Pith. sign in

REVIEW 5 major objections 4 minor 51 references

Efficient Burst Super-Resolution with One-step Diffusion

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Burst super-resolution can be reduced to a single distilled diffusion step, cutting runtime to 1.6% of the prior baseline while keeping distortion and perceptual quality scores close to the original.

desk verdict Useful engineering recipe for one-step burst SR, but the 'maintaining SR quality' claim is contradicted on the real BurstSR dataset. read the letter →

arxiv 2507.13607 v1 pith:WJCREFF3 submitted 2025-07-18 cs.CV

Kento Kawai , Takeru Oba , Kyotaro Tokoro , Kazutoshi Akita , Norimichi Ukita This is my paper · ORCID
classification cs.CV
keywords burstsuper-resolutiondiffusionmodelsone-stepknowledgedistillationconsistencyearly-stepskippingEDMsamplerperceptualquality
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

Burst super-resolution aims to turn several photos of the same scene, such as smartphone burst shots, into one high-resolution image. Deterministic burst SR models are fast but tend to oversmooth fine texture, while diffusion models restore sharp detail at the cost of many denoising steps. This paper claims that the two can be combined: start from a fast deterministic burst SR image, add a small amount of noise to put it at an intermediate diffusion step, and distill the diffusion reverse process into a single consistency-model step. The result, E-BSRD, is reported to cut runtime to 1.6% of the 100-step diffusion baseline, from 27.2 seconds to 0.44 seconds per image, while keeping distortion metrics and perceptual metrics close to the baseline. If correct, this makes diffusion-based perceptual-quality burst SR practical for interactive use.

What carries the argument

The load-bearing mechanism is the early-step skip combined with consistency distillation. A deterministic burst SR network (Burstormer or BIPNet) produces an initial high-resolution estimate $x'_0$; Eq. (1) adds Gaussian noise scaled by $\sqrt{1-\bar\alpha_\tau}$ so the estimate lands at the $\tau$-th step of the diffusion trajectory, exactly the trick used by SDEdit. The reverse process, implemented with EDM's second-order ODE solver and conditioned on aligned multi-scale burst features via spatial feature transformation, then refines fine details. A consistency model distilled from this teacher maps the noisy latent directly to the clean image in one step. The paper identifies $\sigma_{\max}$ as the decisive hyperparameter: with the default EDM value the one-step model drifts and changes global color, while $\sigma_{\max}=0.03$ keeps the output anchored to the initial estimate.

What would settle it

Feed E-BSRD a burst whose deterministic init has a deliberately induced systematic defect, such as a strong color cast or extreme blur, and compare the one-step output with the 40-step teacher output. The assumption predicts the teacher can largely correct the defect while the one-step model mostly inherits it; if the one-step model corrects it just as well, the assumption is not the limiting factor. Quantitatively, the gap in FID and LPIPS between $T_{\mathrm{CM}}=1$ and the teacher on such a corrupted-init test set settles the question.

Watch

Extended reading notes

Core claim

The paper's central claim is that a diffusion model for burst SR does not need many denoising steps if the reverse process is initialized from a deterministic burst SR estimate rather than from random noise. E-BSRD takes an initial image from Burstormer or BIPNet, perturbs it with noise via $x'_\tau = \sqrt{\bar\alpha_\tau}\,x'_0 + \sqrt{1-\bar\alpha_\tau}\,\epsilon$, runs a second-order EDM sampler with a small maximum noise level $\sigma_{\max}=0.03$, and distills the resulting teacher into a consistency model so that inference takes a single step. On the SyntheticBurst benchmark, the one-step model keeps the teacher's PSNR, SSIM, and LPIPS and remains close on FID, while runtime falls to 0.44 seconds per image, 1.6% of the BSRD baseline. The paper further argues that the key parameter enabling this is $\sigma_{\max}$, because a tiny added noise keeps the deterministic init close to the target and lets the distilled model act as a fine-detail refiner rather than a generator.

Load-bearing premise

The method relies on the assumption that adding a small amount of diffusion noise to the deterministic initial burst SR image makes it statistically close to the true noisy latent at that diffusion step, so the reverse process can start from it; if the initial image has systematic errors such as oversmoothing or color shifts, the one-step distilled model has no way to undo them.

Editorial extensions

If this is right

  • One-step E-BSRD makes diffusion-based burst SR fast enough for interactive or on-device use, at roughly 0.44 seconds per burst instead of 27 seconds.
  • The early-step skip means the diffusion model only has to refine details, so the quality ceiling is largely set by the deterministic burst SR backbone that supplies the initial image.
  • The $\sigma_{\max}$ sweep gives a practical recipe: tune the maximum noise level to the task rather than using the value designed for image generation from random noise.
  • The observation that consistency-model iterations beyond $T_{\mathrm{CM}}=1$ barely change the output suggests the distilled one-step model has already captured the residual mapping from init to target.

Reading between the lines

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

  • Because the one-step output converges toward the deterministic init as $\sigma_{\max}$ shrinks, we infer that E-BSRD is best understood as a learned residual refiner on top of Burstormer or BIPNet; a direct test would compare it against a simple network trained to predict the residual $\mathrm{HR}-\mathrm{init}$.
  • The paper's numbers on the real BurstSR dataset show a larger FID gap for the distilled one-step model than on SyntheticBurst; a likely consequence is that one-step distillation needs per-dataset $\sigma_{\max}$ and teacher-step tuning, and the real-data gap would grow if the init's color or blur statistics shift.
  • The same early-step skip plus consistency-distillation pattern could transfer to other multi-frame restoration tasks, such as video super-resolution or low-light burst denoising, provided a fast deterministic frame-fusion backbone exists; the paper does not test this.
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

5 major / 4 minor

Summary. The paper proposes E-BSRD, an acceleration of the BSRD burst super-resolution method. E-BSRD combines an EDM-style second-order stochastic sampler with consistency-model distillation so that the reverse diffusion process can run in a single step, starting from an initial burst SR image produced by a deterministic model such as Burstormer or BIPNet. The authors report that E-BSRD reduces runtime to 1.6% of BSRD (0.44 s/image versus 27.2 s/image) while "maintaining the SR quality measured based on image distortion and perceptual quality." Experiments are conducted on SyntheticBurst and the real BurstSR dataset, with PSNR, SSIM, LPIPS, and FID as metrics. The paper also presents sweeps over the noise level sigma_max and the consistency-model iteration count TCM.

Significance. If the central claim were correct, E-BSRD would be a practically important result: it would make diffusion-based burst super-resolution feasible for real-time use while preserving both distortion and perceptual quality. The efficiency gain is clear and the paper includes systematic parameter sweeps and comparisons on standard benchmarks. However, the quality-maintenance claim is contradicted by the paper's own results on the real BurstSR dataset, and the favorable distortion scores at small sigma_max are partly a construction of the initialization procedure rather than a demonstrated diffusion-model advantage. The current contribution is therefore better described as an efficiency-quality tradeoff, and the significance is substantially lower than claimed.

major comments (5)
  1. [§4.4, Table 5] The central claim in the Abstract and Sec. 4.2 that SR quality is "maintained" at 1.6% runtime is not supported by the paper's own results on the real BurstSR dataset. With the BIPNet initializer and TCM=1, E-BSRD reports PSNR 48.29 vs. 50.54 for BSRD, SSIM 0.980 vs. 0.984, LPIPS 0.054 vs. 0.047, and FID 71.80 vs. 43.34; with the Burstormer initializer, PSNR 47.64 vs. 49.45, LPIPS 0.060 vs. 0.050, and FID 77.30 vs. 48.51. The FID regression of roughly 25--30 points is large. Quality is maintained only on the SyntheticBurst distortion metrics, not on the real-image benchmark, so the headline claim as stated is falsified by Table 5.
  2. [§3.2, Eq. (1), Table 2] The improved distortion scores at small sigma_max are substantially a construction artifact. With sigma_max=0.03, Eq. (1) yields x'_tau approximately equal to the initial burst SR image x'_0 up to a tiny perturbation, so the one-step model is constrained to stay near the deterministic initializer. The paper itself concedes in Sec. 4.3 that "the effect of the diffusion model is reduced as sigma_max decreases." Thus the PSNR/SSIM gains over BSRD at sigma_max=0.03 mainly reflect closeness to the initializer, not a diffusion-model advantage. To substantiate the perceptual-quality role of the diffusion model, the authors should compare E-BSRD against the initializer alone at the chosen sigma_max and demonstrate, with proper statistical tests, that the diffusion output is perceptually better than both the initializer and BSRD.
  3. [§4.4, Tables 4--5] The statement in Sec. 4.4 that the quality gaps are "not significant" is unsubstantiated: no confidence intervals, error bars, or statistical tests are reported for any metric. Moreover, sigma_max and tau are selected by sweeping the test sets in Tables 2 and 4 with no held-out validation split, so the reported numbers are optimized on the evaluation set. On the BurstSR test set of 882 images, FID differences of 20--30 points cannot be dismissed as insignificant without a statistical analysis; at minimum, bootstrap confidence intervals should be reported.
  4. [§4.2, Table 1] The runtime comparison is not sufficiently specified to support the 1.6% claim. Table 1 does not state the GPU/CPU hardware, the batch size, or whether the listed BSRD and E-BSRD runtimes include the cost of the initial burst SR model (Burstormer/BIPNet) and the burst feature encoder. Since both methods rely on an initializer/encoder, the quoted 0.44 s/image may not be an end-to-end time, and the claimed savings relative to BSRD may be overstated.
  5. [§3.2, Eq. (1), §4.1] The paper does not specify how the EDM noise parameter sigma_max relates to the DDPM-style \bar{\alpha}_\tau used in Eq. (1), nor how sigma_max is set during training versus inference. Because the entire method depends on noising the initial SR image to the correct intermediate level, this missing mapping prevents reproduction and makes it impossible to assess whether the "optimized" parameters are selected on training or test data.
minor comments (4)
  1. [§4.4] There is a typo in Sec. 4.4: "Bursotrmer" should be "Burstormer." Reference [19] also contains the typo "Deep back-projectinetworks."
  2. [Figure 3] The Figure 3 caption is grammatically awkward: "Images in the lower are reconstructed by EDM conditioned by those in the upper" should be reworded, for example to "The lower images are reconstructed by EDM conditioned on the upper images."
  3. [Table 4] The BSRD (+BS) SSIM value of 0.915 in Table 4 is inconsistent with the surrounding values: Burstormer is 0.985, E-BSRD (+BS) is around 0.98, and BSRD (+BIP) is 0.984. Please check whether this is a typo, as it affects the comparison.
  4. [§4.1] Section 4.1 says "We follow all training details and parameters of the original EDM," while Sec. 3.2 states that sigma_max is empirically optimized for the early-step skipping process. These statements should be reconciled so the reader knows which EDM parameters are changed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained, though the central quality claim is empirically contradicted on the real BurstSR dataset.

full rationale

The paper's derivation chain is not circular. E-BSRD uses Eq. (1) to add diffusion noise to the deterministic initial burst SR image x'_0 and then runs a one-step consistency-model reverse process; the fact that smaller sigma_max makes the output closer to x'_0 is explicitly acknowledged by the authors ('Both PSNR and SSIM improve as σmax decreases... This is natural because the effect of the diffusion model is reduced as σmax decreases'). This is a stated design trade-off, not a concealed identification of output with input. The method is a finetuned/distilled extension of BSRD [42], which is a prior paper by three of the five authors, but BSRD is re-evaluated here with the authors' code and used as a reproducible baseline and teacher, so the self-citation is not load-bearing in a circular way. No uniqueness theorem or ansatz is smuggled in via self-citation; the EDM and consistency-model components are cited from independent external work [22, 40]. The real concern is empirical, not circular: on the BurstSR real-image benchmark, Table 5 shows E-BSRD (+BIP, TCM=1) is worse than BSRD on PSNR (48.29 vs. 50.54), SSIM (0.980 vs. 0.984), LPIPS (0.054 vs. 0.047), and FID (71.80 vs. 43.34), which contradicts the abstract's claim of 'maintaining the SR quality'. That is a correctness/validity problem, not a circularity problem, and it does not involve the paper deriving its conclusion from its own premises by construction.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three tuned hyperparameters (sigma_max=0.03, tau=40, T_CM=1), the early-step skip assumption that noisy initial SR images approximate true diffused latents, and the transferability of EDM and CM training recipes from prior work. No new entities are introduced. The distortion-metric results are partly determined by the choice of small sigma_max, which reduces the diffusion model's contribution toward the deterministic initializer.

free parameters (3)
  • sigma_max = 0.03 (inferred from Tables 2 and 3)
    Maximum noise level for EDM early-step skip; tuned per dataset/test set to balance PSNR and perceptual metrics; not explicitly stated in text.
  • tau (EDM steps) = 40
    Number of EDM steps; chosen based on Fig. 5 to maintain quality while reducing runtime.
  • T_CM (consistency model iterations) = 1
    Number of CM iterative steps; chosen because quality is unchanged for small T_CM (Tables 3 and 5).
assumptions (3)
  • domain assumption x'_tau approximates x_tau when diffusion noise is added to the deterministic initial SR image (Sec. 2.3, Eq. (1)).
    The early-step skip assumes small differences between x'_0 and x_0 are drowned by noise; with small sigma_max this noise is tiny.
  • domain assumption BSRD is a valid teacher for consistency-model distillation under early-step skip.
    The paper asserts CM distillation from a teacher that starts from an initial SR image avoids quality degradation, but provides no theoretical or ablative support beyond the reported numbers.
  • domain assumption EDM preconditioning and training parameters transfer to the burst-SR conditional setting (Sec. 4.1).
    'We follow all training details and parameters of the original EDM' without verifying transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Burst Super-Resolution with One-step Diffusion." pith.science (2026). https://pith.science/paper/WJCREFF3

@misc{pith2026250713607,
  author       = {Pith},
  title        = {Pith review of: Efficient Burst Super-Resolution with One-step Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJCREFF3}},
  note         = {Machine review of arXiv:2507.13607}
}
read the original abstract

While burst Low-Resolution (LR) images are useful for improving their Super Resolution (SR) image compared to a single LR image, prior burst SR methods are trained in a deterministic manner, which produces a blurry SR image. Since such blurry images are perceptually degraded, we aim to reconstruct sharp and high-fidelity SR images by a diffusion model. Our method improves the efficiency of the diffusion model with a stochastic sampler with a high-order ODE as well as one-step diffusion using knowledge distillation. Our experimental results demonstrate that our method can reduce the runtime to 1.6 % of its baseline while maintaining the SR quality measured based on image distortion and perceptual quality.

Figures

Figures reproduced from arXiv: 2507.13607 by the authors.

Figure 1
Figure 1. Comparison between prior SR model and our burst SR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Examples reconstructed by EDM. Images in the lower [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Examples of SR images reconstructed with different [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Effect of the number of diffusion steps, [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: E-BSRD using CM. This figure shows an example with [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Visual results on the SyntheticBurst dataset. Best viewed with zoom-in. While clear textures, such as the boundary lines of texts, [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Visual results on the BurstSR dataset. Best viewed with zoom-in. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 42 canonical work pages

  1. [1]

    Deep burst super-resolution

    Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Deep burst super-resolution. In CVPR, 2021. 1, 5

    work page 2021

  2. [2]

    NTIRE 2021 challenge on burst super- resolution: Methods and results

    Goutam Bhat et al. NTIRE 2021 challenge on burst super- resolution: Methods and results. In CVPRW, 2021. 3

    work page 2021

  3. [3]

    NTIRE 2022 burst super-resolution chal- lenge

    Goutam Bhat et al. NTIRE 2022 burst super-resolution chal- lenge. In CVPRW, 2022. 1, 3

    work page 2022

  4. [4]

    Perception pri- oritized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception pri- oritized training of diffusion models. In CVPR, 2022. 4, 8

    work page 2022

  5. [5]

    Improving diffusion models for inverse prob- lems using manifold constraints

    Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse prob- lems using manifold constraints. In NeurIPS, 2022. 2

    2022

  6. [6]

    Come-closer-diffuse-faster: Accelerating conditional diffu- sion models for inverse problems through stochastic contrac- tion

    Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffu- sion models for inverse problems through stochastic contrac- tion. In CVPR, 2022. 1, 2

    work page 2022

  7. [7]

    Parallel diffusion models of operator and image for blind inverse problems

    Hyungjin Chung, Jeongsol Kim, Sehui Kim, and Jong Chul Ye. Parallel diffusion models of operator and image for blind inverse problems. In CVPR, 2023. 2

    work page 2023

  8. [8]

    Diffusion pos- terior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mc- Cann, Marc Louis Klasky, and Jong Chul Ye. Diffusion pos- terior sampling for general noisy inverse problems. In ICLR,

    work page

Show all 51 references
  1. [9]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 2

    2017

  2. [10]

    Image super-resolution using deep convolutional net- works

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional net- works. IEEE Trans. Pattern Anal. Mach. Intell., 38(2):295– 307, 2016. 1

    2016

  3. [11]

    Burst image restoration and enhancement

    Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fa- had Shahbaz Khan, and Ming-Hsuan Yang. Burst image restoration and enhancement. In CVPR, 2022. 1, 3, 6

    2022

  4. [12]

    Burstormer: Burst image restoration and enhancement transformer

    Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fa- had Shahbaz Khan, and Ming-Hsuan Yang. Burstormer: Burst image restoration and enhancement transformer. In CVPR, 2023. 1, 3, 5, 6

    2023

  5. [13]

    AIM 2020 challenge on video extreme super-resolution: Methods and results

    Dario Fuoli et al. AIM 2020 challenge on video extreme super-resolution: Methods and results. In ECCV Workshop,

    2020

  6. [14]

    AIM 2019 challenge on image extreme super-resolution: Methods and results

    Shuhang Gu et al. AIM 2019 challenge on image extreme super-resolution: Methods and results. In ICCVW, 2019. 1

    2019

  7. [15]

    A differentiable two-stage alignment scheme for burst image reconstruction with large shift

    Shi Guo, Xi Yang, Jianqi Ma, Gaofeng Ren, and Lei Zhang. A differentiable two-stage alignment scheme for burst image reconstruction with large shift. In CVPR, 2022. 1, 2

    2022

  8. [16]

    Deep back-projection networks for super-resolution

    Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In CVPR, 2018. 1

    2018

  9. [17]

    Recurrent back-projection network for video super- resolution

    Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video super- resolution. In CVPR, 2019. 2

    2019

  10. [18]

    Space-time-aware multi-resolution video enhance- ment

    Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Space-time-aware multi-resolution video enhance- ment. In CVPR, 2020. 2

    2020

  11. [19]

    Deep back-projectinetworks for single image super- resolution

    Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Deep back-projectinetworks for single image super- resolution. IEEE Trans. Pattern Anal. Mach. Intell., 43(12): 4323–4337, 2021. 1

    2021

  12. [20]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. In NIPS, 2017. 5

    2017

  13. [21]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In NeurIPS, 2020. 1, 2

    2020

  14. [22]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 1, 2, 4, 5

    2022

  15. [23]

    Denoising diffusion restoration models

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In NeurIPS,

  16. [24]

    Consistency trajectory mod- els: Learning probability flow ODE trajectory of diffusion

    Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Mu- rata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory mod- els: Learning probability flow ODE trajectory of diffusion. In ICLR, 2024. 1, 2

    2024

  17. [25]

    Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi

    Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 1

    2017

  18. [26]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3

    2021

  19. [27]

    Video swin transformer

    Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. InCVPR,

  20. [28]

    EBSR: feature enhanced burst super-resolution with deformable alignment

    Ziwei Luo, Lei Yu, Xuan Mo, Youwei Li, Lanpeng Jia, Hao- qiang Fan, Jian Sun, and Shuaicheng Liu. EBSR: feature enhanced burst super-resolution with deformable alignment. In CVPR Workshops, 2021. 1, 2

    2021

  21. [29]

    BSRT: improving burst super-resolution with swin transformer and flow-guided deformable alignment

    Ziwei Luo, Youwei Li, Shen Cheng, Lei Yu, Qi Wu, Zhihong Wen, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. BSRT: improving burst super-resolution with swin transformer and flow-guided deformable alignment. In CVPR Workshops,

  22. [30]

    Adaptive feature consolidation network for burst super- resolution

    Nancy Mehta, Akshay Dudhane, Subrahmanyam Murala, Syed Waqas Zamir, Salman Khan, and Fahad Shahbaz Khan. Adaptive feature consolidation network for burst super- resolution. In CVPRW, 2022. 1, 3

    2022

  23. [31]

    Sdedit: Guided image synthesis and editing with stochastic differential equa- tions

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equa- tions. In ICLR, 2022. 3

    2022

  24. [32]

    Diffusion model based posterior sampling for noisy linear inverse problems

    Xiangming Meng and Yoshiyuki Kabashima. Diffusion model based posterior sampling for noisy linear inverse problems. CoRR, abs/2211.12343, 2022. 2

    2022 arXiv

  25. [33]

    Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration

    Naoki Murata, Koichi Saito, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Yuki Mitsufuji, and Stefano Ermon. Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration. In ICML, 2023. 2

    2023

  26. [34]

    NTIRE 2019 challenge on video super- resolution: Methods and results

    Seungjun Nah et al. NTIRE 2019 challenge on video super- resolution: Methods and results. In CVPR Workshop, 2019. 2

    2019

  27. [35]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2

    2021

  28. [36]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 2

    2022

  29. [37]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Sali- mans, David J. Fleet, and Mohammad Norouzi. Image super- resolution via iterative refinement.IEEE Trans. Pattern Anal. Mach. Intell., 45(4):4713–4726, 2023. 2

    2023

  30. [38]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In ICLR, 2021. 2

    2021

  31. [39]

    Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. In ICLR, 2021. 2

    2021

  32. [40]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 1, 2, 4, 7

    2023

  33. [41]

    Ntire 2018 challenge on single image super-resolution: Methods and results

    Radu Timofte et al. Ntire 2018 challenge on single image super-resolution: Methods and results. In CVPRW, 2018. 1

    2018

  34. [42]

    Burst super-resolution with diffusion models for improving perceptual quality

    Kyotaro Tokoro, Kazutoshi Akita, and Norimichi Ukita. Burst super-resolution with diffusion models for improving perceptual quality. In IJCNN, 2024. 1, 2, 3, 4, 6, 7, 8

    2024

  35. [43]

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. CoRR, abs/2305.07015,

    arXiv

  36. [44]

    Recovering realistic texture in image super-resolution by deep spatial feature transform

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 3

    2018

  37. [45]

    Kot, and Bihan Wen

    Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C. Kot, and Bihan Wen. Sinsr: Diffusion-based image super- resolution in a single step. In CVPR, 2024. 1, 2

    2024

  38. [46]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600–612, 2004. 5

    2004

  39. [47]

    H2FA R-CNN: holistic and hierarchical feature align- ment for cross-domain weakly supervised object detection

    Yunqiu Xu, Yifan Sun, Zongxin Yang, Jiaxu Miao, and Yi Yang. H2FA R-CNN: holistic and hierarchical feature align- ment for cross-domain weakly supervised object detection. In CVPR, 2022. 3

    2022

  40. [48]

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

    Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super- resolution by residual shifting. In NeurIPS, 2023. 1, 2

    2023

  41. [49]

    Effi- cient diffusion model for image restoration by residual shift- ing

    Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Effi- cient diffusion model for image restoration by residual shift- ing. IEEE Trans. Pattern Anal. Mach. Intell., 47(1):116–130,

  42. [50]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5

    2018

  43. [51]

    Hierarchical instance feature alignment for 2d image- based 3d shape retrieval

    Heyu Zhou, Weizhi Nie, Wenhui Li, Dan Song, and An-An Liu. Hierarchical instance feature alignment for 2d image- based 3d shape retrieval. In IJCAI, 2020. 3

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.