Pith. sign in

REVIEW 4 major objections 5 minor 72 references

Decoupling Cross-Modality Manifold Discrepancy: Leveraging Visible Diffusion Priors for Infrared Super-Resolution

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A frozen visible-spectrum diffusion model can be steered into the infrared manifold by two lightweight modules that correct global distribution and local structure, yielding infrared super-resolution with better perceptual and structural co

desk verdict A useful, well-engineered dual-module extension of ResShift for infrared SR with strong headline numbers, but the 'manifold steering' story is asserted rather than demonstrated and two load-bearing schedules are missing. read the letter →

arxiv 2607.21174 v1 pith:YUFDCJV6 submitted 2026-07-23 cs.CV

Yunpeng Hua , Hongwei Yu , Jiawei Li , Qiankun Liu , Huimin Ma , Jiansheng Chen This is my paper
classification cs.CV
keywords infraredimagesuper-resolutiondiffusionmodelscross-modaldiscrepancygenerativepriorsmanifoldalignmentglobalrepresentationmodulationlocalstructurerefinementResShift
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

The paper claims that the core obstacle to infrared image super-resolution is not resolution recovery but a mismatch between the visible-spectrum priors baked into pre-trained diffusion models and the infrared data distribution. Rather than fine-tuning or retraining the diffusion backbone, Shift-IISR keeps the backbone frozen and attaches two lightweight modules: GRM, which extracts infrared-specific global features and injects them into the time embedding at each sampling step, and LSR, which adds mean-normalized Sobel edge signals to the latent state at each step. Jointly these are supposed to progressively steer the denoising trajectory from the visible manifold toward the infrared manifold, improving global distributional and structural consistency while retaining most of the generative capacity of the original model. The paper supports this with quantitative gains in SSIM/LPIPS and distributional metrics on three infrared benchmarks, plus downstream detection and segmentation improvements.

What carries the argument

The central mechanism is a dual-path modulation of a frozen ResShift diffusion backbone. GRM (Global Representation Modulation) uses a weight-sharing convolutional extractor and a co-trained modality classifier to isolate infrared-specific global features from the LR infrared latent; these are projected and added to the time embedding via a gated, scheduling function. LSR (Local Structure Refinement) applies a Sobel operator to the LR latent, mean-normalizes the resulting edge map, and injects it directly into the intermediate latent at each reverse step via a time-decaying schedule. Together they correct global statistical drift and local geometric drift without fine-tuning the backbone.

What would settle it

Set both gamma_shift(t) and gamma_str(t) to zero across all timesteps; if the outputs still match the full Shift-IISR results, the modules are not doing the claimed work. Alternatively, feed the same LR infrared image through the frozen backbone with only GRM active and measure the distance (e.g., FID or Bhattacharyya distance in latent space) between predicted latents and true infrared latents: if the projection pushes latents off the infrared manifold, artifacts should appear.

Watch

Extended reading notes

Core claim

Shift-IISR's central claim is that cross-modality discrepancy in IISR can be decoupled into a global distribution shift and a local structural mismatch, and that both can be corrected without altering the frozen pre-trained diffusion backbone. GRM learns a relative modality shift by contrasting infrared and same-scene visible latent features, then injects the infrared-specific prior into the UNet's time embedding through a scheduling function; LSR injects normalized edge maps into the intermediate latents. The authors argue that the joint action of these two modulations progressively steers the diffusion trajectory from the visible manifold to the infrared manifold, suppressing visually plau

Load-bearing premise

The load-bearing premise is that a classifier-trained projection of the infrared LR latent, added to the time embedding, can reorient a frozen visible-prior diffusion model's predictions into the infrared manifold without fine-tuning—and that adding mean-normalized Sobel edges to the latent at every reverse step does not push the latent outside the distribution the frozen network expects; the two control schedules are never specified, so both bets remain untested in isolation

Editorial extensions

If this is right

  • Infrared image super-resolution no longer requires fine-tuning a diffusion model on scarce infrared data; a frozen visible prior can be adapted with small auxiliary modules.
  • The explicit decoupling of global distribution and local structure provides a recipe for adapting generative priors to other modality gaps where paired same-scene data exist.
  • Reported gains in SSIM, LPIPS, and histogram-based distributional metrics indicate that the approach improves perceptual and structural fidelity, not just pixel-wise accuracy.
  • Downstream object detection and semantic segmentation results improve, suggesting the reconstructed infrared images carry more usable semantic information.

Reading between the lines

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

  • Beyond the paper, the same decoupling strategy might transfer to other cross-modal restoration tasks, such as medical or SAR imaging, where a pre-trained visible-domain diffusion prior is available and the target modality has limited data.
  • The two control schedules, gamma_shift(t) and gamma_str(t), are never specified in the paper; systematically ablating them would test the steering claim and is a natural next step for reproducibility.
  • Because GRM relies on paired same-scene visible latents as references, the method inherits the need for aligned multi-modal training data; an unsupervised variant that learns the modulo shift without visible references would be a meaningful extension.
  • If the frozen-backbone assumption holds, the framework suggests a more general principle: generative priors can be redirected across modalities by additive corrections in the latent space, rather than by retraining or fine-tuning.
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

4 major / 5 minor

Summary. The paper proposes Shift-IISR, a diffusion-based infrared image super-resolution framework built on a frozen ResShift backbone. It introduces two auxiliary modules: Global Representation Modulation (GRM), which learns an infrared-specific global representation via a contrastive classifier on paired infrared/visible latents and injects it into the time embedding, and Local Structure Refinement (LSR), which injects mean-normalized Sobel edge maps into the intermediate latents at each reverse step. The method is evaluated on M3FD, RoadScene, and TNO, with quantitative comparisons, an ablation study, and downstream detection/segmentation experiments. The central claim is that GRM and LSR jointly and progressively steer the diffusion trajectory from the visible manifold to the infrared manifold, improving both distributional and structural consistency.

Significance. If the central claim holds, Shift-IISR is a practical and modular way to adapt a visible-domain diffusion prior to infrared super-resolution without fine-tuning the generative backbone, with potential benefits for downstream infrared perception tasks. The manuscript has concrete strengths: it provides a code link, uses a frozen backbone (which is a scientifically clean setup), includes downstream detection/segmentation evaluation, and reports an ablation of the two proposed modules. However, the paper's mechanistic claim about progressive manifold steering is not directly evidenced by the reported experiments, and the reproducibility of the method is hampered by unspecified temporal schedules and an apparent inconsistency in the training objective for the LSR forward process.

major comments (4)
  1. [§6, §4.1, Eq. (4)] The central claim that GRM and LSR 'progressively steer the diffusion trajectory from the visible manifold toward the infrared manifold' is asserted but never directly tested. All reported evidence (Table 1, Fig. 5, Figs. 6–9) concerns final reconstructed images. No analysis of intermediate latents z_t along the reverse chain, no measurement of proximity to visible versus infrared manifolds, and no comparison to a version in which GRM/LSR act only at the final step. The observed gains could therefore arise from generic conditioning or an edge-preserving post-processor rather than progressive reorientation. Please add trajectory-level analysis (e.g., per-step latent distances to reference manifolds, or a 'guidance only at final step' ablation) to support the mechanism.
  2. [§4.2 Eq. (8), §4.3 Eqs. (11), (14)] The temporal control schedules γ_shift(t) and γ_str(t) are never specified. The text only says γ_shift decreases monotonically and γ_str decays to zero; the exact functional form, ranges, and hyperparameters are omitted. These schedules control the magnitude of both interventions at every reverse step and are essential for reproducibility and for assessing whether the latent stays on the expected manifold. Provide the explicit schedules (linear/cosine, min/max, etc.) in the main text or supplementary, together with a sensitivity analysis.
  3. [§3 Eq. (3) vs §4.3 Eqs. (12)–(13)] There is an inconsistency between the training loss and the LSR-modified forward process. Equation (3) defines L_simple using the original residual-shift latent z_t from Eq. (2). Section 4.3 then states that LSR is 'consistently integrated into both the forward and reverse processes' and redefines the forward latent as Eq. (13). If training uses the modified forward process, Eq. (3) must be rewritten accordingly; if it does not, the claim of consistent integration is not reflected in the optimization. Please clarify which forward process is used to compute L_simple when LSR is active, and report the corrected training loss if needed.
  4. [§5.2, Fig. 5] Distributional consistency is a central contribution, but the only supporting evidence is a single-image histogram comparison with three scalar distances (Cosine Similarity, Bhattacharyya Distance, L1 Distance). No aggregate statistics over the test sets, no confidence intervals, and no per-method comparison table for these distributional metrics are provided. The claim of improved global distribution consistency needs quantitative support across the full evaluation datasets. Add a table reporting mean ± std for these metrics over Set5/Set15/Set20 (and other datasets if used) for all compared methods.
minor comments (5)
  1. [§5.1] The text says the model is evaluated on M3FD, RoadScene, and TNO, but Table 1 and Figures 4–5 only report M3FD subsets. Please state where the RoadScene/TNO results appear, or add a summary table in the main text.
  2. [Fig. 3] Caption typo: 'The architectures' is written as 'he architectures'.
  3. [§5.3, Table 2] The ablation table reports only PSNR/LPIPS/SSIM, not the distributional metrics that are central to the paper. Also, no standard deviations or significance tests are reported. Please include distributional metrics in the ablation and report variability over multiple runs or seeds.
  4. [§4.2, Eqs. (5)–(7)] It should be stated explicitly whether the same-scene visible latent z_y^{vis} is used only during training (as the dashed training-only paths in Fig. 2 suggest) or also at test time. This affects reproducibility and the practical applicability of the method.
  5. [§4.2, §5.1] Important hyperparameters are missing: the value of λ in Eq. (9), the initialization of g beyond '-2.0' (is it used as a scalar bias?), and the exact schedule ranges. Please report these in the experimental settings.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GRM/LSR are supervised/deterministic conditioning modules evaluated on external benchmarks, not constructions equivalent to their inputs.

full rationale

The derivation chain is self-contained in the sense required by the circularity check. GRM's δshift (Eq. 7) is learned from paired infrared/visible latents through the classifier objective L_cls (Eq. 6) and injected into the time embedding (Eq. 8); LSR's δstr is a deterministic mean-normalized Sobel map of the LR latent (Eqs. 10-14). Neither term is defined in terms of the reported PSNR/SSIM/LPIPS, histogram distances, or downstream mAP/mIoU, and none of those benchmark metrics appears in L_total (Eq. 9). The training objective includes the ResShift denoising loss and the modality-classification loss; the evaluation metrics are external and independent. The authors' self-citations ([19], [20], [22], [23]) support only general application statements (fusion, detection, embodied intelligence) and are not load-bearing for the GRM/LSR mechanism. No uniqueness theorem or ansatz is imported from the authors' prior work. Two non-circular limitations are worth flagging: the schedules γ_shift(t) and γ_str(t) are only described qualitatively ('decreases monotonically', 'decays to zero') and never specified, and the central 'trajectory steering' claim (Section 6) is supported only by final-image metrics and a single histogram rather than by analyzing intermediate latents z_t. These are reproducibility/evidence gaps, not reductions of the claimed predictions to the method's inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two learned feature-injection mechanisms, not on new physical entities. The main free parameters are the gating scalar, the loss weight, and two unspecified time schedules. The strongest domain assumptions are the transferability of a frozen visible-pretrained latent diffusion model to infrared and the sufficiency of latent-space edge injection as a structural prior.

free parameters (4)
  • learnable gating g = initialized to -2.0, trained
    Controls the sigmoid-gated injection strength of δshift in Eq. (7). No sensitivity analysis is given in the main text.
  • λ classification loss weight = not stated in main text
    Balances L_simple and L_cls in Eq. (9). The value affects training and is never reported.
  • γ_shift(t) schedule = monotone decreasing, functional form unspecified
    Modulates the GRM injection over timesteps in Eq. (8). The paper calls it 'predefined' but gives no formula.
  • γ_str(t) schedule = strongest at early sampling, decays to zero, form unspecified
    Modulates the LSR injection in Eq. (11). No closed form is supplied.
assumptions (5)
  • standard math ResShift residual-shifting formulation (Eqs. 1-3) correctly models the latent degradation process and its pretrained VQGAN/UNet backbone transfers to infrared LR latents.
    The paper builds directly on ResShift [56] without re-deriving it; all subsequent equations inherit its validity.
  • domain assumption Paired same-scene infrared and visible images are available and sufficiently aligned for contrastive modality learning in latent space.
    GRM uses visible latent z_vis_y as a training benchmark. M3FD/MSRS pairs must be co-registered, which is not verified in the paper.
  • domain assumption The visible-bias error can be corrected additively via time-embedding modulation δshift with no backbone fine-tuning.
    Core design choice of the frozen-backbone method in Eq. (8); no proof or drift analysis is offered.
  • domain assumption Mean-normalized Sobel edges of the LR latent are a sufficient structural prior and adding them to z_t at each sampling step does not push latents out of distribution.
    LSR Eqs. (10)-(14) assume this; no evidence is given about out-of-distribution behavior.
  • ad hoc to paper γ_shift(t) and γ_str(t) can be chosen as monotone time schedules that stabilize guidance.
    The schedules are central to the method but their functional forms are never specified, making them ad hoc design choices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupling Cross-Modality Manifold Discrepancy: Leveraging Visible Diffusion Priors for Infrared Super-Resolution." pith.science (2026). https://pith.science/paper/YUFDCJV6

@misc{pith2026260721174,
  author       = {Pith},
  title        = {Pith review of: Decoupling Cross-Modality Manifold Discrepancy: Leveraging Visible Diffusion Priors for Infrared Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YUFDCJV6}},
  note         = {Machine review of arXiv:2607.21174}
}
read the original abstract

Infrared image super-resolution (IISR) mitigates the limitations imposed by low spatial resolution. Existing methods have recognized that IISR should preserve consistency in global distribution and structural information while enhancing image clarity. However, these methods are either insufficient or overly intrusive, a problem that becomes even more pronounced in diffusion-based models. To address these issues, we propose a dual-path diffusion-based framework for IISR, termed Shift-IISR. The proposed method is designed to improve the consistency of IISR results while preserving the generative capacity of diffusion models. Specifically, we develop a Global Representation Modulation (GRM) module to extract modality-specific information from infrared imagery and guide the global distribution of the diffusion model toward the ground truth. In addition, we introduce a Local Structure Refinement (LSR) module to encourage the model to focus on structural information at each step of the iterative denoising process. Extensive experiments demonstrate that the proposed method effectively improves distributional and structural consistency while maintaining competitive super-resolution performance. The source code of the proposed Shift-IISR can be available at https://github.com/Assassink8/Shift-IISR.

Figures

Figures reproduced from arXiv: 2607.21174 by the authors.

Figure 1
Figure 1. The left illustrates the limitations of a pre-trained diffusion model for infrared image super-resolution. Shift-IISR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 2
Figure 2. Overall architecture of Shift-IISR: cross-modal discrepancies between visible and infrared domains are addressed by [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: he architectures of the feature extractor [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 4
Figure 4. Figure 4: Qualitative comparison of different super-resolution methods on the M [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of Grayscale Histogram Distributions Between Reconstructed Images and Ground Truth. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Quantitative results of different SR models on down [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of generated results under different ablation settings. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of Detection Performance for Infrared Image Super-Resolution Methods [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Comparison of Segmentation Performance for Infrared Image Super-Resolution Methods [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 3 linked inside Pith

  1. [1]

    Anil Bhattacharyya. 1943. On a measure of divergence between two statistical populations defined by their probability distribution.Bulletin of the Calcutta Mathematical Society35 (1943), 99–110

    1943

  2. [2]

    Bing Cao, Yiming Sun, Pengfei Zhu, and Qinghua Hu. 2023. Multi-Modal Gated Mixture of Local-to-Global Experts for Dynamic Image Fusion. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 23555–23564

    2023

  3. [3]

    Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. 2023. Video Super- Resolution Transformer. arXiv:2106.06847 [cs.CV] https://arxiv.org/abs/2106. 06847

    arXiv 2023

  4. [4]

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. 2021. Pre-Trained Image Processing Transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 12299–12310

    2021

  5. [5]

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. 2018. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. InProceedings of the European Conference on Computer Vision (ECCV)

    2018

  6. [6]

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. 2023. Activating More Pixels in Image Super-Resolution Transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 22367–22377

    2023

  7. [7]

    Zheng Chen, Haotong Qin, Yong Guo, Xiongfei Su, Xin Yuan, Linghe Kong, and Yulun Zhang. 2024. Binarized diffusion model for image super-resolution. Advances in Neural Information Processing Systems37 (2024), 30651–30669

    2024

  8. [8]

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. 2014. Learning a deep convolutional network for image super-resolution. InEuropean conference on computer vision. Springer, Springer International Publishing, Cham, 184–199

    2014

Show all 72 references
  1. [9]

    Patrick Esser, Robin Rombach, and Bjorn Ommer. 2021. Taming Transformers for High-Resolution Image Synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 12873–12883

    2021

  2. [10]

    Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. 2025. Mambairv2: Attentive state space restoration. InProceedings of the Computer Vision and Pattern Recognition Conference. 28124–28133

    2025

  3. [11]

    Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia

  4. [12]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  5. [13]

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-to- image translation with conditional adversarial networks. InProceedings of the IEEE conference on computer vision and pattern recognition. 1125–1134

    2017

  6. [14]

    Zhiying Jiang, Zengxi Zhang, Jinyuan Liu, Xin Fan, and Risheng Liu. 2024. Mul- tispectral image stitching via global-aware quadrature pyramid regression.IEEE Transactions on Image Processing33 (2024), 4288–4302

    2024

  7. [15]

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. 2023. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10124–10134

    2023

  8. [16]

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. 2022. Denoising diffusion restoration models.Advances in neural information processing systems 35 (2022), 23593–23606

    2022

  9. [17]

    Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunning- ham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. 2017. Photo-realistic single image super-resolution using a generative adversarial network. InProceedings of the...

    2017

  10. [18]

    Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. 2022. Srdiff: Single image super-resolution with diffusion probabilistic models.Neurocomputing479 (2022), 47–59

    2022

  11. [19]

    Jiawei Li, Jiansheng Chen, Jinyuan Liu, and Huimin Ma. 2023. Learning a graph neural network with cross modality interaction for image fusion. InProceedings of the 31st ACM international conference on multimedia. 4471–4479

    2023

  12. [20]

    Jiawei Li, Jiansheng Chen, Jinyuan Liu, Hongwei Yu, Xinlong Ding, and Huimin Ma. 2026. Multi-Scale Spatial Channel Joint Representation for General Multi- Modality Image Fusion With Self-Supervision.IEEE Transactions on Multimedia (2026)

    2026

  13. [21]

    Jing Li, Hongtao Huo, Chang Li, Renhua Wang, Chenhong Sui, and Zhao Liu

  14. [22]

    Jiawei Li, Jinyuan Liu, Shihua Zhou, Qiang Zhang, and Nikola K Kasabov. 2023. Gesenet: A general semantic-guided network with couple mask ensemble for medical image fusion.IEEE Transactions on Neural Networks and Learning Systems 35, 11 (2023), 16248–16261

    2023

  15. [23]

    Jiawei Li, Hongwei Yu, Jiansheng Chen, Xinlong Ding, Jinlong Wang, Jinyuan Liu, Bochao Zou, and Huimin Ma. 2025. A 2rnet: Adversarial attack resilient network for robust infrared and visible image fusion. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39...

    2025

  16. [24]

    Xingyuan Li, Jinyuan Liu, Zhixin Chen, Yang Zou, Long Ma, Xin Fan, and Risheng Liu. 2024. Contourlet residual for prompt learning enhanced infrared image super-resolution. InEuropean Conference on Computer Vision. Springer, 270–288

    2024

  17. [25]

    Xingyuan Li, Zirui Wang, Yang Zou, Zhixin Chen, Jun Ma, Zhiying Jiang, Long Ma, and Jinyuan Liu. 2025. Difiisr: A diffusion model with gradient guidance for infrared image super-resolution. InProceedings of the Computer Vision and Pattern Recognition Conference. 7534–7544

    2025

  18. [26]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. Swinir: Image restoration using swin transformer. InProceedings of the IEEE/CVF international conference on computer vision. 1833–1844

    2021

  19. [27]

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee

  20. [28]

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. 2024. Diffbir: Toward blind image restora- tion with generative diffusion prior. InEuropean conference on computer vision. Springer, 430–448

    2024

  21. [29]

    Jinyuan Liu, Zihang Chen, Zhu Liu, Zhiying Jiang, Long Ma, Xin Fan, and Risheng Liu. 2025. Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark. InThe Thirty-ninth Annual Conference on Neural Information Pro- cessing Systems. https://openreview.net/forum?...

    2025

  22. [30]

    Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. 2022. Target-aware dual adversarial learning and a multi- scenario multi-modality benchmark to fuse infrared and visible for object detec- tion. InProceedings of the IEEE/CVF conference ...

    2022

  23. [31]

    Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. 2020. Latticenet: Towards lightweight image super-resolution with lattice block. In European conference on computer vision. Springer, 272–289

    2020

  24. [32]

    Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas B Schön

  25. [33]

    Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. 2020. Structure-preserving super resolution with gradient guidance. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7769–7778

    2020

  26. [34]

    Jiayi Ma, Yong Ma, and Chang Li. 2019. Infrared and visible image fusion methods and applications: A survey.Information fusion45 (2019), 153–178

    2019

  27. [35]

    Brian B Moser, Arundhati S Shanbhag, Federico Raue, Stanislav Frolov, Sebastian Palacio, and Andreas Dengel. 2024. Diffusion models, image super-resolution, and everything: A survey.IEEE Transactions on Neural Networks and Learning Systems36, 7 (2024), 11793–11813

    2024

  28. [36]

    Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh

  29. [37]

    Feiwei Qin, Zhengwei Shen, Ruiquan Ge, Kai Zhang, Fei Lin, Yeru Wang, Juan M Gorriz, Ahmed Elazab, and Changmiao Wang. 2025. InfraFFN: A Feature Fusion Network leveraging dual-path convolution and self-attention for infrared image super-resolution.Knowledge-Based Systems310 (2...

    2025

  30. [38]

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. 2016. You only look once: Unified, real-time object detection. InProceedings of the IEEE conference on computer vision and pattern recognition. 779–788

    2016

  31. [39]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  32. [40]

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. 2022. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 conference proceedings. 1–10

    2022

  33. [41]

    Understanding the latent space of diffusion models through the lens of riemannian geometry.Advances in Neural Information Processing Systems36 (2023), 24129–24142

    2023

  34. [42]

    Gerard Salton, Anita Wong, and Chung-Shu Yang. 1975. A vector space model for automatic indexing.Commun. ACM18, 11 (1975), 613–620

    1975

  35. [43]

    Shuyao Shang, Zhengyang Shan, Guangxing Liu, LunQian Wang, XingHua Wang, Zekai Zhang, and Jinglin Zhang. 2024. Resdiff: Combining cnn and diffusion model for image super-resolution. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 8975–8983

    2024

  36. [44]

    Zhenning Shi, Haoshuai Zheng, Chen Xu, Changsheng Dong, Bin Pan, Xueshuo Xie, Along He, Tao Li, and Huazhu Fu. 2024. Resfusion: Denoising diffusion probabilistic models for image restoration based on prior residual noise.Advances in Neural Information Processing Systems37 (202...

    2024

  37. [45]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502(2020)

    2020 arXiv

  38. [46]

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. 2022. Image super-resolution via iterative refinement.IEEE transactions on pattern analysis and machine intelligence45, 4 (2022), 4713–4726

    2022

  39. [47]

    Alexander Toet. 2017. The TNO multiband image data collection.Data in brief 15 (2017), 249

    2017

  40. [48]

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. 2024. Exploiting diffusion prior for real-world image super- resolution.International Journal of Computer Vision132, 12 (2024), 5929–5949

    2024

  41. [49]

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. 2021. Real-esrgan: Train- ing real-world blind super-resolution with pure synthetic data. InProceedings of the IEEE/CVF international conference on computer vision. 1905–1914

    2021

  42. [50]

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. 2018. Esrgan: Enhanced super-resolution generative adversarial networks. InProceedings of the European conference on computer vision (ECCV) workshops. 0–0

    2018

  43. [51]

    Linfeng Tang, Chunyu Li, and Jiayi Ma. 2025. Mask-DiFuser: A Masked Diffusion Model for Unified Unsupervised Image Fusion.IEEE Transactions on Pattern MM’26, November 10–14, 2026, Rio de Janeiro, Brazil Yunpeng Hua et al. Analysis and Machine Intelligence(2025), 1–18

    2025

  44. [52]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 4 (2004), 600–612

    2004

  45. [53]

    Zhihao Wang, Jian Chen, and Steven CH Hoi. 2020. Deep learning for image super-resolution: A survey.IEEE transactions on pattern analysis and machine intelligence43, 10 (2020), 3365–3387

    2020

  46. [54]

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. 2023. Diffir: Efficient diffusion model for image restoration. InProceedings of the IEEE/CVF international conference on computer vision. 13095–13105

    2023

  47. [55]

    Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. 2020. U2Fusion: A unified unsupervised image fusion network.IEEE transactions on pattern analysis and machine intelligence44, 1 (2020), 502–518

    2020

  48. [56]

    Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. 2024. Sinsr: diffusion-based image super-resolution in a single step. InProceedings of the IEEE/CVF conference on computer vision and pattern recogniti...

    2024

  49. [57]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shah- baz Khan, and Ming-Hsuan Yang. 2022. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5728–5739

    2022

  50. [58]

    Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. 2021. Designing a practical degradation model for deep blind image super-resolution. InProceedings of the IEEE/CVF international conference on computer vision. 4791–4800

    2021

  51. [59]

    Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. 2024. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2856–2865

    2024

  52. [60]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional con- trol to text-to-image diffusion models. InProceedings of the IEEE/CVF international conference on computer vision. 3836–3847

    2023

  53. [61]

    Zongsheng Yue, Jianyi Wang, and Chen Change Loy. 2023. Resshift: Efficient diffusion model for image super-resolution by residual shifting.Advances in neural information processing systems36 (2023), 13294–13307

    2023

  54. [62]

    Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. 2022. Efficient long-range attention network for image super-resolution. InEuropean conference on computer vision. Springer, 649–667

    2022

  55. [63]

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. 2018. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV). 286–301

    2018

  56. [64]

    Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. 2018. Residual dense network for image super-resolution. InProceedings of the IEEE conference on computer vision and pattern recognition. 2472–2481

    2018

  57. [65]

    Yongbing Zhang, Yulun Zhang, Jian Zhang, and Qionghai Dai. 2015. CCR: Clus- tering and collaborative representation for fast single image super-resolution. IEEE Transactions on Multimedia18, 3 (2015), 405–417

    2015

  58. [66]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  59. [72]

    Haifeng Zhong, Fan Tang, Hyung Jin Chang, Xingyu Zhu, and Yixing Gao. 2025. DarkSeg: Infrared-Driven Semantic Segmentation for Garment Grasping De- tection in Low-Light Conditions. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 17680–17687

    2025

  60. [2017]

    In Proceedings of the IEEE conference on computer vision and pattern recognition workshops

    Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 136–144

  61. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  62. [2020]

    Multigrained attention network for infrared and visible image fusion.IEEE Transactions on Instrumentation and Measurement70 (2020), 1–12

    2020

  63. [2023]

    arXiv preprint arXiv:2301.11699(2023)

    Image restoration with mean-reverting stochastic differential equations. arXiv preprint arXiv:2301.11699(2023)

    2023 arXiv

  64. [2024]

    InEuropean conference on computer vision

    Mambair: A simple baseline for image restoration with state-space model. InEuropean conference on computer vision. Springer, 222–241

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.