pith. machine review for the scientific record. sign in

arxiv: 2604.15654 · v1 · submitted 2026-04-17 · 💻 cs.CV

Recognition: unknown

From Zero to Detail: A Progressive Spectral Decoupling Paradigm for UHD Image Restoration with New Benchmark

Chen Zhao , Yunzhe Xu , Zhizhou Chen , Enxuan Gu , Kai Zhang , Xiaoming Liu , Jian Yang , Ying Tai
Authors on Pith no claims yet

Pith reviewed 2026-05-10 09:21 UTC · model grok-4.3

classification 💻 cs.CV
keywords UHD image restorationspectral decouplingprogressive restorationKolmogorov-Arnold Networkimage benchmarkfrequency enhancementdetail refinementERR framework
0
0 comments X

The pith

A progressive spectral decoupling into zero-, low-, and high-frequency stages with cooperative sub-networks achieves superior UHD image restoration.

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

The paper establishes that UHD image restoration benefits from decomposing the process into three frequency-specific stages rather than a single end-to-end network. It introduces the ERR framework, which deploys a zero-frequency enhancer to capture global priors, a low-frequency restorer for main content structures, and a high-frequency refiner that uses a frequency-windowed Kolmogorov-Arnold Network to recover fine textures. The work also supplies a new benchmark dataset of 82,126 diverse UHD images to enable consistent evaluation. If the staged approach holds, it would mean higher-fidelity outputs for tasks involving high-resolution content with intricate details.

Core claim

By decomposing UHD image restoration into zero-frequency enhancement, low-frequency restoration, and high-frequency refinement, the ERR framework integrates ZFE for holistic mappings, LFR for coarse-scale reconstruction, and HFR with FW-KAN for detail recovery, delivering superior performance across UHD restoration tasks on the new LSUHDIR benchmark as confirmed by experiments and module ablations.

What carries the argument

The ERR framework that couples a zero-frequency enhancer (ZFE) using global priors, a low-frequency restorer (LFR) focused on coarse information, and a high-frequency refiner (HFR) employing frequency-windowed Kolmogorov-Arnold Network (FW-KAN) to handle fine textures through progressive spectral decoupling.

If this is right

  • The ZFE, LFR, and HFR modules each address distinct frequency bands and their removal reduces overall restoration quality according to the ablations.
  • The LSUHDIR dataset provides a standardized large-scale testbed that future methods can use for fair comparison on diverse UHD scenes.
  • The frequency-windowed Kolmogorov-Arnold Network component enables targeted recovery of intricate high-frequency details that standard convolutions struggle with.
  • The overall pipeline supports multiple restoration tasks including denoising, deblurring, and enhancement while maintaining efficiency through specialization.

Where Pith is reading between the lines

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

  • The progressive frequency separation could extend naturally to video sequences by applying the same staged processing across frames to maintain temporal consistency.
  • Replacing traditional layers with frequency-windowed Kolmogorov-Arnold Networks in other high-resolution vision models might reduce parameter counts while preserving detail fidelity.
  • The benchmark construction process highlights the need for content-diverse UHD data in other domains such as satellite or medical imaging.

Load-bearing premise

That breaking the restoration into separate zero-, low-, and high-frequency stages with dedicated sub-networks will reliably outperform existing single-network end-to-end methods on varied UHD images.

What would settle it

Quantitative comparison of the ERR method against leading end-to-end UHD restoration models on the LSUHDIR dataset using standard metrics such as PSNR, SSIM, and LPIPS to measure whether the staged approach yields measurable gains.

Figures

Figures reproduced from arXiv: 2604.15654 by Chen Zhao, Enxuan Gu, Jian Yang, Kai Zhang, Xiaoming Liu, Ying Tai, Yunzhe Xu, Zhizhou Chen.

Figure 1
Figure 1. Figure 1: Our core motivation. Based on the observations in (a) and (b), we decompose the complex UHD restoration problem [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the proposed methods and the currently existing multi-stage coarse-to-fine paradigm. (a) multi [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Framework of our ERR. From a progressive spectral perspective, ERR consists of three collaborative sub-networks: [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of the PSNR curve with spectrum exchange [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Analysis of Zero-Frequency Coefficients across Various Tasks and Methods. In the figure, a larger distance from [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparison of zero-frequency components. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Zero-frequency visualization. We swap the zero [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Architecture of the global perception transformer [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Architecture of the residual state space block (RSSB). [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: Left: We showcase representative low-quality (failure) cases with their corresponding scores in two key dimensions—detail richness and content complex￾ity—underscoring the necessity of such filtering. Right: In contrast, our LSUHDIR samples demonstrate superior texture fidelity and higher semantic complexity. Regularization. In the final stage, our objective is to inject high-frequency information, with a… view at source ↗
Figure 12
Figure 12. Figure 12: Visual comparison with linear system. The visual [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of the GLCM score distributions [PITH_FULL_IMAGE:figures/full_fig_p009_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of the shannon entropy score distribu [PITH_FULL_IMAGE:figures/full_fig_p009_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: A visualization comparison between the resize and [PITH_FULL_IMAGE:figures/full_fig_p010_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Visual comparison with other SOTA methods on the UHD-LL [ [PITH_FULL_IMAGE:figures/full_fig_p011_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Visual comparison with other SOTA methods on the 4K-Rain13k [ [PITH_FULL_IMAGE:figures/full_fig_p011_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Visual comparison with other SOTA methods on the UHD-Haze [ [PITH_FULL_IMAGE:figures/full_fig_p011_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Visual comparison on the UHDM [2]. TABLE 15: Quantitative comparison for underwater image enhancement. Methods UIEB [103] LSUI [104] PSNR SSIM PSNR SSIM Params U-shape [104] 16.01 0.8180 23.26 0.8241 65.6M DM-water [123] 21.79 0.8517 24.15 0.8716 - CECF [124] 21.35 0.7748 26.12 0.8664 - HCLR [125] 22.24 0.9002 26.64 0.8815 4.87M MambaIR [84] 21.00 0.8618 24.90 0.8771 16.7M IERR 24.25 0.9318 27.54 0.8871 5… view at source ↗
Figure 21
Figure 21. Figure 21: Visual comparison with other SOTA methods on our UHD-Noise. [PITH_FULL_IMAGE:figures/full_fig_p013_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Visual comparison with other SOTA methods on our UHD-JPEG. [PITH_FULL_IMAGE:figures/full_fig_p013_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Effect of different frequency cutoffs k. TABLE 16: Ablation study with different architecture on our ERR. PR refers to progressive residual, as indicated by the yellow arrows between each stage in [PITH_FULL_IMAGE:figures/full_fig_p014_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Visual ablation results for the zero-frequency part. [PITH_FULL_IMAGE:figures/full_fig_p014_24.png] view at source ↗
Figure 26
Figure 26. Figure 26: Analysis of the model complexity. TABLE 21: Ablation study with architecture on our IERR. Method GPP Down Lg LEM ZR PSNR↑ SSIM↑ I ✓ × × × × 27.64 0.931 II ✓ ✓ × × × 27.63 0.932 III ✓ ✓ ✓ × × 27.71 0.931 IV ✓ ✓ ✓ ✓ × 27.81 0.932 IERR ✓ ✓ ✓ ✓ ✓ 27.87 0.932 degraded performance—highlighting the essential role of these components. We then ablate the global and local AvP branches within the AAP unit individual… view at source ↗
read the original abstract

Ultra-high-definition (UHD) image restoration poses unique challenges due to the high spatial resolution, diverse content, and fine-grained structures present in UHD images. To address these issues, we introduce a progressive spectral decomposition for the restoration process, decomposing it into three stages: zero-frequency \textbf{enhancement}, low-frequency \textbf{restoration}, and high-frequency \textbf{refinement}. Based on this formulation, we propose a novel framework, \textbf{ERR}, which integrates three cooperative sub-networks: the zero-frequency enhancer (ZFE), the low-frequency restorer (LFR), and the high-frequency refiner (HFR). The ZFE incorporates global priors to learn holistic mappings, the LFR reconstructs the main content by focusing on coarse-scale information, and the HFR adopts our proposed frequency-windowed Kolmogorov-Arnold Network (FW-KAN) to recover fine textures and intricate details for high-fidelity restoration. To further advance research in UHD image restoration, we also construct a large-scale, high-quality benchmark dataset, \textbf{LSUHDIR}, comprising 82{,}126 UHD images with diverse scenes and rich content. Our proposed methods demonstrate superior performance across a range of UHD image restoration tasks, and extensive ablation studies confirm the contribution and necessity of each module. Project page: https://github.com/NJU-PCALab/ERR.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a progressive spectral decoupling paradigm for UHD image restoration that decomposes the process into zero-frequency enhancement (ZFE for global priors), low-frequency restoration (LFR for coarse content), and high-frequency refinement (HFR using a proposed frequency-windowed Kolmogorov-Arnold Network for fine details). These are integrated into the ERR framework with three cooperative sub-networks. The work also introduces the LSUHDIR benchmark dataset containing 82,126 diverse UHD images and claims superior performance across UHD restoration tasks, with ablations confirming the necessity of each module.

Significance. If the empirical results hold after controlling for capacity, the work would offer a structured frequency-aware approach to UHD restoration that could improve detail recovery over monolithic end-to-end networks, while the new large-scale dataset would provide a valuable community resource for high-resolution image restoration research.

major comments (2)
  1. [Abstract] Abstract: The central claim of superior performance due to progressive spectral decoupling is not isolated from the increased model capacity introduced by deploying three distinct sub-networks (ZFE, LFR, HFR). Without explicit comparisons to capacity-matched or scaled baselines (e.g., single-network models with equivalent parameters or FLOPs), gains cannot be confidently attributed to the frequency-specific stages rather than total compute.
  2. [Ablation studies] Ablation studies: The manuscript states that ablations confirm the contribution and necessity of each module, yet provides no indication that these controls hold total parameter count or FLOPs fixed across variants. This leaves open the possibility that observed improvements stem from added capacity rather than the cooperative spectral decomposition.
minor comments (1)
  1. [Abstract] The acronym ERR is introduced without expansion in the abstract, which reduces immediate clarity for readers unfamiliar with the framework.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below and outline planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of superior performance due to progressive spectral decoupling is not isolated from the increased model capacity introduced by deploying three distinct sub-networks (ZFE, LFR, HFR). Without explicit comparisons to capacity-matched or scaled baselines (e.g., single-network models with equivalent parameters or FLOPs), gains cannot be confidently attributed to the frequency-specific stages rather than total compute.

    Authors: We acknowledge that the current presentation does not isolate the contribution of progressive spectral decoupling from the total model capacity of the three-subnetwork architecture. While the design intentionally assigns specialized roles—ZFE for global priors, LFR for coarse content, and HFR with the frequency-windowed KAN for fine details—the referee is correct that direct attribution requires capacity-controlled comparisons. In the revised manuscript we will add experiments comparing ERR against single-network baselines scaled to match both parameter count and FLOPs, allowing readers to evaluate whether the frequency-specific stages provide benefits beyond increased compute. revision: yes

  2. Referee: [Ablation studies] Ablation studies: The manuscript states that ablations confirm the contribution and necessity of each module, yet provides no indication that these controls hold total parameter count or FLOPs fixed across variants. This leaves open the possibility that observed improvements stem from added capacity rather than the cooperative spectral decomposition.

    Authors: We agree that the ablation studies as currently reported do not hold total parameter count or FLOPs fixed, which limits the strength of claims about the necessity of the cooperative spectral decomposition. The existing ablations demonstrate the effect of removing or altering individual sub-networks, but do not adjust architecture dimensions to equalize capacity. We will revise the ablation section to include capacity-matched variants (e.g., by scaling channel widths or depths in ablated models) and report both parameter counts and FLOPs, thereby providing clearer evidence that performance differences arise from the frequency-aware design rather than capacity differences. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical architecture and benchmark validation

full rationale

The paper proposes an empirical framework (ERR) that decomposes UHD restoration into zero-, low-, and high-frequency stages implemented via cooperative sub-networks (ZFE, LFR, HFR with FW-KAN) plus a new dataset (LSUHDIR). Superiority claims rest on external benchmark comparisons and internal ablations, not on any equation or parameter that reduces to its own fitted inputs by construction. No self-citation load-bearing uniqueness theorems, no ansatz smuggled via prior work, and no renaming of known results as new derivations appear in the provided text. The derivation chain is a standard engineering pipeline whose outputs are falsifiable against independent test sets.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that frequency-band decomposition is a natural and effective way to factor UHD restoration; network weights and training hyperparameters are free parameters but are standard for deep learning methods.

free parameters (1)
  • network hyperparameters and training schedule
    Standard deep-learning weights and learning rates fitted to the new dataset and tasks.
axioms (1)
  • domain assumption UHD images can be meaningfully decomposed into zero-, low-, and high-frequency components that can be restored independently by specialized sub-networks.
    Invoked in the progressive spectral decomposition formulation in the abstract.

pith-pipeline@v0.9.0 · 5569 in / 1239 out tokens · 61644 ms · 2026-05-10T09:21:40.650927+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. L2P: Unlocking Latent Potential for Pixel Generation

    cs.CV 2026-05 unverdicted novelty 6.0

    L2P repurposes pre-trained LDMs for direct pixel generation via large-patch tokenization and shallow-layer training on synthetic data, matching source performance with 8-GPU training and enabling native 4K output.

Reference graph

Works this paper leans on

128 extracted references · 26 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    Embedding fourier for ultra-high-definition low-light image enhancement

    C. Li, C.-L. Guo, M. Zhou, Z. Liang, S. Zhou, R. Feng, and C. C. Loy, “Embedding Fourier for ultra-high-definition low-light image enhancement,”arXiv preprint arXiv:2302.11831, 2023

    work page arXiv 2023

  2. [2]

    Towards ef- ficient and scale-robust ultra-high-definition image demoiréing,

    X. Yu, P . Dai, W. Li, L. Ma, J. Shen, J. Li, and X. Qi, “Towards ef- ficient and scale-robust ultra-high-definition image demoiréing,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 646–662

    2022

  3. [3]

    Ultra-high-definition image dehazing via multi-guided bilateral learning,

    Z. Zheng, W. Ren, X. Cao, X. Hu, T. Wang, F. Song, and X. Jia, “Ultra-high-definition image dehazing via multi-guided bilateral learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 16 180–16 189

    2021

  4. [4]

    Multi- scale separable network for ultra-high-definition video deblur- ring,

    S. Deng, W. Ren, Y. Yan, T. Wang, F. Song, and X. Cao, “Multi- scale separable network for ultra-high-definition video deblur- ring,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 14 030–14 039

    2021

  5. [5]

    UHDNeRF: Ultra-high-definition neural radiance fields,

    Q. Li, F. Li, J. Guo, and Y. Guo, “UHDNeRF: Ultra-high-definition neural radiance fields,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 23 097–23 108

    2023

  6. [6]

    Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer,

    H. Sun, Y. Zhang, L. Xu, S. Jin, and Y. Chen, “Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2025, pp. 7087–7095

    2025

  7. [7]

    Correlation matching transformation Transformers for UHD image restoration,

    C. Wang, J. Pan, W. Wang, G. Fu, S. Liang, M. Wang, X.-M. Wu, and J. Liu, “Correlation matching transformation Transformers for UHD image restoration,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2024, pp. 5336–5344

    2024

  8. [8]

    Towards ultra-high-definition image deraining: A benchmark and an effi- cient method,

    H. Chen, X. Chen, C. Wu, Z. Zheng, J. Pan, and X. Fu, “Towards ultra-high-definition image deraining: A benchmark and an effi- cient method,”IEEE Trans. Multimedia (TMM), pp. 1–13, 2026

    2026

  9. [9]

    SwinIR: Image restoration using Swin Transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “SwinIR: Image restoration using Swin Transformer,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 1833–1844

    2021

  10. [10]

    Restormer: Efficient Transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.- H. Yang, “Restormer: Efficient Transformer for high-resolution image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 5728–5739

    2022

  11. [11]

    Learn- ing multi-scale spatial-frequency features for image denoising,

    X. Zhao, C. Zhao, X. Hu, H. Zhang, Y. Tai, and J. Yang, “Learn- ing multi-scale spatial-frequency features for image denoising,” arXiv preprint arXiv:2506.16307, 2025

    work page arXiv 2025

  12. [12]

    More realistic and accurate precipitation nowcasting with conditional rectified flow transformers,

    Y. Zhou, C. Zhao, F. Ji, R. Hang, Q. Liu, and X.-T. Yuan, “More realistic and accurate precipitation nowcasting with conditional rectified flow transformers,”Engineering Applications of Artificial Intelligence, vol. 165, p. 113402, 2026

    2026

  13. [13]

    Multi-cropping contrastive learning and domain consistency for unsupervised image-to-image translation,

    C. Zhao, W.-L. Cai, Z. Yuan, and C.-W. Hu, “Multi-cropping contrastive learning and domain consistency for unsupervised image-to-image translation,”IET Image Processing, vol. 19, no. 1, p. e70006, 2025

    2025

  14. [14]

    O-mamba: O- shape state-space model for underwater image enhancement,

    C. Dong, C. Zhao, W. Cai, B. Yang, and Y. Guo, “O-mamba: O- shape state-space model for underwater image enhancement,” inChinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 2025, pp. 168–182

    2025

  15. [15]

    Cycle contrastive adver- sarial learning with structural consistency for unsupervised high- quality image deraining transformer,

    C. Zhao, W. Cai, C. Hu, and Z. Yuan, “Cycle contrastive adver- sarial learning with structural consistency for unsupervised high- quality image deraining transformer,”Neural Networks, vol. 178, p. 106428, 2024

    2024

  16. [16]

    Spectral normalization and dual contrastive regularization for image-to-image translation,

    C. Zhao, W.-L. Cai, and Z. Yuan, “Spectral normalization and dual contrastive regularization for image-to-image translation,” The Visual Computer, vol. 41, no. 1, pp. 129–140, 2025

    2025

  17. [17]

    Addsr: Accelerating diffusion-based blind super- resolution with adversarial diffusion distillation.arXiv preprint arXiv:2404.01717, 2024

    R. Xie, C. Zhao, K. Zhang, Z. Zhang, J. Zhou, J. Yang, and Y. Tai, “Addsr: Accelerating diffusion-based blind super- resolution with adversarial diffusion distillation,”arXiv preprint arXiv:2404.01717, 2024

    work page arXiv 2024

  18. [18]

    Exploiting multimodal spatial-temporal patterns for video object tracking,

    X. Hu, Y. Tai, X. Zhao, C. Zhao, Z. Zhang, J. Li, B. Zhong, and J. Yang, “Exploiting multimodal spatial-temporal patterns for video object tracking,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 3581–3589

    2025

  19. [19]

    Ultrahr-100k: Enhancing uhr image synthesis with a large-scale high-quality dataset,

    C. Zhao, E. Ci, Y. Xu, T. Fan, S. Guan, Y. Ge, J. Yang, and Y. Tai, “Ultrahr-100k: Enhancing uhr image synthesis with a large-scale high-quality dataset,”Advances in Neural Information Processing Systems, 2025

    2025

  20. [20]

    Luve: Latent-cascaded ultra-high-resolution video generation with dual frequency experts.arXiv preprint arXiv:2602.11564, 2026

    C. Zhao, J. Chen, H. Li, Z. Kang, S. Lu, X. Wei, K. Zhang, J. Yang, and Y. Tai, “Luve: Latent-cascaded ultra-high-resolution video generation with dual frequency experts,”arXiv preprint arXiv:2602.11564, 2026

    work page arXiv 2026

  21. [21]

    TF-ICON: Diffusion- based training-free cross-domain image composition,

    S. Lu, Y. Liu, and A. W.-K. Kong, “TF-ICON: Diffusion- based training-free cross-domain image composition,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 2294–2305

    2023

  22. [22]

    Learning a physical- aware diffusion model based on Transformer for underwater image enhancement,

    C. Zhao, C. Dong, W. Cai, and Y. Wang, “Learning a physical- aware diffusion model based on Transformer for underwater image enhancement,”IEEE Trans. Geosci. Remote Sens. (TGRS), vol. 64, pp. 1–14, 2026

    2026

  23. [23]

    Adapt or perish: Adaptive sparse Transformer with attentive feature refinement for image restoration,

    S. Zhou, D. Chen, J. Pan, J. Shi, and J. Yang, “Adapt or perish: Adaptive sparse Transformer with attentive feature refinement for image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 2952–2963

    2024

  24. [24]

    Efficient and explicit modelling of image hierar- chies for image restoration,

    Y. Li, Y. Fan, X. Xiang, D. Demandolx, R. Ranjan, R. Timofte, and L. Van Gool, “Efficient and explicit modelling of image hierar- chies for image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 18 278–18 289

    2023

  25. [25]

    Distilling semantic priors from SAM to efficient image restoration models,

    Q. Zhang, X. Liu, W. Li, H. Chen, J. Liu, J. Hu, Z. Xiong, C. Yuan, and Y. Wang, “Distilling semantic priors from SAM to efficient image restoration models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 25 409–25 419

    2024

  26. [26]

    HomoFormer: Homogenized Transformer for image shadow removal,

    J. Xiao, X. Fu, Y. Zhu, D. Li, J. Huang, K. Zhu, and Z.-J. Zha, “HomoFormer: Homogenized Transformer for image shadow removal,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 25 617–25 626

    2024

  27. [27]

    Migc: Multi- instance generation controller for text-to-image synthesis,

    D. Zhou, Y. Li, F. Ma, X. Zhang, and Y. Yang, “Migc: Multi- instance generation controller for text-to-image synthesis,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 6818–6828

    2024

  28. [28]

    Mace: Mass IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 17 concept erasure in diffusion models,

    S. Lu, Z. Wang, L. Li, Y. Liu, and A. W.-K. Kong, “Mace: Mass IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 17 concept erasure in diffusion models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 6430–6440

    2024

  29. [29]

    Rethinking 3d convolution inℓ p-norm space,

    L. Zhang, Y. Zhong, J. Wang, Z. Min, L. Liuet al., “Rethinking 3d convolution inℓ p-norm space,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  30. [30]

    Gapt-dar: Category-level garments pose tracking via integrated 2d deformation and 3d reconstruction,

    L. Zhang, M. Xu, J. Wang, Q. Yu, L. Yang, Y. Li, C. Lu, R. Wang, and L. Liu, “Gapt-dar: Category-level garments pose tracking via integrated 2d deformation and 3d reconstruction,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22 638–22 647

    2025

  31. [31]

    Rˆ 2-art: Category-level articulation pose estimation from single rgb image via cascade render strategy,

    L. Zhang, H. Jiang, Y. Huo, Y. Zhong, J. Wang, X. Wang, R. Wang, and L. Liu, “Rˆ 2-art: Category-level articulation pose estimation from single rgb image via cascade render strategy,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 9, 2025, pp. 9985–9993

    2025

  32. [32]

    Vocapter: Voting- based pose tracking for category-level articulated object via inter- frame priors,

    L. Zhang, Z. Han, Y. Zhong, Q. Yu, X. Wuet al., “Vocapter: Voting- based pose tracking for category-level articulated object via inter- frame priors,” inACM Multimedia 2024, 2024

    2024

  33. [33]

    U-cope: Taking a further step to universal 9d category-level object pose estimation,

    L. Zhang, W. Meng, Y. Zhong, B. Kong, M. Xu, J. Du, X. Wang, R. Wang, and L. Liu, “U-cope: Taking a further step to universal 9d category-level object pose estimation,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 254–270

    2025

  34. [34]

    Bidedpo: Condi- tional image generation with simultaneous text and condition alignment.arXiv preprint arXiv:2511.19268, 2025

    D. Zhou, M. Li, Z. Yang, Y. Lu, Y. Xu, Z. Wang, Z. Huang, and Y. Yang, “Bidedpo: Conditional image generation with simultaneous text and condition alignment,”arXiv preprint arXiv:2511.19268, 2025

    work page arXiv 2025

  35. [35]

    Wave-Mamba: Wavelet state space model for ultra-high-definition low-light image en- hancement,

    W. Zou, H. Gao, W. Yang, and T. Liu, “Wave-Mamba: Wavelet state space model for ultra-high-definition low-light image en- hancement,” inProc. ACM Int. Conf. Multimedia (ACM MM), 2024, pp. 1534–1543

    2024

  36. [36]

    Ultra- high-definition low-light image enhancement: A benchmark and Transformer-based method,

    T. Wang, K. Zhang, T. Shen, W. Luo, B. Stenger, and T. Lu, “Ultra- high-definition low-light image enhancement: A benchmark and Transformer-based method,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2023, pp. 2654–2662

    2023

  37. [37]

    arXiv preprint arXiv:2305.10028 , year=

    D. Zhou, Z. Yang, and Y. Yang, “Pyramid diffusion models for low-light image enhancement,”arXiv preprint arXiv:2305.10028, 2023

    work page arXiv 2023

  38. [38]

    Migc++: Advanced multi-instance generation controller for image synthesis,

    D. Zhou, Y. Li, F. Ma, Z. Yang, and Y. Yang, “Migc++: Advanced multi-instance generation controller for image synthesis,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  39. [39]

    3dis: Depth-driven decoupled instance synthesis for text-to-image generation.arXiv preprint arXiv:2410.12669, 2024

    D. Zhou, J. Xie, Z. Yang, and Y. Yang, “3dis: Depth-driven decoupled instance synthesis for text-to-image generation,”arXiv preprint arXiv:2410.12669, 2024

    work page arXiv 2024

  40. [40]

    Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image mod- els,

    D. Zhou, M. Li, Z. Yang, and Y. Yang, “Dreamrenderer: Taming multi-instance attribute control in large-scale text-to-image mod- els,”arXiv preprint arXiv:2503.12885, 2025

    work page arXiv 2025

  41. [41]

    3dis-flux: simple and efficient multi-instance generation with dit ren- dering.arXiv preprint arXiv:2501.05131, 2025

    D. Zhou, J. Xie, Z. Yang, and Y. Yang, “3dis-flux: simple and efficient multi-instance generation with dit rendering,”arXiv preprint arXiv:2501.05131, 2025

    work page arXiv 2025

  42. [42]

    Con- textgen: Contextual layout anchoring for identity-consistent multi-instance generation.arXiv preprint arXiv:2510.11000,

    R. Xu, D. Zhou, F. Ma, and Y. Yang, “Contextgen: Contextual lay- out anchoring for identity-consistent multi-instance generation,” arXiv preprint arXiv:2510.11000, 2025

    work page arXiv 2025

  43. [43]

    PhyEdit: Towards Real-World Object Manipulation via Physically-Grounded Image Editing

    R. Xu, D. Zhou, X. Shen, F. Ma, and Y. Yang, “Phyedit: Towards real-world object manipulation via physically-grounded image editing,”arXiv preprint arXiv:2604.07230, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Learning non-uniform-sampling for ultra-high-definition image enhancement,

    W. Yu, Q. Zhu, N. Zheng, J. Huang, M. Zhou, and F. Zhao, “Learning non-uniform-sampling for ultra-high-definition image enhancement,” inProc. ACM Int. Conf. Multimedia (ACM MM), 2023, pp. 1412–1421

    2023

  45. [45]

    P-BiC: Ultra-high-definition image moiré patterns removal via patch bilateral compensation,

    Z. Xiao, Z. Lu, and X. Wang, “P-BiC: Ultra-high-definition image moiré patterns removal via patch bilateral compensation,” in Proc. ACM Int. Conf. Multimedia (ACM MM), 2024, pp. 8365–8373

    2024

  46. [46]

    DreamUHD: Frequency enhanced variational autoencoder for ultra-high- definition image restoration,

    Y. Liu, D. Li, J. Xiao, Y. Bao, S. Xu, and X. Fu, “DreamUHD: Frequency enhanced variational autoencoder for ultra-high- definition image restoration,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2025, pp. 5712–5720

    2025

  47. [47]

    Neural discrimination-prompted Transformers for efficient UHD image restoration and enhancement,

    C. Wang, J. Pan, L. Wang, W. Wang, and Y. Yang, “Neural discrimination-prompted Transformers for efficient UHD image restoration and enhancement,”Int. J. Comput. Vis. (IJCV), vol. 134, no. 3, p. 84, 2026

    2026

  48. [48]

    Deep learning-driven ultra-high-definition image restoration: A survey,

    L. Wang, W. Zhou, C. Wang, K.-M. Lam, Z. Su, and J. Pan, “Deep learning-driven ultra-high-definition image restoration: A survey,”arXiv preprint arXiv:2505.16161, 2025

    work page arXiv 2025

  49. [49]

    Empowering resampling operation for ultra-high- definition image enhancement with model-aware guidance,

    W. Yu, J. Huang, B. Li, K. Zheng, Q. Zhu, M. Zhou, and F. Zhao, “Empowering resampling operation for ultra-high- definition image enhancement with model-aware guidance,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 25 722–25 731

    2024

  50. [50]

    Signals and the frequency domain,

    C.-Z. Lee, “Signals and the frequency domain,” https://web.stanford.edu/class/archive/engr/engr40m.1178/ slides/signals.pdf, 2017

    2017

  51. [51]

    Multi-scale progressive fusion network for single image deraining,

    K. Jiang, Z. Wang, P . Yi, C. Chen, B. Huang, Y. Luo, J. Ma, and J. Jiang, “Multi-scale progressive fusion network for single image deraining,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 8346–8355

    2020

  52. [52]

    Multi-stage progressive image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Multi-stage progressive image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 14 821–14 831

    2021

  53. [53]

    Benchmarking ultra-high-definition image super- resolution,

    K. Zhang, D. Li, W. Luo, W. Ren, B. Stenger, W. Liu, H. Li, and M.-H. Yang, “Benchmarking ultra-high-definition image super- resolution,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 14 769–14 778

    2021

  54. [54]

    Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution,

    L. Wang, C. Wang, J. Pan, X. Liu, W. Zhou, X. Sun, W. Wang, and Z. Su, “Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution,”IEEE Trans. Circuits Syst. Video Technol. (TCSVT), vol. 36, no. 2, pp. 2052–2068, 2026

    2052

  55. [55]

    From zero to detail: Deconstructing ultra-high-definition image restoration from progressive spectral perspective,

    C. Zhao, Z. Chen, Y. Xu, E. Gu, J. Li, Z. Yi, Q. Wang, J. Yang, and Y. Tai, “From zero to detail: Deconstructing ultra-high-definition image restoration from progressive spectral perspective,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 17 935–17 946

    2025

  56. [56]

    RefineAnything: Multimodal Region-Specific Refinement for Perfect Local Details

    D. Zhou, Y. Li, Z. Yang, and Y. Yang, “Refineanything: Multi- modal region-specific refinement for perfect local details,”arXiv preprint arXiv:2604.06870, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  57. [57]

    Image super-resolution using deep convolutional networks,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Image super-resolution using deep convolutional networks,”IEEE Trans. Pattern Anal. Mach. Intell. (TP AMI), vol. 38, no. 2, pp. 295–307, 2015

    2015

  58. [58]

    Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,

    K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,”IEEE Trans. Image Process. (TIP), vol. 26, no. 7, pp. 3142–3155, 2017

    2017

  59. [59]

    Robust watermarking using generative priors against image editing: From benchmarking to advances

    S. Lu, Z. Zhou, J. Lu, Y. Zhu, and A. W.-K. Kong, “Robust wa- termarking using generative priors against image editing: From benchmarking to advances,”arXiv preprint arXiv:2410.18775, 2024

    work page arXiv 2024

  60. [60]

    Eraseanything: Enabling concept erasure in rec- tified flow transformers,

    D. Gao, S. Lu, W. Zhou, J. Chu, J. Zhang, M. Jia, B. Zhang, Z. Fan, and W. Zhang, “Eraseanything: Enabling concept erasure in rec- tified flow transformers,” inForty-second International Conference on Machine Learning, 2025

    2025

  61. [61]

    Set you straight: Auto- steering denoising trajectories to sidestep unwanted concepts,

    L. Li, S. Lu, Y. Ren, and A. W.-K. Kong, “Set you straight: Auto- steering denoising trajectories to sidestep unwanted concepts,” in Proceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 9257–9266

    2025

  62. [62]

    arXiv preprint arXiv:2510.02253 (2025)

    Z. Zhou, S. Lu, S. Leng, S. Zhang, Z. Lian, X. Yu, and A. W.- K. Kong, “Dragflow: Unleashing dit priors with region based supervision for drag editing,”arXiv preprint arXiv:2510.02253, 2025

    work page arXiv 2025

  63. [63]

    Exposing and defending the achilles’ heel of video mixture-of-experts,

    S. Wang, Q. Liu, Y. Lyu, N. Li, Z. He, and C. Shan, “Exposing and defending the achilles’ heel of video mixture-of-experts,”arXiv preprint arXiv:2602.01369, 2026

    work page arXiv 2026

  64. [64]

    Efficient robustness assessment via adversarial spatial-temporal focus on videos,

    X. Wei, S. Wang, and H. Yan, “Efficient robustness assessment via adversarial spatial-temporal focus on videos,”IEEE Trans. Pattern Anal. Mach. Intell. (TP AMI), vol. 45, no. 9, pp. 10 898–10 912, 2023

    2023

  65. [65]

    Fast adversarial training with weak-to-strong spatial- temporal consistency in the frequency domain on videos,

    S. Wang, H. Liu, Y. Lyu, X. Hu, Z. He, W. Wang, C. Shan, and L. Wang, “Fast adversarial training with weak-to-strong spatial- temporal consistency in the frequency domain on videos,”IEEE Transactions on Information Forensics and Security, vol. 21, pp. 681– 696, 2025

    2025

  66. [66]

    Runawayevil: Jailbreaking the image-to-video gener- ative models,

    S. Wang, R. Qian, Y. Lyu, Q. Liu, L. Zou, J. Qin, S. Liu, and C. Shan, “Runawayevil: Jailbreaking the image-to-video gener- ative models,”arXiv preprint arXiv:2512.06674, 2025

    work page arXiv 2025

  67. [67]

    FCANet: Frequency channel attention networks,

    Z. Qin, P . Zhang, F. Wu, and X. Li, “FCANet: Frequency channel attention networks,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 783–792

    2021

  68. [68]

    Frequency separation for real-world super-resolution,

    M. Fritsche, S. Gu, and R. Timofte, “Frequency separation for real-world super-resolution,” inProc. IEEE/CVF Int. Conf. Comput. Vis. Worksh. (ICCVW), 2019, pp. 3599–3608

    2019

  69. [69]

    Intriguing findings of frequency selection for image deblurring,

    X. Mao, Y. Liu, F. Liu, Q. Li, W. Shen, and Y. Wang, “Intriguing findings of frequency selection for image deblurring,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2023, pp. 1905–1913. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 18

    2023

  70. [70]

    Diffusion model for camouflaged object detection,

    Z. Chen, R. Gao, T.-Z. Xiang, and F. Lin, “Diffusion model for camouflaged object detection,”arXiv preprint arXiv:2308.00303, 2023

    work page arXiv 2023

  71. [71]

    Ragd: Regional-aware diffusion model for text-to- image generation,

    Z. Chen, Y. Li, H. Wang, Z. Chen, Z. Jiang, J. Li, Q. Wang, J. Yang, and Y. Tai, “Ragd: Regional-aware diffusion model for text-to- image generation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 19 331–19 341

    2025

  72. [72]

    arXiv preprint arXiv:2503.23461 (2025)

    N. Du, Z. Chen, S. Gao, Z. Chen, X. Chen, Z. Jiang, J. Yang, and Y. Tai, “Textcrafter: Accurately rendering multiple texts in complex visual scenes,”arXiv preprint arXiv:2503.23461, 2025

    work page arXiv 2025

  73. [73]

    Dip: Taming diffusion models in pixel space

    Z. Chen, J. Zhu, X. Chen, J. Zhang, X. Hu, H. Zhao, C. Wang, J. Yang, and Y. Tai, “Dip: Taming diffusion models in pixel space,” arXiv preprint arXiv:2511.18822, 2025

    work page arXiv 2025

  74. [74]

    Adaptive guid- ance learning for camouflaged object detection,

    Z. Chen, X. Zhang, T.-Z. Xiang, and Y. Tai, “Adaptive guid- ance learning for camouflaged object detection,”arXiv preprint arXiv:2405.02824, 2024

    work page arXiv 2024

  75. [75]

    arXiv preprint arXiv:2509.21278 , year=

    S. Lu, Z. Lian, Z. Zhou, S. Zhang, C. Zhao, and A. W.-K. Kong, “Does FLUX already know how to perform physically plausible image composition?”arXiv preprint arXiv:2509.21278, 2025

    work page arXiv 2025

  76. [76]

    A general spatial-frequency learning framework for multimodal image fusion,

    M. Zhou, J. Huang, K. Yan, D. Hong, X. Jia, J. Chanussot, and C. Li, “A general spatial-frequency learning framework for multimodal image fusion,”IEEE Trans. Pattern Anal. Mach. Intell. (TP AMI), vol. 47, no. 7, pp. 5281–5298, 2025

    2025

  77. [77]

    Toward sufficient spatial-frequency interaction for gradient-aware underwater im- age enhancement,

    C. Zhao, W. Cai, C. Dong, and Z. Zeng, “Toward sufficient spatial-frequency interaction for gradient-aware underwater im- age enhancement,” inProc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), 2024, pp. 3220–3224

    2024

  78. [78]

    Image restoration via fre- quency selection,

    Y. Cui, W. Ren, X. Cao, and A. Knoll, “Image restoration via fre- quency selection,”IEEE Trans. Pattern Anal. Mach. Intell. (TP AMI), vol. 46, no. 2, pp. 1093–1108, 2023

    2023

  79. [79]

    Resolution-robust large mask inpainting with Fourier convolutions,

    R. Suvorov, E. Logacheva, A. Mashikhin, A. Remizova, A. Ashukha, A. Silvestrov, N. Kong, H. Goka, K. Park, and V . Lempitsky, “Resolution-robust large mask inpainting with Fourier convolutions,” inProc. IEEE/CVF Winter Conf. Appl. Com- put. Vis. (WACV), 2022, pp. 2149–2159

    2022

  80. [80]

    Fourmer: An efficient global modeling paradigm for image restoration,

    M. Zhou, J. Huang, C.-L. Guo, and C. Li, “Fourmer: An efficient global modeling paradigm for image restoration,” inProc. Int. Conf. Mach. Learn. (ICML), 2023, pp. 42 589–42 601

    2023

Showing first 80 references.