Pith. sign in

REVIEW 3 major objections 58 references

PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration

T0 review · 3 major / 0 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read A training-free diffusion method recovers natural light, color, and identity from real dark faces by coupling physical lighting rules with style-stripped facial structure.

desk verdict Solid training-free joint LL-BFR guidance with a useful real benchmark; the AdaIN decoupling claim is only lightly stress-tested but the overall empirical package is strong enough to take seriously. read the letter →

arxiv 2603.24969 v2 pith:N7INGFV5 submitted 2026-03-26 cs.CV

Yilin Ni , Wenjie Li , Zhengxue Wang , Juncheng Li , Guangwei Gao , Jian Yang This is my paper · ORCID
classification cs.CV
keywords low-lightenhancementblindfacerestorationdiffusionmodelstraining-freeguidanceRetinexstyle-agnosticstructuralinjectionWildDark-Face
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

Real low-light faces carry compound damage: darkness, noise, blur, and washed-out color. Cascading separate enhancers and face restorers usually amplifies noise or erases identity, while generic joint networks lack face-specific knowledge and leave residual blur. PASDiff treats restoration as guided sampling of a pre-trained diffusion model. Photometric losses derived from inverse-intensity weighting and Retinex force plausible exposure and intrinsic color. Style-Agnostic Structural Injection then pulls high-frequency facial geometry from an off-the-shelf restorer while statistically stripping that restorer’s own lighting bias, so identity details can coexist with the physical constraints. On both synthetic data and a new 700-image real-world set (WildDark-Face), the method produces brighter, more natural, and more identity-preserving faces than prior cascaded or joint baselines, all without any task-specific training.

What carries the argument

Style-Agnostic Structural Injection (SASI): Adaptive Instance Normalization that matches the mean and variance of an off-the-shelf face-restoration prior to the current diffusion estimate, so the structural loss supplies identity geometry without re-introducing the prior’s laboratory lighting or color.

What would settle it

Replace AdaIN alignment with a direct pixel or feature MSE to the external face prior; if the resulting images still keep both natural illumination and identity fidelity, the claimed style-filtering step is unnecessary.

Watch

Extended reading notes

Core claim

Joint low-light enhancement and blind face restoration can be solved without retraining by steering an unconditional diffusion model with two orthogonal energy terms: physics-based photometric constraints (spatially varying exposure plus Retinex reflectance) that recover natural illumination and color, and a statistic-aligned structural loss that injects only high-frequency facial geometry from an external prior after AdaIN has removed that prior’s lighting style.

Load-bearing premise

That simply matching mean and variance fully separates a face prior’s high-frequency structure from its lighting and color biases, so the structural term never fights the physical lighting constraints.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. PASDiff is a training-free diffusion guidance framework for joint low-light enhancement and blind face restoration. It steers an unconditional ImageNet DDPM with a multi-objective energy: (i) photometric terms from inverse-intensity exposure maps and Retinex reflectance anchors, and (ii) Style-Agnostic Structural Injection (SASI), which AdaIN-aligns an off-the-shelf face restorer (OSDFace) so that only high-frequency structure is injected. The authors also release WildDark-Face (700 real low-light face crops). On synthetic FFHQ and WildDark-Face, PASDiff reports better PSNR/LPIPS/Deg/LMD, no-reference IQA, face-recognition accuracy, and user preference than cascaded LLIE+BFR pipelines and several joint restorers.

Significance. Joint real-world low-light face restoration is practically important (surveillance, night photography) and poorly served by pure cascades or generic joint models. A training-free multi-objective guidance recipe that reuses existing Retinex and face-restoration networks is useful if it truly balances illumination, color, and identity. The WildDark-Face benchmark and the promised code/dataset release are concrete community assets. Empirical gains on recognition accuracy (Table 1) and identity metrics (Table 2) are meaningful if they survive stronger stress tests of the SASI decoupling claim.

major comments (3)
  1. Sec. 3.2, Eq. (10) and Table 5 / Fig. 9: The central narrative is that AdaIN fully strips photometric style from Φ while preserving identity geometry, so L_stru does not fight L_exp/L_ref. The only direct evidence is a modest MSE-vs-SASI ablation on synthetic FFHQ. That does not measure residual illumination/chromatic leakage under WildDark-Face photon starvation, nor how much of the large Deg./recognition gains (Table 1: 71.43% vs ~64%) come from re-injecting Φ’s high-quality identity manifold rather than geometry consistent with the physical constraints. Please add (i) a real-data style-leakage diagnostic (e.g., low-frequency color/illumination distance between Φ(x̂0) and x̂align, or a controlled swap of Φ’s lighting), and (ii) an identity-metric breakdown with/without SASI on a held-out real subset with proxy labels or human identity ranking.
  2. Sec. 3.2, Eq. (12) and implementation details: Balancing weights λ_exp=1200, λ_ref=0.03, λ_stru=10000 (plus α,β,s,N) are extreme and purely empirical. Supplement Fig. S4 shows only visual stability under small perturbations. Because the “orthogonal physical vs structural” claim depends on these scales not letting one term dominate, report quantitative sensitivity (PSNR/LPIPS/Deg/LMD and at least one no-ref metric) over a wider grid, and state a selection protocol that is not tuned to the test set.
  3. Table 2 and Sec. 4.2: On WildDark-Face the superiority claim for “identity consistency” rests on no-reference IQA, FID, visuals, and a 20-image user study; Deg/LMD and the recognition protocol (Table 1) require GT and appear synthetic-only. Either (a) add a real-world identity protocol (e.g., same-person pairs from DarkFace scenes, or forced-choice identity ranking against cascades), or (b) clearly scope the identity claim to synthetic data and treat real-world identity as qualitative only.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: training-free multi-objective guidance with fixed empirical weights, independent losses, and external-benchmark evaluation; nothing reduces by construction to its inputs.

full rationale

PASDiff defines an energy E_total = λ_exp L_exp + λ_ref L_ref + λ_stru L_stru (Eqs. 7–12) whose three terms are constructed from independent external modules (inverse-intensity map from the input, a frozen Retinex net R, and an off-the-shelf face restorer Φ) and then used only to shift the reverse-process mean of a pre-trained unconditional DDPM. The free scalars (λ_exp=1200, λ_ref=0.03, λ_stru=10000, α=0.55, β=0.15) are stated once as empirical constants and never re-fitted to any reported metric. Restored images are scored against held-out synthetic FFHQ pairs and the newly constructed WildDark-Face set using standard external metrics (PSNR/LPIPS/Deg./LMD/MUSIQ/etc.) and an independent face-recognition model; none of those scores is algebraically forced by the loss definitions or by any self-citation. The AdaIN alignment (Eq. 10) is an explicit design choice whose effect is measured by ablation, not a tautology. No uniqueness theorem, no fitted-then-predicted quantity, and no load-bearing self-citation appear in the derivation chain. The paper is therefore free of the six circularity patterns.

Assumptions & free parameters 5 free parameters · 3 assumptions · 2 invented entities

The central empirical claim rests on a handful of hand-chosen scalars that balance the multi-objective energy, on the validity of Retinex decomposition and AdaIN as style filters, and on the assumption that the chosen off-the-shelf priors remain reliable under extreme low light. No new physical constants or particles are invented; the free parameters and domain assumptions are the main external load.

free parameters (5)
  • λ_exp = 1200
    Weight on exposure loss; set empirically to 1200 to balance magnitudes.
  • λ_ref = 0.03
    Weight on reflectance loss; set empirically to 0.03.
  • λ_stru = 10000
    Weight on structural loss; set empirically to 10000.
  • α, β (exposure map) = 0.55, 0.15
    Base exposure and amplitude for inverse-intensity map M_exp; chosen from statistical observations of low-light distributions.
  • guidance scale s and gradient steps N = N=2 (accelerated)
    Control strength and number of energy-gradient updates per reverse step; N=2 used for speed-quality trade-off.
assumptions (3)
  • domain assumption Retinex theory: image = reflectance ∘ illumination, and reflectance extracted by a pre-trained network is a reliable chromatic anchor even under extreme low light.
    Invoked in Sec. 3.2 to define L_ref; standard in low-light literature but approximate for real sensors.
  • ad hoc to paper AdaIN statistic matching fully strips photometric style while preserving high-frequency facial structure from the off-the-shelf prior.
    Core of SASI (Eq. 10); justified by the claim that identity lives in high-frequency variations.
  • domain assumption Unconditional ImageNet DDPM prior plus the chosen face-restoration and Retinex networks supply sufficient generative and structural capacity for the joint task.
    Foundation of the training-free claim; standard transfer assumption in diffusion guidance papers.
invented entities (2)
  • Style-Agnostic Structural Injection (SASI)
    purpose: Extract high-frequency facial geometry from an off-the-shelf restorer while discarding its lighting/color bias via AdaIN alignment.
    New algorithmic construct introduced in Sec. 3.2; no independent external validation beyond the paper’s ablations.
  • WildDark-Face benchmark
    purpose: Provide 700 real-world low-light face crops with compound degradations for evaluation.
    New dataset constructed by cropping DarkFace; useful but defined by the authors’ filtering choices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration." pith.science (2026). https://pith.science/paper/N7INGFV5

@misc{pith2026260324969,
  author       = {Pith},
  title        = {Pith review of: PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7INGFV5}},
  note         = {Machine review of arXiv:2603.24969}
}
read the original abstract

Face images captured in real-world low light suffer multiple degradations-low illumination, blur, noise, and low visibility, etc. Existing cascaded solutions often suffer from severe error accumulation, while generic joint models lack explicit facial priors and struggle to resolve clear face structures. In this paper, we propose PASDiff, a Physics-Aware Semantic Diffusion with a training-free manner. To achieve a plausible illumination and color distribution, we leverage inverse intensity weighting and Retinex theory to introduce photometric constraints, thereby reliably recovering visibility and natural chromaticity. To faithfully reconstruct facial details, our Style-Agnostic Structural Injection (SASI) extracts structures from an off-the-shelf facial prior while filtering out its intrinsic photometric biases, seamlessly harmonizing identity features with physical constraints. Furthermore, we construct WildDark-Face, a real-world benchmark of 700 low-light facial images with complex degradations. Extensive experiments demonstrate that PASDiff significantly outperforms existing methods, achieving a superior balance among natural illumination, color recovery, and identity consistency. Code and dataset will be available at https://github.com/IVIPLab/PASDiff.

Figures

Figures reproduced from arXiv: 2603.24969 by the authors.

Figure 1
Figure 1. Visual comparison on degraded synthetic and real-world inputs (a). Cascaded paradigms (b) L-Diff [20]→DiffBIR [35] and (c) DiffBIR [35]→L-Diff [20] suffer from severe noise amplification and structural collapse, respectively. Generic joint models like DarkIR [12] (d) and FDN [40] (e) struggle with complex degradations, leading to residual blur and detail loss. (f) Our PASDiff achieves superior perceptual quality wit… view at source ↗
Figure 2
Figure 2. From the phase reconstructions and statistical metrics, it can be seen that end-to-end learning strategies [35] inherently adhere to the degraded intensity distri￾bution, suffering from compromised facial identities and severe color shifts. Existing guidance approaches [36] correct global illumination and chromaticity, but lack seman￾tic guidance for fine geometries. In contrast, PASDiff elegantly integrates physica… view at source ↗
Figure 3
Figure 3. Overall framework of the proposed training-free PASDiff. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Left: Visualization of physical priors (Ex￾posure and Reflectance maps). Right: Subjective preferences from the user study. (a) Spatially-Varying Exposure Guidance. A critical challenge in low-light en￾hancement is the extreme dynamic range: blindly boosting global bri…
Figure 5
Figure 5. Figure 5: Visual comparison on the synthetic FFHQ dataset. low-light conditions. Following the protocol in [33, 34], the first stage simulates general face degradation: the ground truth y is sequentially degraded by Gaus￾sian blur (σ ∈ [0.1, 5]), downsampling (r ∈ [1, 4]), addit…
Figure 6
Figure 6. Figure 6: Visual comparison on the real-world WildDark-Face dataset. 4.2 Comparisons with Existing Methods Since existing methods typically treat low-light enhancement and face restora￾tion in isolation, there is currently no end-to-end solution specifically tailored for this jo…
Figure 7
Figure 7. Figure 7: Visual ablation on physical constraints [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Visual ablation on structural injection. observed, cascaded methods exhibit specific limitations depending on their ex￾ecution order. In the L-Diff [20]→DiffBIR [35] sequence, the restoration model tends to hallucinate unnatural textures based on noise amplified by the…
Figure 9
Figure 9. Figure 9: Visual ablation on style-agnostic design. scenes. Conversely, removing Lref deprives the model of a reliable chromatic anchor. The results suffer from a profound loss of chromatic information, failing to recover their intrinsic colors. This confirms that both terms are…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 5 linked inside Pith

  1. [1]

    In: ICONIP

    Bai, J., Yin, Y., He, Q., Li, Y., Zhang, X.: RetinexMamba: Retinex-based Mamba for low-light image enhancement. In: ICONIP. pp. 427–442. Springer (2024)

    2024

  2. [2]

    In: ICCV

    Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One- stage retinex-based transformer for low-light image enhancement. In: ICCV. pp. 12504–12513 (2023)

    2023

  3. [3]

    In: CVPR

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

    2018

  4. [4]

    In: CVPR

    Chung, H., Kim, J., Kim, S., Ye, J.C.: Parallel diffusion models of operator and image for blind inverse problems. In: CVPR. pp. 6059–6069 (2023)

    2023

  5. [5]

    arXiv preprint arXiv:2209.14687 (2022)

    Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sam- pling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022)

    arXiv 2022

  6. [6]

    In: ECCV

    Conde, M.V., Geigle, G., Timofte, R.: InstructIR: High-quality image restoration following human instructions. In: ECCV. pp. 1–21. Springer (2024)

    2024

  7. [7]

    In: CVPR

    Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR. pp. 4690–4699 (2019)

    2019

  8. [8]

    NeurIPS 34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS 34, 8780–8794 (2021)

    2021

Show all 58 references
  1. [9]

    IEEE transactions on pattern analysis and machine intelligence44(5), 2567–2581 (2020)

    Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unify- ing structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence44(5), 2567–2581 (2020)

    2020

  2. [10]

    In: CVPR

    Dong, L., Fan, Q., Guo, Y., Wang, Z., Zhang, Q., Chen, J., Luo, Y., Zou, C.: TSD- SR: One-step diffusion with target score distillation for real-world image super- resolution. In: CVPR. pp. 23174–23184 (2025)

    2025

  3. [11]

    In: CVPR

    Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Gen- erative diffusion prior for unified image restoration and enhancement. In: CVPR. pp. 9935–9946 (2023)

    2023

  4. [12]

    In: CVPR

    Feijoo, D., Benito, J.C., Garcia, A., Conde, M.V.: DarkIR: Robust low-light image restoration. In: CVPR. pp. 10879–10889 (2025)

    2025

  5. [13]

    In: CVPR

    Fu, Z., Yang, Y., Tu, X., Huang, Y., Ding, X., Ma, K.K.: Learning a simple low- light image enhancer from paired low-light instances. In: CVPR. pp. 22252–22261 (2023) 16 Yilin Ni et al

    2023

  6. [14]

    In: CVPR

    Garber, T., Tirer, T.: Image restoration by denoising diffusion models with itera- tively preconditioned guidance. In: CVPR. pp. 25245–25254 (2024)

    2024

  7. [15]

    In: CVPR

    Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: CVPR. pp. 1780–1789 (2020)

    2020

  8. [16]

    IEEE Transactions on Image Processing26(2), 982–993 (2016)

    Guo, X., Li, Y., Ling, H.: LIME: Low-light image enhancement via illumination map estimation. IEEE Transactions on Image Processing26(2), 982–993 (2016)

    2016

  9. [17]

    He, C., Fang, C., Zhang, Y., Li, K., Tang, L., You, C., Xiao, F., Guo, Z., Li, X.: Reti-Diff: Illumination degradation image restoration with Retinex-based Latent Diffusion Model (2025)

    2025

  10. [18]

    In: NeurIPS

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS. vol. 30 (2017)

    2017

  11. [19]

    NeurIPS33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS33, 6840–6851 (2020)

    2020

  12. [20]

    In: ECCV

    Jiang, H., Luo, A., Liu, X., Han, S., Liu, S.: Lightendiffusion: Unsupervised low- light image enhancement with latent-retinex diffusion models. In: ECCV. pp. 161–

  13. [21]

    IEEE Transactions on Image Processing6(7), 965–976 (1997)

    Jobson, D.J., Rahman, Z.u., Woodell, G.A.: A multiscale Retinex for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image Processing6(7), 965–976 (1997)

    1997

  14. [22]

    NeurIPS35, 23593–23606 (2022)

    Kawar, B., Elad, M., Ermon, S., Song, J.: Denoising diffusion restoration models. NeurIPS35, 23593–23606 (2022)

    2022

  15. [23]

    In: ICCV

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV. pp. 5148–5157 (2021)

    2021

  16. [24]

    arXiv preprint arXiv:1908.08239 (2019)

    Kim, D., Kim, M., Kwon, G., Kim, D.S.: Progressive face super-resolution via attention to facial landmark. arXiv preprint arXiv:1908.08239 (2019)

    1908 arXiv

  17. [25]

    arXiv preprint arXiv:2503.08136 (2025)

    Kim, J., Kim, B.S., Ye, J.C.: FlowDPS: Flow-driven posterior sampling for inverse problems. arXiv preprint arXiv:2503.08136 (2025)

    2025 arXiv

  18. [26]

    In: AAAI

    Lan, Y., Cui, Z., Liu, C., Peng, J., Wang, N., Luo, X., Liu, D.: Exploiting diffusion prior for real-world image dehazing with unpaired training. In: AAAI. vol. 39, pp. 4455–4463 (2025)

    2025

  19. [27]

    Scientific american237(6), 108–129 (1977)

    Land, E.H.: The retinex theory of color vision. Scientific american237(6), 108–129 (1977)

    1977

  20. [28]

    In: CVPR

    Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452–17462 (2022)

    2022

  21. [29]

    arXiv preprint arXiv:2207.14273 (2022)

    Li, C., Guo, C., Feng, R., Zhou, S., Loy, C.C.: Cudi: Curve distillation for efficient and controllable exposure adjustment. arXiv preprint arXiv:2207.14273 (2022)

    2022 arXiv

  22. [30]

    In: ACM MM

    Li, W., Guo, H., Liu, X., Liang, K., Hu, J., Ma, Z., Guo, J.: Efficient face super- resolution via wavelet-based feature enhancement network. In: ACM MM. pp. 4515–4523 (2024)

    2024

  23. [31]

    ACM Computing Surveys (2025)

    Li, W., Wang, M., Zhang, K., Li, J., Li, X., Zhang, Y., Gao, G., Ma, Z.: Survey on deep face restoration: From non-blind to blind and beyond. ACM Computing Surveys (2025)

    2025

  24. [32]

    In: NeurIPS (2025)

    Li, W., Wang, X., Guo, H., Gao, G., Ma, Z.: Self-supervised selective-guided dif- fusion model for old-photo face restoration. In: NeurIPS (2025)

    2025

  25. [33]

    In: ECCV

    Li, X., Chen, C., Zhou, S., Lin, X., Zuo, W., Zhang, L.: Blind face restoration via deep multi-scale component dictionaries. In: ECCV. pp. 399–415. Springer (2020)

    2020

  26. [34]

    In: ECCV

    Li, X., Liu, M., Ye, Y., Zuo, W., Lin, L., Yang, R.: Learning warped guidance for blind face restoration. In: ECCV. pp. 272–289 (2018) Physics-Aware Semantic Diffusion 17

    2018

  27. [35]

    In: ECCV

    Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Qiao, Y., Ouyang, W., Dong, C.: DiffBIR: Toward blind image restoration with generative diffusion prior. In: ECCV. pp. 430–448. Springer (2024)

    2024

  28. [36]

    In: AAAI

    Lin, Y., Ye, T., Chen, S., Fu, Z., Wang, Y., Chai, W., Xing, Z., Li, W., Zhu, L., Ding, X.: Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. In: AAAI. vol. 39, pp. 5307–5315 (2025)

    2025

  29. [37]

    IEEE Transactions on Circuits and Systems for Video Technology35(7), 6602–6615 (2025)

    Liu, M., Cui, Y., Ren, W., Zhou, J., Knoll, A.C.: Liednet: A lightweight network for low-light enhancement and deblurring. IEEE Transactions on Circuits and Systems for Video Technology35(7), 6602–6615 (2025)

    2025

  30. [38]

    In: NeurIPS

    Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: PromptIR: Prompting for all- in-one blind image restoration. In: NeurIPS. vol. 36, pp. 24706–24746 (2023)

    2023

  31. [39]

    In: CVPR

    Su, S., Yan, Q., Zhu, Y., Zhang, C., Ge, X., Sun, J., Zhang, Y.: Blindly assess image quality in the wild guided by a self-adaptive hyper network. In: CVPR. pp. 3660–3669 (2020)

    2020

  32. [40]

    IEEE Transactions on Image Processing34, 5184–5199 (2025)

    Tu, L., Wu, J., Wang, C., Meng, D., Jin, Z.: Fourier-based decoupling network for joint low-light image enhancement and deblurring. IEEE Transactions on Image Processing34, 5184–5199 (2025)

    2025

  33. [41]

    In: CVPR

    Wang, J., Gong, J., Zhang, L., Chen, Z., Liu, X., Gu, H., Liu, Y., Zhang, Y., Yang, X.: Osdface: One-step diffusion model for face restoration. In: CVPR. pp. 12626–12636 (2025)

    2025

  34. [42]

    Pattern Recognition166, 111628 (2025)

    Wang, T., Zhang, K., Zhang, Y., Luo, W., Stenger, B., Lu, T., Kim, T.K., Liu, W.: LLDiffusion: Learning degradation representations in diffusion models for low-light image enhancement. Pattern Recognition166, 111628 (2025)

    2025

  35. [43]

    In: CVPR

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

    2021

  36. [44]

    In: ICCV

    Wang, X., Bo, L., Fuxin, L.: Adaptive wing loss for robust face alignment via heatmap regression. In: ICCV. pp. 6971–6981 (2019)

    2019

  37. [45]

    In: ICLR (2023)

    Wang, Y., Yu, J., Zhang, J.: Zero-Shot image restoration using Denoising Diffusion Null-Space Model. In: ICLR (2023)

    2023

  38. [46]

    In: AAAI

    Wang, Y., Wan, R., Yang, W., Li, H., Chau, L.P., Kot, A.: Low-light image en- hancement with normalizing flow. In: AAAI. vol. 36, pp. 2604–2612 (2022)

    2022

  39. [47]

    arXiv preprint arXiv:1808.04560 (2018)

    Wei, C., Wang, W., Yang, W., Liu, J.: Deep Retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560 (2018)

    2018 arXiv

  40. [48]

    Wu, W., Weng, J., Zhang, P., Wang, X., Yang, W., Jiang, J.: URetinex-Net: Retinex-baseddeepunfoldingnetworkforlow-lightimageenhancement.In:CVPR. pp. 5901–5910 (2022)

    2022

  41. [49]

    In: CVPR

    Xu, R., Niu, Y., Li, Y., Xu, H., Liu, W., Chen, Y.: Urwkv: Unified rwkv model with multi-state perspective for low-light image restoration. In: CVPR. pp. 21267–21276 (2025)

    2025

  42. [50]

    In: CVPR

    Yan, Q., Feng, Y., Zhang, C., Pang, G., Shi, K., Wu, P., Dong, W., Sun, J., Zhang, Y.: HVI: A new color space for low-light image enhancement. In: CVPR. pp. 5678– 5687 (2025)

    2025

  43. [51]

    In: CVPR

    Yang,S.,Wu,T.,Shi,S.,Lao,S.,Gong,Y.,Yuan,M.,Wang,M.,Yang,Y.:Maniqa: Multi-dimension attention network for no-reference image quality assessment. In: CVPR. pp. 1191–1200 (2022)

    2022

  44. [52]

    IEEE Transactions on Image Processing29, 5737–5752 (2020)

    Yang, W., Yuan, Y., Ren, W., Liu, J., Scheirer, W.J., Wang, Z., et al.: Advancing imageunderstandinginpoorvisibilityenvironments:Acollectivebenchmarkstudy. IEEE Transactions on Image Processing29, 5737–5752 (2020)

    2020

  45. [53]

    In: ICCV

    Yu, J., Wang, Y., Zhao, C., Ghanem, B., Zhang, J.: Freedom: Training-free energy- guided conditional diffusion model. In: ICCV. pp. 23174–23184 (2023) 18 Yilin Ni et al

    2023

  46. [54]

    In: CVPR

    Zhang, B., Chu, W., Berner, J., Meng, C., Anandkumar, A., Song, Y.: Improving diffusion inverse problem solving with decoupled noise annealing. In: CVPR. pp. 20895–20905 (2025)

    2025

  47. [55]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

    2018

  48. [56]

    NeurIPS35, 30599–30611 (2022)

    Zhou, S., Chan, K., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. NeurIPS35, 30599–30611 (2022)

    2022

  49. [57]

    In: ECCV

    Zhou, S., Li, C., Change Loy, C.: LEDNet: Joint low-light enhancement and de- blurring in the dark. In: ECCV. pp. 573–589. Springer (2022)

    2022

  50. [58]

    In: AAAI

    Zou, W., Gao, H., Ye, T., Chen, L., Yang, W., Huang, S., Chen, H., Chen, S.: Vqcnir: clearer night image restoration with vector-quantized codebook. In: AAAI. vol. 38, pp. 7873–7881 (2024) PASDiff: Physics-Aware Semantic Guidance for Joint Real-world Low-Light Face Enhancement...

    2024

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.