors' own prior methods. Ablations (Tables 3-4, Fig. 8-9) isolate the contribution of δ (η=1) and dual-sample averaging without tautology. No equation equates a claimed prediction to a fitted input by construction; no uniqueness theorem or ansatz is imported from self-citations to force the result; recognition gains are measured downstream on generated images, not on the intermediate fused features themselves. Self-citations (to FASR 2024a, SDE-SR 2022, SRDG 2024b) are normal for an incremental paper and are not required for the SOTA claim against external methods. Absolute cosine similarity of fused LR features remains modest (0.35), but this is quantified and does not create circularity. The derivation chain is therefore self-contained and non-circular.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central empirical claim rests on standard diffusion SDE machinery, a pre-trained AdaFace encoder, and a handful of hand-chosen hyperparameters; the only invented component is the Feature Combiner itself, whose utility is measured by cosine-similarity and downstream recognition gains rather than by independent physical evidence.

free parameters (5)
  • σ_min / σ_max (VE noise schedule) = 0.001 / 348
    Set to 0.001 and 348; control the noise range of the reverse SDE and therefore the quality of generated samples.
  • triplet-loss margin for δ network = 0.495
    Empirically set to 0.495 after observing collapse for larger margins; directly affects how well the Feature Combiner separates identities.
  • η (Feature Combiner residual weight) = 1
    Binary switch (0 or 1) controlling whether the learned residual δ is added; final models use η=1.
  • number of reverse SDE steps = 2000
    Fixed at 2000 Euler-Maruyama steps; trades quality against compute.
  • dropout rates (0.2 in δ, 0.1 in residual blocks) = 0.2 / 0.1
    Chosen by validation; affect generalization of the combiner and the U-Net.
assumptions (4)
  • domain assumption Score-based generative modeling via reverse-time SDE (Song et al. 2021) correctly approximates the data score when trained with the denoising objective (Eq. 6).
    Invoked throughout Section 3.3; the entire generation procedure rests on this unproved but standard assumption.
  • domain assumption AdaFace (ResNet backbone trained on CASIA-WebFace) produces 512-d embeddings whose cosine similarity is a reliable identity metric even for low-resolution inputs.
    Used both for training the Feature Combiner (triplet loss) and for all recognition metrics; Section 4.1.
  • domain assumption Multiple low-resolution frames of the same identity are available at inference time and share a common identity signal that averaging can amplify.
    Core premise of the multi-feature pipeline (Figure 1, Section 3.1); true for video surveillance but not for single-shot capture.
  • ad hoc to paper Arithmetic mean of paired features plus a learned residual (Eq. 1) is a sufficient aggregator; other pooling operators are inferior.
    Justified only by the authors' experimental comparison (Section 3.2); no theoretical guarantee.
invented entities (2)
  • Feature Combiner (FC) module with residual network δ
    purpose: Merge two LR feature vectors into a higher-SNR representation that approximates the mean of the corresponding HR features.
    Introduced in Section 3.2 and Figure 2; its utility is demonstrated only inside this paper via cosine-similarity histograms and downstream recognition gains; no external falsifiable prediction.
  • FASR++ pipeline (ensemble of FCs + dual-sample averaging + multi-feature conditioning of NCSN++)
    purpose: Produce identity-preserving super-resolved faces from one LR reference plus auxiliary LR frames.
    The overall system is the paper's contribution; evaluated solely on the two chosen datasets against selected baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Face Super-Resolution and Recognition Through Multi-Feature Aggregation in Diffusion Models." pith.science (2026). https://pith.science/paper/TT5W22UG

@misc{pith2026260705702,
  author       = {Pith},
  title        = {Pith review of: Robust Face Super-Resolution and Recognition Through Multi-Feature Aggregation in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TT5W22UG}},
  note         = {Machine review of arXiv:2607.05702}
}
read the original abstract

Images acquired in surveillance environments often suffer from conditions such as low resolution, variations in pose, irregular illumination, and occlusions. Due to the low quality of these images, face recognition algorithms often struggle. This major limitation can be addressed by employing super-resolution techniques that enhance the details of the image. However, due to the high degree of difficulty of the problem, most super-resolution algorithms tend to cause distortions in the image and in the individual's identity. Thus, additional information must be incorporated into the processing to improve recognition robustness. In this regard, surveillance cameras can capture multiple images, even at low quality, and the data extracted from these images, such as consecutive video frames, can significantly enhance both super-resolution and facial recognition. In this work, we introduce FASR++, a diffusion-model-based super-resolution algorithm. It leverages a reference low-resolution image and features extracted from multiple auxiliary low-quality images to generate a super-resolved output, minimizing distortions in the individual's identity. Our approach recovers facial features without explicitly providing soft attributes or computing a function gradient to guide the reconstruction process. FASR++ generates high-quality images that can considerably improve performance in face recognition tasks when used as a pre-processing step. We validate our approach on two standard face recognition datasets and attain state-of-the-art results for verification, face recognition, and image quality metrics such as PSNR, SSIM, and LPIPS.

Figures

Figures reproduced from arXiv: 2607.05702 by the authors.

Figure 1
Figure 1. Overview of the proposed method. At inference time, a set of 𝑁+1 low-resolution images is collected from an individual. The images LR1, . . . , LR𝑁 are processed to extract feature representations F1, . . . , F𝑁 , which are then aggregated through an ensemble of Feature Combiner (FC) modules to produce the merged feature vector F𝑀. The reference image LR0 is jointly integrated with F𝑀 into the diffusion model to gen… view at source ↗
Figure 2
Figure 2. The FC module consists of the mean of the input features and a refinement network 𝛿. The mean operation was adopted as a baseline because the input low-resolution images, and consequently their extracted features, often contain noise Input features F𝑖, F𝑖+1 ∈ R 512 Concatenate [F𝑖, F𝑖+1] ∈ R 1024 Linear(1024→1024) + BN + ReLU + Dropout(0.2) Linear(1024→512) + BN + ReLU + Dropout(0.2) Linear(512→512) 𝜂 · 𝛿( [F𝑖, F𝑖+1… view at source ↗
Figure 3
Figure 3. Time and features encoding. Illustration of the conditioning mechanism in the proposed diffusion model. The first block represents the image feature maps of shape (𝐵, 𝐶, 𝐻, 𝑊), where 𝐵 is the batch size. In this example, we illustrate a single element from the batch for clarity. The time encoding and feature encoding tensors are broadcast along the spatial dimensions so that their values remain constant across 𝐻 and… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Efficacy of FASR++ for face reconstruction. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Histogram of the similarity score distributions for the two [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results for the CelebA Dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Qualitative Results for the Quis-Campi Dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: CMC curves on CelebA (left) and Quis-Campi (right) datasets. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 10
Figure 10. Figure 10: Failure cases. The first row presents results from SRDG [dos Santos et al., 2024b], FASR++ (ours), and ground truth (GT) images, while the second row presents results from SDE-SR [dos Santos et al., 2022], FASR++ (ours), and GT images. Competing interests The authors …
Figure 9
Figure 9. Figure 9: Performance metrics of FASR++ as a function of the number of training images. The top plot illustrates the image quality metrics (PSNR and SSIM), whereas the bottom plot presents the recognition metrics (Rank-1 and Rank-5 accuracies). applications in real-world surveil…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

123 extracted references · 123 canonical work pages

  1. [1]

    2006 , publisher=

    Writing for science , author=. 2006 , publisher=

    work page 2006

  2. [2]

    2001 , publisher=

    Guide to writing empirical papers, theses, and dissertations , author=. 2001 , publisher=

    work page 2001

  3. [3]

    Beall, Jeffrey , journal=. What. 2017 , publisher=

    work page 2017

  4. [4]

    The state of

    Piwowar, Heather and Priem, Jason and Larivi. The state of. PeerJ , volume=. 2018 , publisher=

    work page 2018

  5. [5]

    Contemporary empirical methods in software engineering , pages=

    Open science in software engineering , author=. Contemporary empirical methods in software engineering , pages=. 2020 , publisher=

    work page 2020

  6. [6]

    Proceedings - International Conference on Pattern Recognition , keywords =

    Ahmed, Nasir Udin and Cvetkovic, Slobodan and Siddiqi, Erfanul Hoque and Nikiforov, Andrey and Nikiforov, Ilia , doi =. Proceedings - International Conference on Pattern Recognition , keywords =

    work page

  7. [7]

    IEEE Transactions on Information Forensics and Security , keywords =

    Proen. IEEE Transactions on Information Forensics and Security , keywords =. doi:10.1109/TIFS.2017.2771230 , issn =

    work page doi:10.1109/tifs.2017.2771230 2017

  8. [8]

    Periocular Recognition Using CNN Features Off-the-Shelf , year=

    Hernandez-Diaz, Kevin and Alonso-Fernandez, Fernando and Bigun, Josef , booktitle=. Periocular Recognition Using CNN Features Off-the-Shelf , year=

    work page

Show all 123 references
  1. [9]

    Unconstrained periocular biometric acquisition and recognition using COTS PTZ camera for uncooperative and non-cooperative subjects , year=

    Juefei-Xu, Felix and Savvides, Marios , booktitle=. Unconstrained periocular biometric acquisition and recognition using COTS PTZ camera for uncooperative and non-cooperative subjects , year=

  2. [10]

    2016 , note =

    A survey on periocular biometrics research , journal =. 2016 , note =. doi:10.1016/j.patrec.2015.08.026 , author =

    2016 doi

  3. [11]

    Best Regions for Periocular Recognition with NIR and Visible Images , journal =

    Alonso-Fernandez, Fernando and Bigun, Josef , year =. Best Regions for Periocular Recognition with NIR and Visible Images , journal =

  4. [12]

    and Santos, Gil , doi =

    Proenca, Hugo and Neves, Joao C. and Santos, Gil , doi =. IJCB 2014 - 2014 IEEE/IAPR International Joint Conference on Biometrics , title =

    2014

  5. [13]

    Noisy iris image matching by using multiple cues , volume =

    Tan, Tieniu and Zhang, Xiaobo and Sun, Zhenan and Zhang, Hui , year =. Noisy iris image matching by using multiple cues , volume =. Pattern Recognition Letters - PRL , doi =

  6. [14]

    Pattern Recognition Letters , keywords =

    Wang, Qi and Zhang, Xiangde and Li, Mingqi and Dong, Xiaopeng and Zhou, Qunhua and Yin, Yu , doi =. Pattern Recognition Letters , keywords =. doi:10.1016/j.patrec.2011.08.014 , volume =

    2011 doi

  7. [15]

    Deep Periocular Representation Aiming Video Surveillance , volume =

    Luz, Eduardo and Moreira, Gladston and Zanlorensi, Luiz and Menotti, David , year =. Deep Periocular Representation Aiming Video Surveillance , volume =. Pattern Recognition Letters , doi =

  8. [16]

    and Luz, Eduardo and Zanlorensi, Luiz A

    Silva, Pedro H. and Luz, Eduardo and Zanlorensi, Luiz A. and Menotti, David and Moreira, Gladston , booktitle=. Multimodal Feature Level Fusion based on Particle Swarm Optimization with Deep Transfer Learning , year=

  9. [17]

    and Raghavendra, R

    Raja, Kiran B. and Raghavendra, R. and Venkatesh, Sushma and Busch, Christoph , doi =. Pattern Recognition Letters , keywords =. doi:10.1016/j.patrec.2016.12.025 , volume =

    2016 doi

  10. [18]

    The BTAS Competition on Mobile Iris Recognition , year=

    Zhang, Man and Zhang, Qi and Sun, Zhenan and Zhou, Shujuan and Ahmed, Nasir Uddin , booktitle=. The BTAS Competition on Mobile Iris Recognition , year=

  11. [19]

    and Raghavendra, R

    Sequeira, Ana and Chen, Lulu and Wild, Peter and Ferryman, James and Alonso-Fernandez, Fernando and Raja, Kiran B. and Raghavendra, R. and Busch, Christoph and Bigun, Joseph , booktitle=. Cross-Eyed - Cross-Spectral Iris/Periocular Recognition Database and Competition , year=

  12. [20]

    and Chen, Lulu and Ferryman, James and Wild, Peter and Alonso-Fernandez, Fernando and Bigun, Josef and Raja, Kiran B

    Sequeira, Ana F. and Chen, Lulu and Ferryman, James and Wild, Peter and Alonso-Fernandez, Fernando and Bigun, Josef and Raja, Kiran B. and Raghavendra, R. and Busch, Christoph and de Freitas Pereira, Tiago and Marcel, Sebastien and Behera, Sushree Sangeeta and Gour, Mahesh and...

    2017

  13. [21]

    Pattern Recognit

    DeepIris: Learning pairwise filter bank for heterogeneous iris verification , author=. Pattern Recognit. Lett. , year=

  14. [22]

    DeepIrisNet: Deep iris representation with applications in iris recognition and cross-sensor iris recognition , year=

    Gangwar, Abhishek and Joshi, Akanksha , booktitle=. DeepIrisNet: Deep iris representation with applications in iris recognition and cross-sensor iris recognition , year=

  15. [23]

    Iris Recognition With Off-the-Shelf CNN Features: A Deep Learning Perspective , year=

    Nguyen, Kien and Fookes, Clinton and Ross, Arun and Sridharan, Sridha , journal=. Iris Recognition With Off-the-Shelf CNN Features: A Deep Learning Perspective , year=

  16. [24]

    A multi-biometric iris recognition system based on a deep learning approach , volume =

    Al-Waisy, Alaa and Qahwaji, Rami and Ipson, Stanley and Al-Fahdawi, Shumoos and Nagem, Tarek , year =. A multi-biometric iris recognition system based on a deep learning approach , volume =. Pattern Analysis and Applications , doi =

  17. [25]

    Convolutional Neural Network-Based Periocular Recognition in Surveillance Environments , year=

    Kim, Min Cheol and Koo, Ja Hyung and Cho, Se Woon and Baek, Na Rae and Park, Kang Ryoung , journal=. Convolutional Neural Network-Based Periocular Recognition in Surveillance Environments , year=

  18. [26]

    Sparse Feature Extraction for Pose-Tolerant Face Recognition , year=

    Abiantun, Ramzi and Prabhu, Utsav and Savvides, Marios , journal=. Sparse Feature Extraction for Pose-Tolerant Face Recognition , year=

  19. [27]

    and Kakadiaris, Ioannis A

    Bilgazyev, Emil and Efraty, Boris and Shah, Shishir K. and Kakadiaris, Ioannis A. , booktitle=. Improved face recognition using super-resolution , year=

  20. [28]

    and Ma, Yi , journal=

    Yang, Jianchao and Wright, John and Huang, Thomas S. and Ma, Yi , journal=. Image Super-Resolution Via Sparse Representation , year=

  21. [29]

    2019 , issn =

    Pattern Recognition , volume =. 2019 , issn =. doi:10.1016/j.patcog.2019.01.032 , author =

    2019 doi

  22. [30]

    2016 , booktitle =

    Perceptual Losses for Real-Time Style Transfer and Super-Resolution , author =. 2016 , booktitle =. doi:10.1007/978-3-319-46475-6\_43 , isbn =

    2016 doi

  23. [31]

    Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network , booktitle =

    Ledig, Christian and Theis, Lucas and Huszar, Ferenc and Caballero, Jose and Cunningham, Andrew and Acosta, Alejandro and Aitken, Andrew and Tejani, Alykhan and Totz, Johannes and Wang, Zehan and Shi, Wenzhe , year =. Photo-Realistic Single Image Super-Resolution Using a Gener...

  24. [32]

    ODE-Inspired Network Design for Single Image Super-Resolution , year=

    He, Xiangyu and Mo, Zitao and Wang, Peisong and Liu, Yang and Yang, Mingyuan and Cheng, Jian , booktitle=. ODE-Inspired Network Design for Single Image Super-Resolution , year=

  25. [33]

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops , pages =

    Lim, Bee and Son, Sanghyun and Kim, Heewon and Nah, Seungjun and Lee, Kyoung Mu , doi =. IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops , pages =. 1707.02921 , isbn =

    arXiv

  26. [34]

    Structure-Preserving Super Resolution With Gradient Guidance , year=

    Ma, Cheng and Rao, Yongming and Cheng, Yean and Chen, Ce and Lu, Jiwen and Zhou, Jie , booktitle=. Structure-Preserving Super Resolution With Gradient Guidance , year=

  27. [35]

    2020 , journal=

    Video Super Resolution Based on Deep Learning: A Comprehensive Survey , author=. 2020 , journal=

    2020

  28. [36]

    IEEE/CVF conference on computer vision and pattern recognition , pages=

    Deep face super-resolution with iterative collaboration between attentive recovery and landmark estimation , author=. IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [37]

    IEEE international conference on computer vision , pages=

    Deformable convolutional networks , author=. IEEE international conference on computer vision , pages=

  30. [38]

    Pattern Recognition , volume=

    Deformable face net for pose invariant face recognition , author=. Pattern Recognition , volume=. 2020 , publisher=

    2020

  31. [39]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Fsrnet: End-to-end learning face super-resolution with facial priors , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  32. [40]

    Wang, Xintao and Chan, Kelvin CK and Yu, Ke and Dong, Chao and Change Loy, Chen , booktitle=

  33. [41]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Tdan: Temporally-deformable alignment network for video super-resolution , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  34. [42]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    BasicVSR: The search for essential components in video super-resolution and beyond , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  35. [43]

    ieee/cvf conference on computer vision and pattern recognition , pages=

    Pulse: Self-supervised photo upsampling via latent space exploration of generative models , author=. ieee/cvf conference on computer vision and pattern recognition , pages=

  36. [44]

    2018 , booktitle =

    Super-Resolving Very Low-Resolution Face Images with Supplementary Attributes , author =. 2018 , booktitle =. doi:10.1109/CVPR.2018.00101 , keywords =

    2018 doi

  37. [45]

    Deep Learning Face Attributes in the Wild , year=

    Liu, Ziwei and Luo, Ping and Wang, Xiaogang and Tang, Xiaoou , booktitle=. Deep Learning Face Attributes in the Wild , year=

  38. [46]

    ACM Computing Surveys (CSUR) , volume=

    Deep learning-based face super-resolution: A survey , author=. ACM Computing Surveys (CSUR) , volume=. 2021 , publisher=

    2021

  39. [47]

    2023 , journal =

    Image Super-Resolution via Iterative Refinement , author =. 2023 , journal =

    2023

  40. [48]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Kawar, Bahjat and Vaksman, Gregory and Elad, Michael , year =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  41. [49]

    IEEE International Conference on Computer Vision , pages=

    Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution , author=. IEEE International Conference on Computer Vision , pages=

  42. [50]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Super-fan: Integrated facial landmark localization and super-resolution of real-world low resolution faces in arbitrary poses with gans , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  43. [51]

    IEEE Transactions on Image Processing , volume=

    Learning spatial attention for face super-resolution , author=. IEEE Transactions on Image Processing , volume=. 2020 , doi=

    2020

  44. [52]

    European Conference on Computer Vision (ECCV) , pages=

    Attribute-guided face generation using conditional cyclegan , author=. European Conference on Computer Vision (ECCV) , pages=. 2018 , doi=

    2018

  45. [53]

    IEEE Conference on Computer Vision and Pattern Recognition workshops , pages=

    Attribute augmented convolutional neural network for face hallucination , author=. IEEE Conference on Computer Vision and Pattern Recognition workshops , pages=. 2018 , doi=

    2018

  46. [54]

    SwinIR: Image Restoration Using Swin Transformer , year=

    Liang, Jingyun and Cao, Jiezhang and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu , booktitle=. SwinIR: Image Restoration Using Swin Transformer , year=

  47. [55]

    Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =

    Multi-Feature Aggregation in Diffusion Models for Enhanced Face Super-Resolution , author =. Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =. 2024 , month =. doi:10.1109/SIBGRAPI62404.2024.10716316 , issn =

    2024 doi

  48. [56]

    AssemblyAI , title =

  49. [57]

    and Kanade, T

    Baker, S. and Kanade, T. , journal=. Limits on super-resolution and how to break them , year=

  50. [58]

    International Conference on Machine Learning (ICML) , pages=

    Improved denoising diffusion probabilistic models , author=. International Conference on Machine Learning (ICML) , pages=. 2021 , doi=

    2021

  51. [59]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Improved techniques for training score-based generative models , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2020 , booktitle =

    2020

  52. [60]

    Score-based generative models with L

    Yoon, Eun Bi and Park, Keehun and Kim, Sungwoong and Lim, Sungbin , journal=. Score-based generative models with L. 2023 , doi=

    2023

  53. [61]

    Deep Cascaded Bi-Network for Face Hallucination

    Zhu, Shizhan and Liu, Sifei and Loy, Chen Change and Tang, Xiaoou. Deep Cascaded Bi-Network for Face Hallucination. European Conference on Computer Vision (ECCV). 2016

    2016

  54. [62]

    Neurocomputing , volume =

    Haoying Li and Yifan Yang and Meng Chang and Shiqi Chen and Huajun Feng and Zhihai Xu and Qi Li and Yueting Chen , year =. Neurocomputing , volume =. doi:10.1016/j.neucom.2022.01.029 , issn =

    2022 doi

  55. [63]

    International Conference on Learning Representations (ICLR) , month=

    Score-based generative modeling through stochastic differential equations , author=. International Conference on Learning Representations (ICLR) , month=. 2021 , pages=

    2021

  56. [64]

    2019 , booktitle =

    Generative Modeling by Estimating Gradients of the Data Distribution , author =. 2019 , booktitle =

    2019

  57. [65]

    International Conference on Machine Learning (ICML) , volume =

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics , author =. International Conference on Machine Learning (ICML) , volume =. 2015 , month =

    2015

  58. [66]

    Denoising Diffusion Probabilistic Models , volume =

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , booktitle =. Denoising Diffusion Probabilistic Models , volume =. 2020 , doi=

    2020

  59. [67]

    Stochastic Processes and their Applications , volume=

    Reverse-time diffusion equation models , author=. Stochastic Processes and their Applications , volume=. 1982 , publisher=

    1982

  60. [68]

    Towards Real-World Blind Face Restoration with Generative Facial Prior , year=

    Wang, Xintao and Li, Yu and Zhang, Honglun and Shan, Ying , booktitle=. Towards Real-World Blind Face Restoration with Generative Facial Prior , year=

  61. [69]

    arXiv preprint , year=

    Gotta Go Fast When Generating Data with Score-Based Models , author=. arXiv preprint , year=

  62. [70]

    arXiv preprint , year=

    Modify the improved Euler scheme to integrate stochastic differential equations , author=. arXiv preprint , year=

  63. [71]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Score-based Generative Modeling in Latent Space , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  64. [72]

    International Conference on Learning Representations (ICLR) , year=

    Denoising diffusion implicit models , author=. International Conference on Learning Representations (ICLR) , year=

  65. [73]

    Geophysical Journal International , volume=

    Bayesian seismic inversion: a fast sampling Langevin dynamics Markov chain Monte Carlo method , author=. Geophysical Journal International , volume=. 2021 , publisher=

    2021

  66. [74]

    Journal of Statistical Physics , volume=

    Optimal non-reversible linear drift for the convergence to equilibrium of a diffusion , author=. Journal of Statistical Physics , volume=. 2013 , publisher=

    2013

  67. [75]

    Journal of Statistical Physics , volume=

    Attaining the optimal Gaussian diffusion acceleration , author=. Journal of Statistical Physics , volume=. 2014 , publisher=

    2014

  68. [76]

    Proceedings of Machine Learning Research , volume=

    Breaking Reversibility Accelerates Langevin Dynamics for Non-Convex Optimization , author=. Proceedings of Machine Learning Research , volume=

  69. [77]

    Diffusion Models Beat

    Dhariwal, Prafulla and Nichol, Alexander , year =. Diffusion Models Beat. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  70. [78]

    Nanxin Chen and Yu Zhang and Heiga Zen (Byungha Chun) and Ron Weiss and Mohammad Norouzi and William Chan , year =

  71. [79]

    European Conference on Computer Vision (ECCV) , pages=

    Learning gradient fields for shape generation , author=. European Conference on Computer Vision (ECCV) , pages=. 2020 , doi=

    2020

  72. [80]

    Chenlin Meng and Yutong He and Yang Song and Jiaming Song and Jiajun Wu and Jun-Yan Zhu and Stefano Ermon , year =

  73. [81]

    2022 , booktitle =

    Solving Inverse Problems in Medical Imaging with Score-Based Generative Models , author=. 2022 , booktitle =

    2022

  74. [82]

    International Conference on Artificial Intelligence and Statistics (AISTATS) , volume =

    Permutation Invariant Graph Generation via Score-Based Generative Modeling , author =. International Conference on Artificial Intelligence and Statistics (AISTATS) , volume =. 2020 , month =

    2020

  75. [83]

    International Conference on Learning Representations (ICLR) , year=

    Score-Based Generative Modeling with Critically-Damped Langevin Diffusion , author=. International Conference on Learning Representations (ICLR) , year=

  76. [84]

    Neural computation , volume=

    A connection between score matching and denoising autoencoders , author=. Neural computation , volume=. 2011 , publisher=

    2011

  77. [85]

    2019 , publisher=

    Applied stochastic differential equations , author=. 2019 , publisher=

    2019

  78. [86]

    2022 , booktitle =

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding , author =. 2022 , booktitle =

    2022

  79. [87]

    Oxford VGGFace Implementation using Keras Functional Framework v2+

  80. [88]

    Nuclear Physics B , volume=

    Correlation functions and computer simulations , author=. Nuclear Physics B , volume=. 1981 , publisher=

    1981

  81. [89]

    Handbook of markov chain monte carlo , volume=

    MCMC using Hamiltonian dynamics , author=. Handbook of markov chain monte carlo , volume=. 2011 , publisher=

    2011

  82. [90]

    and Shechtman, Eli and Wang, Oliver , title=

    Zhang, Richard and Isola, Phillip and Efros, Alexei A. and Shechtman, Eli and Wang, Oliver , title=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  83. [91]

    A Style-Based Generator Architecture for Generative Adversarial Networks , year=

    Karras, Tero and Laine, Samuli and Aila, Timo , booktitle=. A Style-Based Generator Architecture for Generative Adversarial Networks , year=

  84. [92]

    Progressive growing of

    Karras, Tero and Aila, Timo and Laine, Samuli and Lehtinen, Jaakko , booktitle =. Progressive growing of. 2018 , pages=

    2018

  85. [93]

    Numerical Solution of Stochastic Differential Equations , volume =

    Kloeden, Peter and Platen, Eckhard , year =. Numerical Solution of Stochastic Differential Equations , volume =

  86. [94]

    IEEE International Conference on Automatic Face & Gesture Recognition (FG) , pages=

    Low resolution face recognition in surveillance systems using discriminant correlation analysis , author=. IEEE International Conference on Automatic Face & Gesture Recognition (FG) , pages=

  87. [95]

    IEEE Transactions on Information Forensics and Security , volume=

    On low-resolution face recognition in the wild: Comparisons and new techniques , author=. IEEE Transactions on Information Forensics and Security , volume=. 2019 , publisher=

    2019

  88. [96]

    Electronics , volume=

    A Short Survey on Machine Learning Explainability: An Application to Periocular Recognition , author=. Electronics , volume=. 2021 , publisher=

    2021

  89. [97]

    Iberoamerican Congress on Pattern Recognition (CIARP) , volume =

    Multi-Task Learning for Low-Resolution License Plate Recognition , author =. Iberoamerican Congress on Pattern Recognition (CIARP) , volume =. 2019 , month =

    2019

  90. [98]

    International Conference on Computer Vision Theory and Applications (VISAPP) , volume =

    On the Cross-dataset Generalization in License Plate Recognition , author =. International Conference on Computer Vision Theory and Applications (VISAPP) , volume =. 2022 , month =. doi:10.5220/0010846800003124 , isbn =

    2022 doi

  91. [99]

    Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =

    A First Look at Dataset Bias in License Plate Recognition , author =. Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =. 2022 , month =. doi:10.1109/SIBGRAPI55357.2022.9991768 , issn =

    2022 doi

  92. [100]

    2017 , booktitle =

    A Deeper Look at Dataset Bias , author =. 2017 , booktitle =. doi:10.1007/978-3-319-58347-1\_2 , isbn =

    2017 doi

  93. [101]

    2011 , booktitle =

    Unbiased look at dataset bias , author =. 2011 , booktitle =

    2011

  94. [102]

    Grgic, Mislav and Delac, Kresimir and Grgic, Sonja , journal=

  95. [103]

    IET Biometrics , volume=

    Neves, Joao and Moreno, Juan and Proen. IET Biometrics , volume=. 2018 , publisher=

    2018

  96. [104]

    Defying Limits: Super-Resolution Refinement with Diffusion Guidance , booktitle=

    Marcelo. Defying Limits: Super-Resolution Refinement with Diffusion Guidance , booktitle=. 2024 , pages=. doi:10.5220/0012398900003660 , isbn=

    2024 doi

  97. [105]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Implicit diffusion models for continuous super-resolution , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2023 , doi=

    2023

  98. [106]

    Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =

    Face Super-Resolution Using Stochastic Differential Equations , author =. Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =. 2022 , month =. doi:10.1109/SIBGRAPI55357.2022.9991799 , issn =

    2022 doi

  99. [107]

    arXiv preprint , year=

    Learning face representation from scratch , author=. arXiv preprint , year=

  100. [108]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Deep residual learning for image recognition , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2016 , doi=

    2016

  101. [109]

    2022 , doi=

    Kim, Minchul and Jain, Anil K and Liu, Xiaoming , booktitle=. 2022 , doi=

    2022

  102. [110]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Collaborative diffusion for multi-modal face generation and editing , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  103. [111]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    On Distillation of Guided Diffusion Models , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2023 , doi=

    2023

  104. [112]

    IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration , author=. IEEE/CVF Winter Conference on Applications of Computer Vision , pages=. 2024 , doi=

    2024

  105. [113]

    2023 , doi=

    Zhang, Zhixing and Han, Ligong and Ghosh, Arnab and Metaxas, Dimitris and Ren, Jian , booktitle=. 2023 , doi=

    2023

  106. [114]

    ITG Conference on Speech Communication , pages=

    Audio-visual speech enhancement with score-based generative models , author=. ITG Conference on Speech Communication , pages=. 2023 , doi=

    2023

  107. [115]

    2023 , journal =

    Super-Resolution of License Plate Images Using Attention Modules and Sub-Pixel Convolution Layers , author =. 2023 , journal =. doi:10.1016/j.cag.2023.05.005 , issn =

    2023 doi

  108. [116]

    Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =

    Combining Attention Module and Pixel Shuffle for License Plate Super-resolution , author =. Conference on Graphics, Patterns and Images (SIBGRAPI) , volume =. 2022 , month =. doi:10.1109/SIBGRAPI55357.2022.9991753 , issn =

    2022 doi

  109. [117]

    2024 , journal =

    Enhancing License Plate Super-Resolution: A Layout-Aware and Character-Driven Approach , author =. 2024 , journal =. doi:10.1109/SIBGRAPI62404.2024.10716303 , issn =

    2024 doi

  110. [118]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Fourier features let networks learn high frequency functions in low dimensional domains , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2020 , doi=

    2020

  111. [119]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015. 2015

    2015

  112. [120]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Attention is all you need , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2017 , doi=

    2017

  113. [121]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Residual attention network for image classification , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2017 , doi=

    2017

  114. [122]

    In Defense of the Triplet Loss Again: Learning Robust Person Re-Identification with Fast Approximated Triplet Loss and Label Distillation , year=

    Yuan, Ye and Chen, Wuyang and Yang, Yang and Wang, Zhangyang , booktitle=. In Defense of the Triplet Loss Again: Learning Robust Person Re-Identification with Fast Approximated Triplet Loss and Label Distillation , year=

  115. [123]

    arXiv preprint , year=

    Survey of dropout methods for deep neural networks , author=. arXiv preprint , year=

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.