Pith. sign in

REVIEW 3 major objections 4 minor 39 references

LLV-FSR: Exploiting Large Language-Vision Prior for Face Super-resolution

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read LLV-FSR claims that fusing language-vision priors—text captions, descriptions, semantic masks, and depth maps—with pixel features improves face super-resolution beyond prior methods.

desk verdict Novel combination of language-vision priors for face SR, but the ablation undercuts the claim that the priors drive the gain. read the letter →

arxiv 2411.09293 v1 pith:ZGBYNTJ6 submitted 2024-11-14 cs.CV

classification cs.CV
keywords facesuper-resolutionlanguage-visionpriorlargelanguagemodelsimagecaptioningsemanticmaskdepthestimationill-posedinverseproblemMMCelebA-HQ
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

This paper tries to establish that face super-resolution, an ill-posed inverse problem, benefits from pluralistic priors that go beyond pixels: image captions, detailed text descriptions, facial semantic masks, and depth maps. It proposes a framework, LLV-FSR, that generates these priors from the low-resolution input using large pre-trained models and fuses them through a dedicated attention block. The authors report that this fusion outperforms existing face super-resolution methods, with a PSNR of 25.95 dB at ×16 upscaling, 0.43 dB higher than the best comparison method WFEN, and also better identity preservation. A sympathetic reader would care because the result suggests that non-visual, language-based knowledge can regularize and sharpen face reconstruction in a way that purely visual priors have not.

What carries the argument

The key machinery is the Language-Vision Prior Fusion Block (LVPFB), which takes the low-resolution feature map and four prior embeddings—caption embedding, description embedding, semantic-mask feature, and depth-map feature—and processes them through four parallel attention branches: SegA and DepA learn facial-structure-aware spatial attention from mask and depth; CapA applies global average pooling and sigmoid to create global caption attention; DesA uses the description as queries and the LR feature as keys/values in cross-attention. The outputs are concatenated with the LR features and skip-connected. The priors are generated per-image from the LR input: BLIP2 for captions, ChatGPT-4 with a prompt for descriptions, SAM for semantic masks, and DAM for depth maps, with text embeddings extracted by a frozen CLIP text encoder.

What would settle it

Generate ground-truth captions and descriptions from the high-resolution faces and feed those into LLV-FSR instead of the BLIP2/ChatGPT outputs; if PSNR and LPIPS do not improve over the LR-derived priors, then the method's gain is not actually driven by the language content. Alternatively, corrupt the caption with unrelated text and check whether performance collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a unified language-vision prior, composed of a BLIP2-generated caption, a ChatGPT-4-generated description, a SAM-generated semantic mask, and a DAM-generated depth map, can be effectively injected into a transformer-based super-resolution network to yield state-of-the-art face reconstruction. The reported numbers on MMCelebA-HQ are PSNR 29.35 dB at ×8 and 25.95 dB at ×16, with the ×16 value 0.43 dB higher than the second-best method WFEN, along with gains in SSIM, LPIPS, and NIQE. The paper also shows, via identity-distance measurement with a face recognition model, that the super-resolved images retain identity better than baselines. The ablation study attributes these gains to the proposed language-vision prior fusion block, which integrates the four priors through four parallel attention mechanisms instead of simple concatenation.

Load-bearing premise

The whole method rests on the assumption that the captions, descriptions, masks, and depth maps extracted from an extremely low-resolution image are accurate enough to guide reconstruction; the paper does not measure how often these priors are correct.

Editorial extensions

If this is right

  • If the claim holds, language-vision priors become a viable plug-in for other ill-posed restoration tasks beyond faces, such as general image super-resolution or deblurring.
  • The design suggests that descriptions, which carry fine-grained detail, are the most identity-critical prior, potentially guiding future work on text-guided restoration.
  • The reported identity-distance improvement implies that language-vision priors could help downstream face-recognition systems operating on low-resolution imagery.
  • The framework indicates a practical path to using large pre-trained vision-language models without fine-tuning them, by freezing them and only training the fusion network.

Reading between the lines

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

  • A testable extension the authors did not run: replace the automatically generated captions/descriptions with ground-truth captions written from the high-resolution face, which would reveal how much of the gain comes from prior correctness versus prior presence.
  • If caption models hallucinate details not present in the degraded input, the language prior could steer reconstruction toward plausible but wrong facial features; measuring caption accuracy against the HR face would clarify this risk.
  • The same fusion block could be applied to other modalities, such as audio or attribute tags, suggesting a general architecture for multi-modal priors in inverse problems.
  • The reported qualitative gains at ×16, where pixel information is extremely sparse, hint that text priors carry information that survives extreme downsampling better than geometric priors, which is worth testing directly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes LLV-FSR, a face super-resolution framework that augments a transformer-based backbone with language-vision priors: visual priors (semantic mask, depth map) and textual priors (caption, description) are generated from the low-resolution input using pretrained models (SAM, DAM, BLIP2, ChatGPT-4), then injected through a newly designed Language-Vision Prior Fusion Block (LVPFB). The central claim is that this pluralistic prior fusion improves FSR quality beyond existing state-of-the-art methods, supported by comparisons on MMCelebA-HQ at ×8 and ×16, an ablation study of the fusion block, and qualitative visualizations. The paper also reports identity-preservation comparisons using DeepFace cosine distance.

Significance. If the reported gains are robust, the paper proposes a plausible new direction: using large language-vision models to generate semantic and linguistic priors for ill-posed face super-resolution. The LVPFB is a simple, reusable idea, and the authors compare against a broad set of baselines, provide qualitative examples, and include an identity-preservation analysis. The main strengths are the breadth of the empirical comparison and the clear architectural exposition. However, the central attribution of the improvement to the language-vision prior is not yet established: the no-prior baseline already outperforms the prior state of the art, the ablation gains are small and unaccompanied by variance estimates, the prior-generation step is never validated for correctness, and the headline numerical claim is misreported in the text. These issues are fixable with additional experiments and corrections.

major comments (3)
  1. [§4.3, Table 1] The text states that "on ×8 face super-resolution task, the PSNR of our LLV-FSR is 25.95 dB which is 0.43 dB higher than the second-best method WFEN." Table 1 shows the opposite: at ×8, LLV-FSR reaches 29.35 dB and WFEN reaches 29.10 dB (a gain of 0.25 dB), while SFMNet reaches 29.12 dB, so WFEN is not the second-best method at ×8. The numbers 25.95 dB and 0.43 dB correspond to the ×16 column (LLV-FSR 25.95 versus WFEN 25.52). This misattribution of the headline result must be corrected, and the comparison sentence should report the two scales separately.
  2. [§4.4, Table 2] Model 1, which removes LVPFB and all language-vision priors, already attains 29.17 dB PSNR, which is higher than SFMNet (29.12 dB) and WFEN (29.10 dB) in Table 1. Therefore the claimed superiority over prior methods on ×8 mostly comes from the new transformer backbone and not from the language-vision prior. The incremental gains are +0.07 dB for concatenation (Model 2) and +0.18 dB for the full LVPFB, but no error bars, multiple seeds, or statistical tests are reported for any ablation. Without such statistics, these small differences may be run-to-run noise, and the central attribution claim is not supported. The authors should report repeated-run statistics and also provide the ablation at ×16, since the largest advertised gain (0.43 dB) is at ×16 and Table 2 reports only ×8.
  3. [§3.3, §4.4] The language and vision priors are generated by feeding the low-resolution image (16×16 or 32×32) into BLIP2, ChatGPT-4, SAM, and DAM, but the paper never evaluates whether the resulting captions, descriptions, masks, and depth maps are accurate or whether the language models hallucinate facial details. Section 4.4 discusses the effect of removing each prior only qualitatively, without any quantitative prior-correctness metric or oracle experiment (e.g., replacing the generated caption with a ground-truth caption to measure the ceiling). Since Eq. (1) injects EC, ED, FS, and FD into every fusion block, the mechanism is load-bearing: if the caption model hallucinates details from a 16×16 input, the prior can steer reconstruction toward incorrect facial features. The authors should add a quantitative evaluation of prior quality or an oracle ablation to substantiate the claim that the language-vision prior, rather than the backbone alone, drives the improvement.
minor comments (4)
  1. [§4.4, §3.2] There are several typos and prose issues, including "differnt" in Section 4.4, "To summary" instead of "To summarize," "LVFPB" instead of "LVPFB" in the same section, and the paragraph in Section 3.2 beginning "o effectively synergize." These should be corrected.
  2. [References] The reference list contains a placeholder citation "SGPN citesgpn" in Section 2.1, and reference [16] is attributed to "Proceedings of the 29th ACM International Conference on Multimedia, 2024," which appears inconsistent with the conference year and should be verified.
  3. [§3.2, Fig. 3] The notation in Eq. (1) and the description of LVPFB would benefit from explicit tensor dimensions and a definition of L and the feature channels; L is set to 7 only in Section 4.2, and the architecture figure does not clearly show the skip connection and concatenation paths described in the text.
  4. [§4.4, Fig. 9] The ablation for individual priors (Fig. 9) is purely qualitative and the text notes that removal of the caption yields visually similar results; a quantitative table for these prior-removal experiments would help the reader judge the contribution of each prior.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the FSR pipeline is an empirical supervised system benchmarked against external baselines; existing self-citations are contextual and not load-bearing.

full rationale

The paper's derivation chain is an empirical supervised pipeline: a network maps LR to HR under an L1 loss (Eq. 2), with auxiliary language-vision features (caption, description, semantic mask, depth map) extracted by frozen external models (BLIP2, ChatGPT-4, SAM, DAM, CLIP) from the LR input. These priors are fixed external features, not fitted parameters later renamed as predictions, so no step is equivalent to its inputs by construction. The central quantitative claim ('the PSNR of our LLV-FSR is 25.95 dB which is 0.43 dB higher than the second-best method WFEN') is benchmarked in Table 1 against independently trained external methods on a shared test set, and the ablation in Table 2 compares the full model against stripped variants (Model 1 and Model 2); whatever one thinks of the small +0.18 dB gain and the absence of error bars or significance tests, the comparison is empirical and not circular. Self-citations exist (the survey ref [12] in Section 1; the SFMNet baseline ref [25] and FaceFormer ref [29] as comparison methods, each overlapping with the author list) but they are used only as context and comparison targets, not as justification of any load-bearing premise, so they do not constitute circular support. No uniqueness theorem or ansatz is imported from the authors' prior work, and no known result is renamed under new coordinates. The text's apparent misattribution of the 25.95 dB / 0.43 dB figure to the x8 task (Section 4.3) when Table 1 assigns it to x16 is a reporting inconsistency, not a circular step. Genuine empirical concerns (hallucination-prone priors, no prior-correctness evaluation, no error bars) are correctness-risk issues, not circularity, and are out of scope for this pass.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities, particles, or forces. The central claim rests on empirical assumptions about the usefulness of pretrained model outputs and the synthetic degradation model, plus a small set of hand-chosen hyperparameters that are not ablated.

free parameters (3)
  • L (number of fusion blocks) = 7
    Chosen by hand in Section 4.2. No ablation over L is provided, so the contribution of this architectural choice to the central claim is unknown.
  • Learning rate = 2e-4
    Standard Adam learning rate reported in Section 4.2. It affects training but is not fitted to the target result.
  • Training epochs = 30
    Fixed schedule in Section 4.2. No analysis of convergence or overfitting is provided.
assumptions (3)
  • domain assumption Pretrained BLIP2, ChatGPT-4, SAM, and Depth Anything provide useful priors when applied to heavily downsampled LR face images.
    Invoked in Section 3.3, where the LR image is directly fed into these models. The paper provides no quantitative evaluation of prior accuracy on small inputs.
  • domain assumption Bicubic downsampling is the degradation model for training and evaluation.
    Stated in Section 4.2. Real-world low-resolution faces may follow different degradations, so the central claim is scoped to synthetic bicubic downsampling.
  • standard math Standard attention and convolution primitives behave as expected in the fusion block.
    The LVPFB relies on cross-attention, global average pooling, and convolution, which are standard operations in the deep learning literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLV-FSR: Exploiting Large Language-Vision Prior for Face Super-resolution." pith.science (2026). https://pith.science/paper/ZGBYNTJ6

@misc{pith2026241109293,
  author       = {Pith},
  title        = {Pith review of: LLV-FSR: Exploiting Large Language-Vision Prior for Face Super-resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZGBYNTJ6}},
  note         = {Machine review of arXiv:2411.09293}
}
read the original abstract

Existing face super-resolution (FSR) methods have made significant advancements, but they primarily super-resolve face with limited visual information, original pixel-wise space in particular, commonly overlooking the pluralistic clues, like the higher-order depth and semantics, as well as non-visual inputs (text caption and description). Consequently, these methods struggle to produce a unified and meaningful representation from the input face. We suppose that introducing the language-vision pluralistic representation into unexplored potential embedding space could enhance FSR by encoding and exploiting the complementarity across language-vision prior. This motivates us to propose a new framework called LLV-FSR, which marries the power of large vision-language model and higher-order visual prior with the challenging task of FSR. Specifically, besides directly absorbing knowledge from original input, we introduce the pre-trained vision-language model to generate pluralistic priors, involving the image caption, descriptions, face semantic mask and depths. These priors are then employed to guide the more critical feature representation, facilitating realistic and high-quality face super-resolution. Experimental results demonstrate that our proposed framework significantly improves both the reconstruction quality and perceptual quality, surpassing the SOTA by 0.43dB in terms of PSNR on the MMCelebA-HQ dataset.

Figures

Figures reproduced from arXiv: 2411.09293 by the authors.

Figure 1
Figure 1. Comparison of FSR framework. The green part denotes [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. Our method first inputs the LR face image into the pretrained large-scale model to extract [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Framework of language-vision prior fusion block. (a): LVPFB; (b): SegA and DepA; (c): CapA; (d): DesA. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Language-vision prior visualization. (a): HR; (b): LR; (c): Semantic mask; (d): Depth; (e): Caption; (f): Description. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: ×8 FSR results of state-of-the-art methods on MMCelebA-HQ dataset. (a): LR; (b): SRCNN; (c): FSRNet; (d): DIC; (e): SISN; (f): FaceFormer; (g): SFMNet; (h): WFEN; (i): LLV-FSR; (j): HR. performs the existing methods. For example, on ×8 face super-resolution task, the P…
Figure 6
Figure 6. Figure 6: ×16 FSR results of state-of-the-art methods on MMCelebA-HQ dataset. (a): LR; (b): SRCNN; (c): FSRNet; (d): DIC; (e): SISN; (f): FaceFormer; (g): SFMNet; (h): WFEN; (i): LLV-FSR; (j): HR. key facial components. These methods frequently generate distorted faces, with unn…
Figure 7
Figure 7. Figure 7: Identity distance comparison [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Ablation study. (a): LR; (b): Model 1; (c): Model 2; (d): [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

  2. [2]

    Glean: Generative latent bank for image super-resolution and beyond

    Kelvin CK Chan, Xiangyu Xu, Xintao Wang, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for image super-resolution and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3154–3168, 2022. 1

  3. [3]

    Chan, Xiangyu Xu, Xintao Wang, Jinwei Gu, and Chen Change Loy

    Kelvin C.K. Chan, Xiangyu Xu, Xintao Wang, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for im- age super-resolution and beyond. IEEE Transactions on Pat- tern Analysis and Machine Intelligence , 45(3):3154–3168,

  4. [4]

    Y . Chen, Y . Tai, X. Liu, C. Shen, and J. Yang. FSRNet: End-to-end learning face super-resolution with facial priors. In Proceedings of The IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2492–2501, 2018. 2, 5

  5. [5]

    Blind face restoration under extreme conditions: Leveraging 3d-2d prior fusion for superior structural and texture recovery

    Zhengrui Chen, Liying Lu, Ziyang Yuan, Yiming Zhu, Yu Li, Chun Yuan, and Weihong Deng. Blind face restoration under extreme conditions: Leveraging 3d-2d prior fusion for superior structural and texture recovery. Proceedings of the AAAI Conference on Artificial Intelligence, 38(2):1263– 1271, 2024. 1, 2

  6. [6]

    BERT: pre-training of deep bidirectional trans- formers for language understanding.CoRR, abs/1810.04805,

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional trans- formers for language understanding.CoRR, abs/1810.04805,

  7. [7]

    Image super-resolution using deep convolutional net- works

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional net- works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295–307, 2015. 5

  8. [8]

    The elements of end-to-end deep face recogni- tion: A survey of recent advances

    Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recogni- tion: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1–42, 2022. 1

Show all 39 references
  1. [9]

    Ctcnet: A cnn-transformer co- operation network for face image super-resolution

    Guangwei Gao, Zixiang Xu, Juncheng Li, Jian Yang, Tiey- ong Zeng, and Guo-Jun Qi. Ctcnet: A cnn-transformer co- operation network for face image super-resolution. IEEE Transactions on Image Processing, 32:1978–1991, 2023. 2

  2. [10]

    Implicit diffusion models for continuous super-resolution

    Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yan- jing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages...

  3. [11]

    Semi-cycled generative adver- sarial networks for real-world face super-resolution

    Hao Hou, Jun Xu, Yingkun Hou, Xiaotao Hu, Benzheng Wei, and Dinggang Shen. Semi-cycled generative adver- sarial networks for real-world face super-resolution. IEEE Transactions on Image Processing, 32:1184–1199, 2023. 2

  4. [12]

    Deep learning-based face super-resolution: A survey

    Junjun Jiang, Chenyang Wang, Xianming Liu, and Jiayi Ma. Deep learning-based face super-resolution: A survey. ACM Computing Surveys, 55(1):1–36, 2023. 1

  5. [13]

    Prompting visual-language models for efficient video understanding

    Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In European Conference on Computer Vi- sion, pages 105–124. Springer, 2022. 1

  6. [14]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 3, 5

  7. [15]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In In- ternational conference on machine learning , pages 19730– 19742. PMLR, 2023. 3

  8. [16]

    Efficient face super-resolution via wavelet-based feature enhancement network

    Wenjie Li, Heng Guo, Xuannan Liu, Kongming Liang, Jiani Hu, Zhanyu Ma, and Jun Guo. Efficient face super-resolution via wavelet-based feature enhancement network. InProceed- ings of the 29th ACM International Conference on Multime- dia, 2024. 5

  9. [17]

    Face hallucination via split-attention in split-attention network

    Tao Lu, Yuanzhi Wang, Yanduo Zhang, Yu Wang, Liu Wei, Zhongyuan Wang, and Junjun Jiang. Face hallucination via split-attention in split-attention network. In Proceedings of the 29th ACM International Conference on Multimedia , pages 5501–5509, 2021. 5

  10. [18]

    Deep face super-resolution with iterative collabora- tion between attentive recovery and landmark estimation

    Cheng Ma, Zhenyu Jiang, Yongming Rao, Jiwen Lu, and Jie Zhou. Deep face super-resolution with iterative collabora- tion between attentive recovery and landmark estimation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 5569–5578, ...

  11. [19]

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

    Sachit Menon, Alex Damian, McCourt Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceed- ing of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2223–2232, 2020. 2

  12. [20]

    Waveface: Authentic face restoration with efficient frequency recovery

    Yunqi Miao, Jiankang Deng, and Jungong Han. Waveface: Authentic face restoration with efficient frequency recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 6583–6592,

  13. [21]

    completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik. Making a “completely blind” image quality analyzer.IEEE Signal Pro- cessing Letters, 20(3):209–212, 2013. 5

  14. [22]

    Grit: Faster and better image captioning trans- former using dual visual features

    Van-Quang Nguyen, Masanori Suganuma, and Takayuki Okatani. Grit: Faster and better image captioning trans- former using dual visual features. In European Conference on Computer Vision, pages 167–184. Springer, 2022. 3

  15. [23]

    Learning transferable visual models from natural language supervi- sion

    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 supervi- sion. In International conference on machine learning, ...

  16. [24]

    Lightface: A hybrid deep face recognition framework

    Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In Proceedings of Inno- vations in Intelligent Systems and Applications Conference , pages 23–27, 2020. 7

  17. [25]

    Spatial-frequency mutual learning for face super- resolution

    Chenyang Wang, Junjun Jiang, Zhiwei Zhong, and Xian- ming Liu. Spatial-frequency mutual learning for face super- resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22356– 22366, 2023. 2, 5

  18. [26]

    To- wards real-world blind face restoration with generative fa- cial prior

    Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. To- wards real-world blind face restoration with generative fa- cial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9168– 9178, 2021. 2

  19. [27]

    Panini-net: Gan prior based degradation-aware feature interpolation for face restoration

    Yinhuai Wang, Yujie Hu, and Jian Zhang. Panini-net: Gan prior based degradation-aware feature interpolation for face restoration. In Proceedings of the AAAI Conference on Arti- ficial Intelligence, pages 2576–2584, 2022. 2

  20. [28]

    Gan prior based null-space learning for consistent super- resolution

    Yinhuai Wang, Yujie Hu, Jiwen Yu, and Jian Zhang. Gan prior based null-space learning for consistent super- resolution. In Proceedings of the AAAI Conference on Ar- tificial Intelligence, pages 2724–2732, 2023. 2

  21. [29]

    Faceformer: Aggregating global and local representation for face hallucination

    Yuanzhi Wang, Tao Lu, Yanduo Zhang, Zhongyuan Wang, Junjun Jiang, and Zixiang Xiong. Faceformer: Aggregating global and local representation for face hallucination. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2533–2545, 2023. 5

  22. [30]

    Zhou Wang, A. C. Bovik, Hamid Rahim Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Trans. Image Processing , 13 (4):600–612, 2004. 5

  23. [31]

    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 (CVPR), pages 1...

  24. [32]

    Tedigan: Text-guided diverse face image generation and ma- nipulation

    Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and ma- nipulation. In IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), 2021. 5

  25. [33]

    Towards open-world text-guided face image generation and manipulation

    Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Towards open-world text-guided face image generation and manipulation. arxiv preprint arxiv: 2104.08910, 2021. 5

  26. [34]

    A dive into sam prior in image restoration

    Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023. 5

  27. [35]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 5

  28. [36]

    Multiprior learning via neural architec- ture search for blind face restoration

    Yanjiang Yu, Puyang Zhang, Kaihao Zhang, Wenhan Luo, and Changsheng Li. Multiprior learning via neural architec- ture search for blind face restoration. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2

  29. [37]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a percep- tual metric. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, pages 586–595, 2018. 5

  30. [38]

    A survey of deep facial attribute analysis

    Xin Zheng, Yanqing Guo, Huaibo Huang, Yi Li, and Ran He. A survey of deep facial attribute analysis. International Journal of Computer Vision, 128(8):2002–2034, 2020. 1

  31. [39]

    Blind face restora- tion via integrating face shape and generative priors

    Feida Zhu, Junwei Zhu, Wenqing Chu, Xinyi Zhang, Xi- aozhong Ji, Chengjie Wang, and Ying Tai. Blind face restora- tion via integrating face shape and generative priors. In 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 7652–7661, 2022. 1

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.