Pith. sign in

REVIEW 4 major objections 5 minor 43 references

A causal denoiser that separates content from noise in a ViT.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 19:42 UTC pith:SZN4F45Q

load-bearing objection A competent ViT denoiser whose small empirical wins are oversold by causal language the method never actually instantiates. the 4 major comments →

arxiv 2603.01140 v2 pith:SZN4F45Q submitted 2026-03-01 cs.CV

Teacher-Guided Causal Interventions for Image Denoising: Orthogonal Content-Noise Disentanglement in Vision Transformers

Kuai Jiang , Zhaoyan Ding , Guijuan Zhang , Dianjie Lu , Zhuoran Zheng This is my paper
classification cs.CV
keywords image denoisingcausal interventionvision transformercontent-noise disentanglementorthogonality constraintnoise supervisionteacher guidancereal-time inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TCD-Net treats denoising as causal intervention rather than correlation fitting, modeling noisy images as generated by content, noise, and environment. It uses an EBA module to remove global environmental bias, a dual-branch head with an orthogonality constraint to force content and noise representations into perpendicular subspaces, strong noise supervision to anchor the noise branch, and teacher guidance from Google's Nano Banana Pro to keep content on the natural-image manifold. The paper reports state-of-the-art or competitive fidelity on CBSD68, SIDD, DND, and other benchmarks while running at 104.2 FPS on an RTX 5090. A sympathetic reader would care because this is an attempt to make denoising robust to high-frequency texture-vs-grain ambiguity and camera/lighting shifts without extra inference cost.

Core claim

The core claim is that explicit, mechanism-level disentanglement - not just architectural capacity - resolves the content-noise ambiguity in denoising. TCD-Net decomposes the generative process by learning content features Zc and noise features Zn from shared encoder features, enforcing orthogonality between them per token, and supervising the noise branch directly with the residual between noisy and clean images. The teacher prior from Nano Banana Pro is distilled at the feature level during training only, pulling predicted content toward natural-image statistics while adding zero inference overhead. The paper argues this combination suppresses spurious environment-content correlations, red

What carries the argument

The orthogonality constraint (Eq. 10) is the central mechanism: it forces the content and noise feature vectors to be perpendicular at every token, acting as a geometric firewall against information leakage between the two branches. It is supported by three auxiliary mechanisms: (1) Environmental Bias Adjustment (EBA), a per-token de-centering plus bottleneck-MLP residual that suppresses global bias; (2) strong noise supervision with ground-truth noise N_gt = Y - X, which anchors the noise branch and prevents degenerate zero-noise solutions; (3) feature-level teacher guidance from Nano Banana Pro, which regularizes the content branch toward the natural-image manifold. The serial topology of

Load-bearing premise

The load-bearing premise is that content and noise can be strictly separated into orthogonal subspaces, even though the paper's own generative model says noise depends on content (signal-dependent noise); if real sensor noise carries content information, the orthogonality constraint may force the noise branch to discard exactly the information it needs.

What would settle it

A direct test: train TCD-Net on noisy images with known Poisson or Poisson-Gaussian noise (signal-dependent), then measure the noise-prediction error and the angle between Z_c and Z_n at convergence. If the orthogonality constraint forces |⟨Z_c, Z_n⟩| near zero while noise prediction accuracy degrades compared to a model without the constraint, the central claim of universal separation is falsified. Alternatively, inspect the learned projection matrices W_c and W_n: if they are not mutually orthogonal (or nearly so) at convergence, the penalty is being satisfied by a trivial solution (e.g., co

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, denoising models can be made more robust to camera/ISP shifts by explicitly factoring out environmental bias, without requiring the model to memorize nuisance correlations.
  • Real-time deployment is feasible: 104.2 FPS on an RTX 5090 with a straightforward ViT-L backbone and no sampling loop.
  • The explicit noise map output enables downstream tasks like noise-level estimation or noise-aware processing as a byproduct.
  • Teacher-driven priors can be injected at training time only, improving perceptual quality with zero test-time cost, suggesting a general recipe for using generative models as teachers for restoration networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The orthogonality constraint may be too strict for real sensor noise, which is signal-dependent (as the paper's own SCM in Eq. 1 admits: N depends on C). A testable prediction is that enforcing perfect perpendicularity between content and noise features will hurt noise estimation on Poisson or Poisson-Gaussian noise; a relaxed or conditional orthogonality may perform better.
  • The teacher guidance from Nano Banana Pro introduces a dependency on a proprietary model; if that teacher is unavailable or changes, the pipeline's training is affected. One could test substituting a public diffusion model as teacher to see whether the gains are teacher-specific or generic to manifold regularization.
  • The reported FPS is measured on a single RTX 5090 with batch 1 at 256x256; real-world throughput on other hardware and larger resolutions will vary. The paper does not report latency for its overlap-tile inference on high-resolution inputs.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes TCD-Net, a Vision-Transformer-based image denoiser that claims to explicitly disentangle image content from noise via three components: an Environmental Bias Adjustment (EBA) module for de-confounding, a dual-branch head with an orthogonality constraint for content–noise separation, and a teacher-guided prior using Google's Nano Banana Pro (NBP). The method is trained with reconstruction, noise-anchoring, orthogonality, and teacher losses, and is evaluated on synthetic Gaussian denoising (CBSD68, Kodak24, McMaster, Urban100) and real-world denoising (SIDD, DND), with an emphasis on speed (104.2 FPS on RTX 5090). The paper reports strong efficiency and modest fidelity gains, and provides an ablation study of the proposed components.

Significance. If substantiated, the causal-intervention formulation and orthogonal content–noise separation could offer improved robustness to distribution shifts and better handling of high-frequency texture/noise ambiguity, with no added inference cost. The paper provides a concrete architecture, transparent experimental details, and systematic ablations that are generally reproducible in structure. However, the central causal claims are not fully operationalized, and the fidelity advantages over strong baselines are small and inconsistent across benchmarks. The main potential contribution is the combination of explicit noise supervision with a disentanglement regularizer in an efficient ViT denoiser, but this must be validated more carefully before the 'causal intervention' framing is credible.

major comments (4)
  1. [§3.1 Eq. (1) vs. §3.3 Eq. (10)] The SCM in Eq. (1) defines N := f_N(C, E, U_N), making noise depend on content. Eq. (10) penalizes the normalized inner product between content features z_c and noise features z_n, which the text calls a 'strict separation' preventing leakage. If real noise is signal-dependent, an accurate noise estimator must encode content information; the orthogonality constraint would then force the noise branch to discard exactly the information it needs, degrading L_noise, or the constraint is satisfied only weakly, undermining the 'strict separation' claim. No experiment arbitrates this tension: there is no test with known signal-dependent noise, no measurement of content recoverable from z_n, and no ablation reporting L_noise with and without L_ortho. This is load-bearing for the paper's central claim of explicit generative-mechanism decomposition.
  2. [§4.2, Table 1] The abstract states TCD-Net 'outperforms mainstream methods across multiple benchmarks in both fidelity and efficiency,' but Table 1 shows TCD-Net is best on only 6 of 12 synthetic cells, with gains of at most 0.08 dB, and on Kodak24 it trails HAT and MambaIRv2 at every noise level. This overstatement should be corrected, and the text should contextualize the results as competitive or state-of-the-art on a subset of benchmarks, not a blanket claim.
  3. [§4.1, DND fine-tuning] The text says for real-world denoising the authors fine-tune on SIDD/DND 'using dataset-provided noisy/clean pairs for supervision.' This is problematic for DND, which does not publicly release clean ground truth; the benchmark typically requires submitting denoised images to an online server for evaluation. If the authors used a non-public protocol or a distinct training split, they must specify exactly how DND pairs were obtained. Without clarification, the DND PSNR/SSIM results in Table 2 are unsupported.
  4. [§3.3, general causal framing] The paper repeatedly describes EBA, orthogonality, and teacher guidance as 'causal interventions' and 'de-confounding,' but no formal intervention (e.g., a do-operator or counterfactual operation) is defined or implemented. These are architectural regularizers and losses. The causal SCM is used only as motivation. For these claims to be credible, the authors should either show formally how each module approximates an intervention on Eq. (1), or temper the causal language to 'causal-inspired' regularization.
minor comments (5)
  1. [§3.2] The term 'JiT-style ViT' is not defined anywhere; please clarify or remove.
  2. [§4.2, Table 5] The notation '(Indep.)' in Table 5 is not defined in the ablation description. Explain what 'independent' refers to in the orthogonality regularizer.
  3. [§3.4, Eq. (14)] The epsilon in the denominator of Eq. (14) is introduced as 'ε' but not specified; please define it.
  4. [§4.1] The two-stage schedule includes many hyperparameters, but the teacher-sample fraction is never stated. Please provide the sampling ratio for the NBP teacher.
  5. [Throughout] The reliance on a proprietary model, Nano Banana Pro, as a teacher raises reproducibility concerns. If possible, discuss alternatives or provide a version with a public teacher in the ablation.

Circularity Check

0 steps flagged

No circularity found: the empirical pipeline is externally benchmarked; the causal framing is a modeling/regularization choice, not a derivation from its own output.

full rationale

I walked the claimed derivation chain. The paper's SCM in Eq. (1) is a modeling assumption, Eq. (10) is a regularizer, and Eq. (13) is a supervised training objective; none of these defines a benchmark result as a function of a fitted parameter. The central empirical claims (PSNR, SSIM, LPIPS, FPS) are measured against external benchmarks (CBSD68, Kodak24, McMaster, Urban100, SIDD, DND) and compared with published baselines, so they do not reduce to the method's own inputs by construction. The NBP teacher is an external prior cited from independent work and is applied only during training; its ablation contribution is +0.03 dB, so it is not load-bearing. There are no author-only uniqueness theorems or self-citation chains invoked to force the architecture. Possible concerns, such as the tension between the signal-dependent noise in Eq. (1) and strict orthogonality in Eq. (10), or the §4.1 wording about DND ground-truth pairs, are correctness/validity issues rather than circularity: no result is equivalent by construction to its inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The causal variables C, E, N, Y in Eq (1) are asserted, not derived; EBA, orthogonality, and teacher losses are hand-weighted regularizers whose strengths (λ_noise 0.5→0.25, λ_ortho 0.1→0.05, λ_teacher 0.1) are tuned, not predicted. The only external input is the NBP teacher, whose sampling fraction and prompt are unspecified. No new physical entities are introduced; the latent environment confounder E has no identifiable, falsifiable handle in the paper.

free parameters (6)
  • lambda_noise (noise-supervision weight) = 0.5 (stage 1), 0.25 (stage 2)
    Weight of L_noise in Eq (13a); chosen by hand; anchors the noise branch and affects both outputs.
  • lambda_ortho (orthogonality weight) = 0.1 (stage 1), 0.05 (stage 2)
    Weight of L_ortho (Eq 10); hand-chosen; ablation shows topology/ordering matters (±0.17 dB), but strength is not swept.
  • lambda_teacher (NBP guidance weight) = 0.1 (stage 2 only)
    Weight of L_teacher (Eq 12); contributes +0.03 dB in ablation (Table 5), within likely seed noise.
  • teacher-sample fraction = unspecified ('for some samples')
    Fraction of training samples receiving NBP supervision is never stated; an uncontrolled calibration parameter that affects every reported number.
  • two-stage schedule (epochs/lr) = 100 ep @ 2e-4 + 50 ep @ 5e-5
    Hand-chosen schedule; not derived; stage-2 weight changes were tuned together with teacher.
  • Gaussian blending std for overlap tiles = P/4
    Eq (14) blending width chosen by hand; affects high-resolution inference quality.
axioms (4)
  • domain assumption Generative SCM: C := f_C(U_C), E := f_E(U_E), N := f_N(C,E,U_N), Y := f_Y(C,N,E)
    Eq (1) — the entire causal framing and the EBA/orthogonality design target this model, but it is asserted, never tested or identified.
  • domain assumption NBP outputs X_T are useful perceptual targets for denoising
    §3.3 relies on [Zuo et al., 2025] and 'prior evaluations' for the claim that NBP gives high-quality zero-shot restorations; the paper also admits outputs 'may contain plausible but input-inconsistent details,' a tension it hedges by using features only.
  • domain assumption Pixel-space losses alone cannot resolve high-frequency texture/noise ambiguity
    Motivation in §1/§3.3; standard in perceptual-restoration literature, adopted here as justification for L_teacher.
  • ad hoc to paper Orthogonal content/noise subspaces aid, not harm, denoising
    Eq (10)'s 'geometric firewall' assumes separability that the paper's own SCM (N depends on C) contradicts for signal-dependent noise; no experiment validates the assumption.
invented entities (1)
  • Environment factor E (confounder) no independent evidence
    purpose: Latent cause of domain shift in the SCM (Eq 1); the target EBA claims to de-confound.
    No measurement, identifiability argument, or experiment pins down E; EBA's de-centering (Eq 5) subtracts the token-wise channel mean, which has no established link to E, so the entity has no falsifiable handle in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 12230 in / 24868 out tokens · 215553 ms · 2026-08-02T19:42:33.972884+00:00 · methodology

0 comments
read the original abstract

Conventional image denoising models often inadvertently learn spurious correlations between environmental factors and noise patterns. Moreover, due to high-frequency ambiguity, they struggle to reliably distinguish subtle textures from stochastic noise, resulting in over-removed details or residual noise artifacts. We therefore revisit denoising via causal intervention, arguing that purely correlational fitting entangles intrinsic content with extrinsic noise, which directly degrades robustness under distribution shifts. Motivated by this, we propose the Teacher-Guided Causal Disentanglement Network (TCD-Net), which explicitly decomposes the generative mechanism via structured interventions on feature spaces within a Vision Transformer framework. Specifically, our method integrates three key components: (1) An Environmental Bias Adjustment (EBA) module projects features into a stable, de-centered subspace to suppress global environmental bias (de-confounding). (2) A dual-branch disentanglement head employs an orthogonality constraint to force a strict separation between content and noise representations, preventing information leakage. (3) To resolve structural ambiguity, we leverage Nano Banana Pro, Google's reasoning-guided AI image generation model, to guide a causal prior, effectively pulling content representations back onto the natural-image manifold. Extensive experiments demonstrate that TCD-Net outperforms mainstream methods across multiple benchmarks in both fidelity and efficiency, achieving a real-time speed of 104.2 FPS on a single RTX 5090 GPU.

Figures

Figures reproduced from arXiv: 2603.01140 by Dianjie Lu, Guijuan Zhang, Kuai Jiang, Zhaoyan Ding, Zhuoran Zheng.

Figure 1
Figure 1. Figure 1: Efficiency–performance trade-off. PSNR vs. FPS on CBSD68 (σ=15) and SIDD. TCD-Net achieves the best speed– quality trade-off among competing methods. fine textures and noise often appear as high-frequency sig￾nals. Consequently, data-driven denoisers may exploit spuri￾ous correlations between environment factors and noise pat￾terns and may fail to reliably separate subtle textures from random noise, leadin… view at source ↗
Figure 2
Figure 2. Figure 2: Conceptual comparison between conventional denoising and TCD-Net. (a) Conventional denoising is prone to spurious content–noise correlations induced by environmental factors E, leaving residual artifacts. (b) TCD-Net performs causal intervention via EBA, incorporates teacher semantic guidance, and enforces an orthogonality constraint to decouple content and noise for cleaner restoration. 3.1 Causal View an… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of TCD-Net. A ViT backbone with EBA feeds a dual-branch head to predict the restored image Xˆ and noise map Nˆ, trained with orthogonality, noise anchoring, and teacher guidance [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: EBA module. LayerNorm + bottleneck MLP with resid￾ual projection to suppress environment-induced bias and stabilize token representations. 3.3 Causal Interventions: De-confounding, Disentanglement, and NBP Guidance Environmental Bias Adjustment (EBA) de-confounding. Real noisy images often contain global appearance shifts (e.g., illumination/color temperature) that act as environment￾induced confounders E.… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on synthetic and real noise. (a) Synthetic Gaussian denoising (AWGN). (b) Real-world denoising on SIDD. Conventional methods may leave residual noise or oversmooth details due to spurious content–noise correlation. With EBA-based intervention, orthogonal disentanglement, and teacher guidance, TCD-Net restores cleaner results with sharper textures and edges. Net achieves competitive p… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 9 linked inside Pith

  1. [1]

    [Abdelhamedet al., 2018 ] Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692–1700,

    2018

  2. [6]

    Hat: Hybrid attention transformer for image restoration.arXiv preprint arXiv:2309.05239,

    [Chenet al., 2023 ] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration.arXiv preprint arXiv:2309.05239,

    arXiv 2023

  3. [8]

    Image denoising by sparse 3-d transform-domain collaborative filtering.IEEE Transactions on Image Processing, 16(8):2080–2095,

    [Dabovet al., 2007 ] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering.IEEE Transactions on Image Processing, 16(8):2080–2095,

    2007

  4. [13]

    Mambair: A sim- ple baseline for image restoration with state-space model

    [Guo and others, 2024] Hang Guo et al. Mambair: A sim- ple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648,

    Pith/arXiv arXiv 2024

  5. [15]

    Mambairv2: Attentive state space restoration

    [Guoet al., 2025 ] Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. Mambairv2: Attentive state space restoration. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28124–28133, June

    2025

  6. [16]

    Distilling the knowledge in a neural net- work.arXiv preprint arXiv:1503.02531,

    [Hintonet al., 2015 ] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural net- work.arXiv preprint arXiv:1503.02531,

    Pith/arXiv arXiv 2015

  7. [18]

    Perceptual losses for real-time style transfer and super-resolution

    [Johnsonet al., 2016 ] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. InComputer Vision – ECCV 2016, pages 694–711,

    2016

  8. [20]

    SwinIR: Image restoration using swin transformer,

    [Lianget al., 2021 ] Jingyun Liang, Jiezhang Cao, Guo- Cheng Wang, Kai Zhang, Lei Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer,

    2021

  9. [21]

    Enhanced deep residual networks for single image super-resolution

    [Limet al., 2017 ] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1132– 1140,

    2017

  10. [22]

    Diffbir: Towards blind image restoration with generative diffusion prior.arXiv preprint arXiv:2308.15070,

    [Linet al., 2024 ] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior.arXiv preprint arXiv:2308.15070,

    Pith/arXiv arXiv 2024

  11. [23]

    Swin transformer: Hierarchical vision transformer using shifted windows

    [Liuet al., 2021 ] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV), pages 10012–10022,

    2021

  12. [24]

    SGDR: Stochastic gradient descent with warm restarts

    [Loshchilov and Hutter, 2017] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. InInternational Conference on Learning Repre- sentations (ICLR),

    2017

  13. [25]

    [Loshchilov and Hutter, 2019] Ilya Loshchilov and Frank Hutter

    arXiv:1608.03983. [Loshchilov and Hutter, 2019] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInter- national Conference on Learning Representations (ICLR),

    Pith/arXiv arXiv 2019

  14. [26]

    [Mouet al., 2021 ] Chong Mou, Jian Zhang, and Zhuoyuan Wu

    arXiv:1711.05101. [Mouet al., 2021 ] Chong Mou, Jian Zhang, and Zhuoyuan Wu. Dynamic attentive graph learning for image restora- tion. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), pages 4328–4337, October

    Pith/arXiv arXiv 2021

  15. [27]

    Cambridge University Press, 2 edition,

    [Pearl, 2009] Judea Pearl.Causality: Models, Reasoning, and Inference. Cambridge University Press, 2 edition,

    2009

  16. [29]

    Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634,

    [Sch¨olkopfet al., 2021 ] Bernhard Sch¨olkopf, Francesco Lo- catello, Stefan Bauer, Nan Rosemary Ke, Nal Kalch- brenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634,

    2021

  17. [30]

    Vmambair: Visual state space model for image restoration

    [Shiet al., 2024 ] Yuan Shi, Bin Xia, Xiaoyu Jin, Xing Wang, Tianyu Zhao, Xin Xia, Xuefeng Xiao, and Wenming Yang. Vmambair: Visual state space model for image restoration. arXiv preprint arXiv:2403.11423,

    Pith/arXiv arXiv 2024

  18. [31]

    Deep image prior

    [Ulyanovet al., 2018 ] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. InProceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 9446–9454,

    2018

  19. [32]

    Bovik, Hamid R

    [Wanget al., 2004 ] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Trans- actions on Image Processing, 13(4):600–612,

    2004

  20. [34]

    Image blind de- noising using dual convolutional neural network with skip connection.Signal Processing: Image Communication, 138:117365,

    [Wuet al., 2025 ] Wencong Wu, Shicheng Liao, Guannan Lv, Peng Liang, and Yungang Zhang. Image blind de- noising using dual convolutional neural network with skip connection.Signal Processing: Image Communication, 138:117365,

    2025

  21. [35]

    Diffir: Efficient diffusion model for image restoration

    [Xiaet al., 2023 ] Bin Xia, Yulun Zhang, Shiyin Wang, Yi- tong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV),

    2023

  22. [36]

    Multi-stage progres- sive image restoration

    [Zamiret al., 2021 ] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progres- sive image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821–14831,

    2021

  23. [37]

    Restormer: Efficient transformer for high-resolution image restoration

    [Zamiret al., 2022 ] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728–5739,

    2022

  24. [38]

    Color demosaicking by local directional in- terpolation and nonlocal adaptive thresholding.Journal of Electronic Imaging, 20(2):023016,

    [Zhanget al., 2011 ] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional in- terpolation and nonlocal adaptive thresholding.Journal of Electronic Imaging, 20(2):023016,

    2011

  25. [40]

    Plug-and- play image restoration with deep denoiser prior.arXiv preprint arXiv:2008.13751,

    [Zhanget al., 2021 ] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and- play image restoration with deep denoiser prior.arXiv preprint arXiv:2008.13751,

    Pith/arXiv arXiv 2021

  26. [41]

    Practical blind image denoising via swin-conv-UNet and data synthesis.Ma- chine Intelligence Research,

    [Zhanget al., 2023 ] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yulun Zhang, Hao Tang, Deng-Ping Fan, Radu Timofte, and Luc Van Gool. Practical blind image denoising via swin-conv-UNet and data synthesis.Ma- chine Intelligence Research,

    2023

  27. [42]

    Denoising diffusion models for plug-and-play im- age restoration.arXiv preprint arXiv:2305.08995,

    [Zhuet al., 2023 ] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play im- age restoration.arXiv preprint arXiv:2305.08995,

    Pith/arXiv arXiv 2023

  28. [43]

    Is nano banana pro a low-level vision all-rounder? a comprehensive eval- uation on 14 tasks and 40 datasets.arXiv preprint arXiv:2512.15110, 2025

    [Zuoet al., 2025 ] Jialong Zuo, Haoyou Deng, Hanyu Zhou, Jiaxin Zhu, Yicheng Zhang, Yiwei Zhang, Yongxin Yan, Kaixing Huang, Weisen Chen, Yongtai Deng, Rui Jin, Nong Sang, and Changxin Gao. Is nano banana pro a low-level vision all-rounder? a comprehensive eval- uation on 14 tasks and 40 datasets.arXiv preprint arXiv:2512.15110, 2025

    arXiv 2025

  29. [1999]

    [Google DeepMind, 2025] Google DeepMind

    Accessed: 2026-01-10. [Google DeepMind, 2025] Google DeepMind. Introducing nano banana pro. https://blog.google/innovation-and-ai/ products/nano-banana-pro/, November

    2026

  30. [2004]

    Uformer: A general u-shaped transformer for image restoration

    [Wanget al., 2022 ] Zhendong Wang, Xiaodong Cun, Jian- min Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683–17693,

    2022

  31. [2007]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    [Dosovitskiyet al., 2021 ] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Min- derer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognition at scale. InInternational Conference on Learning Re...

    2021

  32. [2009]

    Bench- marking denoising algorithms with real photographs

    [Pl¨otz and Roth, 2017] Tobias Pl¨otz and Stefan Roth. Bench- marking denoising algorithms with real photographs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2750–2759,

    2017

  33. [2011]

    Beyond a gaus- sian denoiser: Residual learning of deep CNN for image denoising

    [Zhanget al., 2017 ] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaus- sian denoiser: Residual learning of deep CNN for image denoising. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3929–3937,

    2017

  34. [2015]

    Single image super-resolution from transformed self-exemplars

    [Huanget al., 2015 ] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197–5206, June

    2015

  35. [2016]

    Noise2void: Learning denoising from single noisy images

    [Krullet al., 2019 ] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void: Learning denoising from single noisy images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2129–2137,

    2019

  36. [2017]

    Real image denoising with feature attention

    [Anwar and Barnes, 2019] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 3155–3164,

    2019

  37. [2018]

    NTIRE 2017 challenge on single image super- resolution: Dataset and study

    [Agustsson and Timofte, 2017] Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super- resolution: Dataset and study. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops,

    2017

  38. [2019]

    Contour detection and hierarchical image segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):898– 916,

    [Arbel´aezet al., 2011 ] Pablo Arbel ´aez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):898– 916,

    2011

  39. [2021]

    Kodak lossless true color im- age suite

    [Franzen, 1999] Rich Franzen. Kodak lossless true color im- age suite. http://r0k.us/graphics/kodak/,

    1999

  40. [2022]

    Simple baselines for image restora- tion

    [Chenet al., 2022b ] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restora- tion. InComputer Vision – ECCV 2022, pages 17–33,

    2022

  41. [2023]

    Conditional positional encodings for vi- sion transformers

    [Chuet al., 2023 ] Xiaojie Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Conditional positional encodings for vi- sion transformers. InInternational Conference on Learn- ing Representations (ICLR),

    2023

  42. [2024]

    Toward convolutional blind denoising of real photographs

    [Guoet al., 2019 ] Shi Guo, Zifei Yan, Kai Zhang, Wang- meng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1712–1722,

    2019

  43. [2025]

    Mamba: Linear- time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752,

    [Gu and Dao, 2023] Albert Gu and Tri Dao. Mamba: Linear- time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752,

    Pith/arXiv arXiv 2023