DroneSR: Rethinking Few-shot Thermal Image Super-Resolution from Drone-based Perspective

Ce Liu; Liang Lin; Xiaopeng Liu; Xingyuan Guo; Yukai Shi; Zhipeng Weng

REVIEW 4 major objections 5 minor 31 references

DroneSR: Rethinking Few-shot Thermal Image Super-Resolution from Drone-based Perspective

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Gaussian posterization tames few-shot thermal super-resolution

desk verdict A useful thermal benchmark and a promising quantization augmentation, but the method description and baseline protocol are too broken to support the central claims as submitted. read the letter →

arxiv 2509.01898 v1 pith:JX56TXST submitted 2025-09-02 cs.CV

Zhipeng Weng , Xiaopeng Liu , Ce Liu , Xingyuan Guo , Yukai Shi , Liang Lin This is my paper · ORCID

classification cs.CV

keywords thermalimagesuper-resolutionfew-shotlearningdiffusionmodelsGaussianquantizationoverfittingmitigationdroneimageryinfraredimaginglatent

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 severe overfitting in few-shot thermal super-resolution can be tamed by a simple preprocessing step: randomly quantizing the pixel values of each training image into a small number of bins selected from a Gaussian. This 'Gaussian quantization' acts as a cheap stochastic degradation that smooths away sensor noise while preserving edges, and it lets a large diffusion-based super-resolution model keep converging when trained on only 200 drone-captured infrared images. To demonstrate the problem and the fix, the authors assemble a multi-source aerial and ground thermal dataset, DroneSR, and show that their two-stage coarse-plus-diffusion pipeline reaches 30.22 dB PSNR at x4 upscaling, 1.85 dB above the best diffusion baseline and 2.77 dB above the same architecture without quantization. If the claim holds, randomized posterization is a viable stabilizer for large generative restoration models under data scarcity.

What carries the argument

The load-bearing mechanism is Gaussian probability-guided quantization, a randomized posterization step: for each channel it draws an integer K from a Gaussian with mean 17 and standard deviation 2, splits the pixel range into K random intervals, and replaces every pixel with a proxy value drawn from its interval. This turns each raw infrared image into one of many possible quantized versions, expanding input diversity and suppressing high-frequency sensor noise before the image reaches the network. The rest of the system is a two-stage architecture: a Swin-Transformer coarse reconstruction network with multi-stage feature fusion (15.8M parameters), a latent diffusion-driven texture refineme

What would settle it

Train the same architecture on the same 200 DroneSR images with and without Gaussian quantization, tracking validation PSNR every 1000 steps; if the quantized curve does not stay above the unquantized curve from the overfitting point onward, the central claim collapses. A second check: replace the 865.9M diffusion refiner with the coarse network alone and see whether the 2.77 dB gain survives.

Watch

Extended reading notes

Core claim

The central discovery is that Gaussian probability-guided quantization, applied to input infrared images before training, converts a data-scarce super-resolution problem into a more learnable one without changing architecture capacity. The method samples an interval count K from a Gaussian, splits each channel's pixel range into K random regions, and replaces pixels by representative values; repeated random sampling diversifies the degradation seen by the network. With this input transform, the authors report that the latent diffusion texture refinement stage stops diverging and that overfitting is visibly reduced in validation curves. Quantitatively, quantization alone accounts for a 2.77 d

Load-bearing premise

The whole two-stage design depends on the unstated assumption that a latent diffusion model pretrained on ordinary RGB photos can be plugged directly into single-channel thermal infrared super-resolution, with no domain adaptation, and still add useful texture.

Editorial extensions

If this is right

If the central claim is right, a cheap input transform can make large diffusion SR models trainable on very small specialized datasets, without shrinking the model.
The DroneSR benchmark gives the community a fixed multi-source thermal testbed for reproducing overfitting behavior in few-shot settings.
The reported gains carry over to x3 and x6 upscaling, so the stabilization is not specific to x4.
Validation curves and ablations indicate that the diffusion refiner's contribution is conditioned on the quantization step; without quantization the same pipeline overfits and scores 26.80 dB.

Reading between the lines

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

A testable extension would be to apply Gaussian quantization to other few-shot restoration tasks, such as visible-light, medical, or satellite imagery, to see whether randomized posterization is a general regularizer or specific to thermal sensors.
The paper leaves implicit that quantization is a stochastic augmentation acting on inputs only; a controlled comparison against other augmentations might show that most of the benefit comes from regularization rather than from the diffusion refiner itself.
If the latent diffusion prior transfers as assumed, the method opens a path to zero-shot detail injection for other single-channel modalities; if it does not, the coarse network plus quantization should reproduce most of the reported results, which is a directly testable claim.

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.

Desk Editor's Note

read the letter

Quick take: the DroneSR dataset and the Gaussian quantization idea are real contributions, but the paper's main claim—that the full pipeline beats existing SR methods—is not supported by the write-up. The best evidence is Table V, a matched ablation showing a 2.77 dB PSNR gain from the quantization trick on the same architecture. That part is honest and worth building on. The dataset, if released with proper provenance, could be a useful community resource.

The soft spots, in order of severity. First, the diffusion refinement stage is described unconditionally: Eqs. 7–10 contain no conditioning variable, no cross-attention, no concatenation of I_coarse into the U-Net. As written, the reverse process samples a latent independent of the input, so it cannot refine the coarse image. Either the equations are wrong or the conditioning is omitted; either way, the 30.22 dB result is not reproducible from the description. This is the load-bearing flaw. Second, the training procedure never defines the LR-to-HR pair. Eq. 5 reconstructs I_quant from I_quant at the same spatial size, and Algorithm 2 only applies pixel-value quantization. The x4 spatial downsampling is never specified. Without that, the "super-resolution" task itself is undefined. Third, the baseline comparison is unfair: StableSR and Resshift are run zero-shot with official models, while Ours, DiffBIR, Real-ESRGAN, and LIRSN are (re)trained on the target data. That protocol gap alone could explain a large chunk of the PSNR difference. Real-ESRGAN and StableSR are also evaluated at x3 and x6 using an x4-trained model. Fourth, the abstract promises a "monitoring mechanism" that never appears in the body. That is a credibility problem rather than a technical one.

One minor point: the discrepancy between 30.22 dB in Table II and 29.57 dB in the ablations is actually reconciled by Table VII, where the infrared denoising module adds 0.65 dB. So that particular concern disappears on a close read.

Who this is for: researchers working on few-shot thermal or drone-based super-resolution, and anyone assembling domain-specific SR benchmarks. The dataset and the quantization augmentation are worth knowing about, but the current version is not a reliable report of a working method.

Recommendation: send to peer review, but with the clear expectation of heavy revision. The core materials merit referee time; the current text should not be accepted. Referees should ask for matched training protocols, a correct description of the diffusion conditioning and the degradation model, and release of code and data.

Referee Report

4 major / 5 minor

Summary. The paper proposes DroneSR, a method for few-shot thermal infrared super-resolution from drone imagery. The pipeline applies a Gaussian-quantization corruption to high-quality infrared images, trains a Swin-based coarse restoration network, a latent diffusion texture-refinement stage, and a Restormer-based infrared denoising module. The authors introduce a DroneSR dataset (200 training images) and report quantitative results at x3/x4/x6 against Real-ESRGAN, LIRSN, DiffBIR, StableSR, and Resshift, with ablations attributing 2.77 dB to quantization and 0.65 dB to the denoising module. The central claim is that the proposed quantization method alleviates overfitting in large-scale generative SR models under few-shot conditions.

Significance. The dataset and the idea of using randomized quantization as a data-augmentation/regularizer for diffusion-based SR under scarce data are potentially relevant to the thermal-imaging and few-shot SR communities. The within-architecture ablation (Table V) is a reasonable first step, and the authors state that code will be released. However, the quantitative contribution is not currently verifiable: the diffusion stage is specified as an unconditional model, the LR/HR training pairs for the x4 task are not defined, and the baseline comparison mixes protocols. If the conditioning and task-construction gaps were closed, the method could still be a credible contribution; in its present form, the central claims are not supported.

major comments (4)

[III-C, Eqs. (6)–(10)] The refinement stage is written as an unconditional LDM. z0 = E(Icoarse) in Eq. (6) is used only to set the starting point of the forward noising; Eqs. (7)–(10) contain no conditioning variable c, no cross-attention to E(Icoarse), no channel-wise concatenation of Icoarse in the U-Net, and no inference-time description of how Icoarse enters reverse sampling. An LDM trained with the unconditional objective in Eq. (10) models the marginal latent distribution, so the sampled z0′ is statistically independent of Icoarse and cannot 'refine' it. Since the full pipeline in Table II and the 29.57 dB baseline in Table VII include this diffusion stage, the described architecture cannot produce the reported input-faithful results. The adaptation of the pretrained VAE/LDM (refs [26],[28]) to single-channel thermal images (channel expansion, normalization, conditioning) is also not described.
[III-B, Eq. (5), Algorithm 2] The x4 SR problem is not specified. Algorithm 2 quantizes IHQ to Iquant and passes Iquant directly to CoarseNet; Eq. (5) minimizes ||Icoarse − Iquant||² at the same spatial size. No bicubic downsampling or other operator creating a spatially smaller LR input is described, and no HR target appears in the loss. The text only says that test LR images are 'obtained by downsampling the original image by a factor of four.' As written, Stage 1 is a same-resolution restoration network for quantization artifacts, not a super-resolution network, which makes the training/test pairs behind Tables I–III irreproducible.
[V-B, Tables I–III] The comparison protocols are not matched across methods. StableSR and Resshift are evaluated zero-shot using official pretrained models, while Ours, DiffBIR, Real-ESRGAN, and LIRSN are retrained/reproduced on DroneSR. For x3 and x6, StableSR (and possibly Real-ESRGAN) is tested with a x4-trained model, and Resshift is excluded. Consequently, the reported margins over these baselines conflate method quality with training-data exposure, domain shift, and scale mismatch. The claim that 'our method outperforms existing super-resolution approaches' requires either retraining all baselines on the same DroneSR split at each scale or clearly separating zero-shot and fine-tuned comparisons.
[Fig. 1 and V-C] The paper's headline claim of significantly mitigating overfitting is not directly measured. The only evidence is a descriptive training-curve figure, and no train-vs-validation gap, early-stopping margin, or overfitting statistic is reported. Table V demonstrates a PSNR/SSIM gain from Gaussian quantization, which is a restoration-quality claim, not an overfitting claim. Please report a quantitative overfitting measure (e.g., gap between training and validation PSNR, or performance at matched convergence) or qualify the conclusion accordingly.

minor comments (5)

[V-C, Table VI] The main text sets diffusion steps to 50 as the 'optimal balance,' but Table VI and its surrounding text state that 10 steps give the best overall results (PSNR 29.65 vs 29.57 at 50 steps). Please reconcile this inconsistency.
[V-C, Tables V and VII] The 'w/ Gaussian Quantization' row in Table V reports the same PSNR (29.57) as the 'w/o Infrared Denoising' row in Table VII, indicating that Table V's positive condition omits IDM. Define the ablation configurations explicitly.
[Abstract and Section I] The abstract contains redundant phrasing ('drone-based, diverse drone-based') and the GitHub URL points to GARLSR rather than DroneSR; align the naming.
[Table IV] Minor typo: '27,71' should be '27.71' for Real-ESRGAN's PSNR.
[III-A, Eq. (4)] The definition of p_j as a 'probability' is unclear; specify its distribution (e.g., U[0,1]) and how it relates to the Gaussian sampling used to determine the number of intervals.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: central SR/overfitting claims rest on controlled ablations and external baselines; only a minor non-load-bearing self-citation and hand-tuned quantization hyperparameters on the same benchmark warrant a small score increase.

full rationale

The paper's central claims—that Gaussian quantization mitigates overfitting and improves infrared SR metrics—are supported by controlled within-architecture ablations (Table V: +2.77 dB with vs. without quantization; Table VII: +0.65 dB for the denoising module) and by comparisons against external baselines (Tables I–III). These quantitative conclusions are empirical and are not derived from the method's own assumptions: the quantization transform is an explicit input-side operation, and the ablation compares identical architectures with and without it. The diffusion/refinement component is taken from prior external work (LDM [28]) and is not used to justify its own effectiveness. The only self-citation is [14] (SwinIBSR, same corresponding author), cited in a list of infrared SR DNN methods; it is not load-bearing. Two internal-consistency issues exist but are not circularity: Eq. 5 trains CoarseNet to reconstruct the same-size quantized image without an explicit spatial downsampling step, and Eqs. 7–10 describe an unconditional latent diffusion loss with no conditioning variable tying the sampled latent to Icoarse. These would undermine reproducibility and the described mechanism's relevance to ×4 SR, but they do not make the empirical comparisons circular. Accordingly, no circular step is identified; the score of 2 reflects one minor non-load-bearing self-citation and the hand-setting of quantization hyperparameters on the same benchmark used to demonstrate success.

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

The paper rests on five unproven premises: posterization preserves SR-critical structure; the RGB-pretrained VAE/LDM transfers to thermal imagery; the 50-image test set is adequate; standard LDM/Swin/Restormer behavior is taken as given; and bicubic-style downsampling proxies real drone degradation. The free parameters are the quantization distribution (mu=17, sigma=2, range [7,27], step 2), the 50 diffusion steps, and the early-stopping rule. No invented entities are introduced; 'Gaussian quantization representation learning' is a name for randomized posterization (Algorithm 1, Eqs 1-4), not a new physical or architectural entity.

free parameters (5)

mu = 17 (Gaussian mean for quantization interval count) = 17
Mean of the Gaussian that sets the number of quantization intervals K (Algorithm 1). Chosen by hand in Sec V-A ('we set the mean mu to 17'); no sensitivity analysis, validated only on the DroneSR test set.
sigma = 2 (Gaussian std for interval count) = 2
Standard deviation of the Gaussian for the interval count; hand-set in Sec V-A with the same lack of sensitivity analysis.
K clamp range [7,27] and step size 2 = [7,27], step 2
Arbitrary clamp and rounding in Algorithm 1 control how aggressively pixels are posterized; not derived and never varied in experiments.
Diffusion inference steps = 50 = 50
Chosen for 'balance' (Sec V-A), but the paper's own ablation (Table VI) shows 10 steps is better on PSNR, SSIM, FSIM and MS-SSIM; the main experiments use a suboptimal setting by the authors' own data.
Early-stopping criterion (loss curves showed stable convergence)
Iteration counts for both stages (100k coarse, 5k diffusion, 100k denoising) were fixed by inspecting validation loss curves (Sec V-A); a hand-applied model-selection rule that is not formalized.

assumptions (5)

domain assumption Posterization (pixel quantization) removes noise while preserving the structural information needed for super-resolution.
Sec III-A and Fig 5 assert this without measurement; the central mechanism's value depends entirely on it.
domain assumption The RGB-pretrained VAE/LDM latent space is a valid generative prior for single-channel thermal infrared images, and the diffusion refinement is effectively connected to the coarse input.
Sec III-C uses a pretrained VAE encoder (ref [26]) and LDM prior (ref [28]) with no domain adaptation and no described conditioning mechanism in Eqs 7-10.
domain assumption A 50-image test set drawn from the collected benchmark is sufficient to rank methods.
Sec IV describes a 50-image test set with no confidence intervals or significance tests (Tables I-III).
standard math Standard LDM/DDPM objective (Eq 10) and the Swin Transformer / Restormer components behave as in their original papers.
The paper adopts Eq 10 from LDM [28] and uses pretrained components without modification; no new math is proved.
domain assumption Bicubic-style x4 downsampling of HR images is a valid proxy for real drone low-quality infrared inputs.
Sec V-B states the LR test image is 'obtained by downsampling the original image by a factor of four', which conflicts with the real-world degradation framing of Secs I and III.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DroneSR: Rethinking Few-shot Thermal Image Super-Resolution from Drone-based Perspective." pith.science (2026). https://pith.science/paper/JX56TXST

@misc{pith2026250901898,
  author       = {Pith},
  title        = {Pith review of: DroneSR: Rethinking Few-shot Thermal Image Super-Resolution from Drone-based Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JX56TXST}},
  note         = {Machine review of arXiv:2509.01898}
}

read the original abstract

Although large scale models achieve significant improvements in performance, the overfitting challenge still frequently undermines their generalization ability. In super resolution tasks on images, diffusion models as representatives of generative models typically adopt large scale architectures. However, few-shot drone-captured infrared training data frequently induces severe overfitting in large-scale architectures. To address this key challenge, our method proposes a new Gaussian quantization representation learning method oriented to diffusion models that alleviates overfitting and enhances robustness. At the same time, an effective monitoring mechanism tracks large scale architectures during training to detect signs of overfitting. By introducing Gaussian quantization representation learning, our method effectively reduces overfitting while maintaining architecture complexity. On this basis, we construct a multi source drone-based infrared image benchmark dataset for detection and use it to emphasize overfitting issues of large scale architectures in few sample, drone-based diverse drone-based image reconstruction scenarios. To verify the efficacy of the method in mitigating overfitting, experiments are conducted on the constructed benchmark. Experimental results demonstrate that our method outperforms existing super resolution approaches and significantly mitigates overfitting of large scale architectures under complex conditions. The code and DroneSR dataset will be available at: https://github.com/wengzp1/GARLSR.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

[26]

Auto-encoding variational bayes,

D. P. Kingma, M. Welling et al. , “Auto-encoding variational bayes,” 2013

2013
[28]

High- resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

2022
[1]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

2020
[2]

Improved denoising diffusion probabilis- tic models,

A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilis- tic models,” in International conference on machine learning . PMLR, 2021, pp. 8162–8171

work page 2021
[3]

Diffusion models beat gans on image synthesis,

P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems , vol. 34, pp. 8780–8794, 2021

2021
[4]

Repaint: Inpainting using denoising diffusion probabilistic models,

A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, “Repaint: Inpainting using denoising diffusion probabilistic models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11 461–11 471

work page 2022
[5]

Exploiting diffusion prior for real-world image super-resolution,

J. Wang, Z. Yue, S. Zhou, K. C. Chan, and C. C. Loy, “Exploiting diffusion prior for real-world image super-resolution,” International Journal of Computer Vision , vol. 132, no. 12, pp. 5929–5949, 2024

work page 2024
[6]

Diffbir: Toward blind image restoration with generative diffusion prior,

X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, Y . Qiao, W. Ouyang, and C. Dong, “Diffbir: Toward blind image restoration with generative diffusion prior,” in European Conference on Computer Vision. Springer, 2024, pp. 430–448

work page 2024

Show all 31 references

[8]

Tinyfl hkd: Enhancing edge ai federated learning with hierarchical knowledge distillation framework,

C.-W. Hung, C.-Y . Tsai, C.-C. Wang, and C.-H. Lee, “Tinyfl hkd: Enhancing edge ai federated learning with hierarchical knowledge distillation framework,” IEEE Sensors Journal , 2025

2025
[9]

Infrared detectors: an overview,

A. Rogalski, “Infrared detectors: an overview,” Infrared physics & technology, vol. 43, no. 3-5, pp. 187–210, 2002

2002
[10]

Comparison of infrared and visible imagery for object tracking: Toward trackers with superior ir performance,

E. Gundogdu, H. Ozkan, H. Seckin Demir, H. Ergezer, E. Akagunduz, and S. Kubilay Pakin, “Comparison of infrared and visible imagery for object tracking: Toward trackers with superior ir performance,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognit...

2015
[11]

Desra: detect and delete the artifacts of gan-based real-world super-resolution models,

L. Xie, X. Wang, X. Chen, G. Li, Y . Shan, J. Zhou, and C. Dong, “Desra: detect and delete the artifacts of gan-based real-world super-resolution models,” arXiv preprint arXiv:2307.02457 , 2023

2023 arXiv
[12]

Super- resolution reconstruction of infrared images based on a convolutional neural network with skip connections,

Y . Zou, L. Zhang, C. Liu, B. Wang, Y . Hu, and Q. Chen, “Super- resolution reconstruction of infrared images based on a convolutional neural network with skip connections,” Optics and Lasers in Engineer- ing, vol. 146, p. 106717, 2021

2021
[13]

Infrared image super-resolution via heterogeneous convolutional wgan,

Y . Huang, Z. Jiang, Q. Wang, Q. Jiang, and G. Pang, “Infrared image super-resolution via heterogeneous convolutional wgan,” in PRICAI 2021: Trends in Artificial Intelligence: 18th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2021, Hanoi, Vietnam, No...

2021
[14]

Swinibsr: Towards real- world infrared image super-resolution,

Y . Shi, N. Chen, Y . Pu, J. Zhang, and L. Yao, “Swinibsr: Towards real- world infrared image super-resolution,” Infrared Physics & Technology, vol. 139, p. 105279, 2024

2024
[15]

Infrared thermal imaging super-resolution via multiscale spatio-temporal feature fusion network,

W. Zhang, X. Sui, G. Gu, Q. Chen, and H. Cao, “Infrared thermal imaging super-resolution via multiscale spatio-temporal feature fusion network,” IEEE Sensors Journal , vol. 21, no. 17, pp. 19 176–19 185, 2021

2021
[16]

Swinipisr: A super-resolution method for infrared polarization imaging sensors via swin transformer,

X. Wu, B. Zhou, X. Wang, J. Peng, P. Lin, R. Cao, and F. Huang, “Swinipisr: A super-resolution method for infrared polarization imaging sensors via swin transformer,” IEEE Sensors Journal, vol. 24, no. 1, pp. 468–477, 2023

2023
[17]

Taylor-guided iterative gradient projection neural network for coal-dust scanning electron microscopy super resolution,

X. An, Z. Wang, S. Teng, and Q. Liang, “Taylor-guided iterative gradient projection neural network for coal-dust scanning electron microscopy super resolution,” IEEE Sensors Journal , 2024

2024
[18]

Improving the spatial resolution of small satellites by implementing a super-resolution algo- rithm based on the optical imaging sensor’s rotation approach,

I. Kazemi, M. A. Najafabadi, and M. Shafiee, “Improving the spatial resolution of small satellites by implementing a super-resolution algo- rithm based on the optical imaging sensor’s rotation approach,” IEEE Sensors Journal, 2024

2024
[19]

Incorporating degradation estimation in light field spatial super-resolution,

Z. Xiao and Z. Xiong, “Incorporating degradation estimation in light field spatial super-resolution,” Computer Vision and Image Understand- ing, vol. 252, p. 104295, 2025

2025
[20]

Event- adapted video super-resolution,

Z. Xiao, D. Kai, Y . Zhang, Z.-J. Zha, X. Sun, and Z. Xiong, “Event- adapted video super-resolution,” in European Conference on Computer Vision. Springer, 2024, pp. 217–235

2024
[21]

Event-based video super-resolution via state space models,

Z. Xiao and X. Wang, “Event-based video super-resolution via state space models,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12 564–12 574. 10 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2025

2025
[22]

Deep gate information bottleneck-based prediction model for complex disease-related micro- ribonucleic acids via heterogeneous biological networks,

Y . Guo, Y . Xin, J. Cao, Y . Xu, and D. Zhou, “Deep gate information bottleneck-based prediction model for complex disease-related micro- ribonucleic acids via heterogeneous biological networks,” Engineering Applications of Artificial Intelligence , vol. 156, p. 111222, 2025

2025
[23]

Learning spatiotemporal embedding with gated convolutional recurrent networks for translation initiation site prediction,

W. Li, Y . Guo, B. Wang, and B. Yang, “Learning spatiotemporal embedding with gated convolutional recurrent networks for translation initiation site prediction,” Pattern Recognition , vol. 136, p. 109234, 2023

2023
[24]

Denoising diffusion implicit models,

J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

2010 arXiv
[25]

Image super-resolution via iterative refinement,

C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,” IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 4, pp. 4713–4726, 2022

2022
[27]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 ...

2015
[29]

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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

2022
[30]

Real-esrgan: Training real- world blind super-resolution with pure synthetic data,

X. Wang, L. Xie, C. Dong, and Y . Shan, “Real-esrgan: Training real- world blind super-resolution with pure synthetic data,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 1905–1914

2021
[31]

Lirsrn: A lightweight infrared image super-resolution network,

C.-A. Lin, T.-J. Liu, and K.-H. Liu, “Lirsrn: A lightweight infrared image super-resolution network,” in 2024 IEEE International Symposium on Circuits and Systems (ISCAS) . IEEE, 2024, pp. 1–5

2024
[32]

Resshift: Efficient diffusion model for image super-resolution by residual shifting,

Z. Yue, J. Wang, and C. C. Loy, “Resshift: Efficient diffusion model for image super-resolution by residual shifting,” Advances in Neural Information Processing Systems , vol. 36, pp. 13 294–13 307, 2023. Zhipeng Weng received the B.S. degree in 2023, from the School of Inform...

2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [26]

Auto-encoding variational bayes,

D. P. Kingma, M. Welling et al. , “Auto-encoding variational bayes,” 2013

2013

[2] [28]

High- resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

2022

[3] [1]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

2020

[4] [2]

Improved denoising diffusion probabilis- tic models,

A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilis- tic models,” in International conference on machine learning . PMLR, 2021, pp. 8162–8171

work page 2021

[5] [3]

Diffusion models beat gans on image synthesis,

P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems , vol. 34, pp. 8780–8794, 2021

2021

[6] [4]

Repaint: Inpainting using denoising diffusion probabilistic models,

A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, “Repaint: Inpainting using denoising diffusion probabilistic models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11 461–11 471

work page 2022

[7] [5]

Exploiting diffusion prior for real-world image super-resolution,

J. Wang, Z. Yue, S. Zhou, K. C. Chan, and C. C. Loy, “Exploiting diffusion prior for real-world image super-resolution,” International Journal of Computer Vision , vol. 132, no. 12, pp. 5929–5949, 2024

work page 2024

[8] [6]

Diffbir: Toward blind image restoration with generative diffusion prior,

X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, Y . Qiao, W. Ouyang, and C. Dong, “Diffbir: Toward blind image restoration with generative diffusion prior,” in European Conference on Computer Vision. Springer, 2024, pp. 430–448

work page 2024