Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Taming Stable Diffusion for Computed Tomography Blind Super-Resolution

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a frozen Stable Diffusion model, guided by LLaVA-Med anatomy captions and low-resolution CT latents through a small side network, outperforms existing CT super-resolution methods on two public datasets by roughly 2…

desk verdict Novel SD-for-CT-SR scheme with LLaVA-Med prompts and ControlNet-style conditioning, but the method as written has a tensor-dimension mismatch in §2.2 that makes the forward pass unbuildable, and the evaluation is synthetic-only with no code. read the letter →

arxiv 2506.11496 v1 pith:F65EZQ36 submitted 2025-06-13 eess.IV cs.CV

Chunlei Li , Yilei Shi , Haoxi Hu , Jingliang Hu , Xiao Xiang Zhu , Lichao Mou This is my paper · ORCID
classification eess.IVcs.CV
keywords CTsuper-resolutionblindStableDiffusionmodelsLLaVA-Medlow-dosemedicalimagesynthesisside-controllingnetwork
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

CT imaging faces a trade-off between resolution and radiation dose, and learning-based super-resolution has been limited by unknown degradations and scarce medical data. This paper claims that a large pre-trained text-to-image model, Stable Diffusion, can be adapted to CT blind super-resolution by conditioning it on both the low-resolution scan and a text caption of the anatomy, generated by LLaVA-Med. Training data are synthesized with the Real-ESRGAN degradation pipeline plus CT-specific clipping and simulated low-dose noise. On the Pancreas and 3D-IRCADb datasets, the reported PSNR gains over the next-best method are about $2.0$ dB at $\times 2$ and $2.2$ dB at $\times 4$, with SSIM improving as well. If these results hold, the framework shows that general-purpose generative priors can be transferred to medical imaging with only a small trainable side network, a step toward high-quality CT at reduced radiation exposure.

What carries the argument

The load-bearing mechanism is the side-controlling strategy: a trainable clone of the Stable Diffusion U-Net encoder, called the condition network $F_{\text{COND}}$, takes the concatenation of the noisy latent $z_t$ and the low-resolution VAE latent $f_{lr}$, along with the CLIP text embedding $f_p$, and skip-connects its per-block outputs into the corresponding decoder blocks of the frozen U-Net as $o_i = D_U^i(o_{i-1}, f_p) + F_{\text{COND}}^i(z'_t, f_p)$. This gives the pre-trained generative model CT-specific spatial constraints from the low-resolution image and semantic constraints from the text while leaving the core generator untouched. The VAE encoder is made learnable to adapt the latent space to CT, and the whole system is trained with the noise-prediction loss of the paper's Equation (7).

What would settle it

Train the pipeline on synthetic degradations and run it on real paired low-dose and high-dose CT volumes from the same patients; if the reported PSNR/SSIM lead over baselines shrinks toward zero, or radiologists judge the reconstructions no more diagnostic, the synthetic-degradation premise fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that Stable Diffusion can be 'tamed' for medical super-resolution by leaving the pre-trained U-Net and text encoder frozen and adding a trainable condition network. In the first stage, LLaVA-Med reads the low-resolution slice and produces a sentence describing the major anatomical structures. In the second stage, the low-resolution slice is encoded by a learnable VAE encoder, concatenated with the noisy latent at each diffusion timestep, and passed together with the CLIP text embedding through the condition network, which adds its features to the decoder blocks of the frozen U-Net. The training loss is the standard noise-prediction objective conditioned on the low-resolution latent, the prompt, and the timestep. The authors report the best quantitative results on both datasets at both scales: on Pancreas, $32.03$ dB PSNR at $\times 2$ and $30.81$ dB at $\times 4$; on 3D-IRCADb, $30.87$ dB and $29.65$ dB. The ablation study attributes a large part of this to the text prompt, whose removal costs more than $2$ dB, and also shows benefits from a learnable visual encoder and pre-trained condition-network initialization.

Load-bearing premise

The evaluation, and the conclusion about reduced radiation dose, depends on the assumption that the Real-ESRGAN degradation plus simulated low-dose noise applied to clipped CT slices faithfully represents what real low-dose CT scans look like.

Editorial extensions

If this is right

  • CT super-resolution can inherit Stable Diffusion's detail-synthesis ability while the core generative network stays frozen, so large-scale medical training data are not required.
  • Fidelity on the two public datasets improves by roughly 2 dB PSNR over the next-best method at both scales, which in a clinical setting would make low-dose acquisitions more diagnostically useful.
  • Text prompts are load-bearing rather than decorative: dropping them costs more than 2 dB on Pancreas, so the semantic captioning stage is part of the method's success, not a minor add-on.
  • The same frozen-diffusion-plus-side-network recipe can be pointed at other imaging modalities by swapping the prompt generator and the degradation model, as the authors propose for MRI and ultrasound.

Reading between the lines

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

  • The paper's clinical promise rests on the synthetic degradation being faithful; a paired low-dose/high-dose CT study would determine whether the ~2 dB advantage transfers to real scanner noise.
  • Because the caption is generated from the same low-resolution image, the text cannot add information about unseen detail; it likely works by biasing reconstruction toward typical anatomy, and a prompt-scrambling probe would separate semantic content from mere conditioning.
  • A text-guided reconstruction that pushes toward typical anatomy may also erase atypical lesions; a reader study with pathology cases is a natural check on this risk.
  • The side-controlling recipe could be reused for other medical inverse problems, such as low-dose PET reconstruction or MRI artifact removal, whenever a diffusion prior and a semantic prompt are available.
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

3 major / 5 minor

Summary. This paper presents a two-stage framework for blind CT super-resolution. In the first stage, LLaVA-Med generates a text description from the low-resolution CT slice. In the second stage, a frozen Stable Diffusion model is adapted with a trainable side-controlling condition network that receives both the low-resolution image encoded by the SD VAE and the text prompt, and injects control signals into the U-Net decoder. The method is trained on simulated low-dose CT slices from Pancreas and 3D-IRCADb and evaluated at ×2 and ×4, reporting PSNR/SSIM gains of roughly 2 dB over the second-best baseline on Pancreas. The authors conclude that the framework can produce high-quality CT images at reduced radiation dose and promise to release code.

Significance. The underlying idea—reusing a large pretrained text-to-image diffusion model for medical imaging via text and image conditioning—is timely and potentially valuable, and the reported gains are large. The manuscript's strengths are the clear problem motivation and the breadth of compared baselines in Table 1. However, the paper currently ships no code, provides no error bars or clinical validation, and, most importantly, contains a conditioning path whose tensor dimensions are unspecified and, as written, incompatible. The empirical claim therefore cannot be independently checked, and the practical claim about reduced radiation dose is not yet supported. If the architectural specification is corrected and the evaluation is strengthened, the contribution would be of interest to the medical imaging community.

major comments (3)
  1. [§2.2, Eqs. (3)–(5)] The conditioning path is not realizable as written. With a VAE downsampling factor of 8, the high-resolution latent z_0 in Eq. (4) has spatial size 64×64 for the 512×512 HR slices, whereas f_lr from Eq. (3) has size 32×32 at ×2 and 16×16 at ×4. The sentence following Eq. (4) states that f_lr and z_t are concatenated to form z'_t, but channel-wise concatenation requires identical spatial dimensions; no resizing, padding, or multi-scale mechanism is described anywhere in §2.2. The same mismatch propagates to Eq. (5), where the condition network outputs are added to U-Net decoder blocks at every level. Please specify the exact alignment/upsampling operation used and state how the condition network accounts for the scale factor; without this, the forward pass and the reported experiments cannot be reproduced.
  2. [§3.1, Table 1] The evaluation is entirely synthetic, and the central practical claim about reduced radiation dose is therefore under-supported. LR images are generated by applying Real-ESRGAN's natural-image degradation pipeline to CT slices clipped to [−135, 215] HU, and low-dose noise is simulated with a single blank-flux setting; no real low-dose CT volumes or clinically acquired pairs are used. Because Real-ESRGAN degradations are not a validated model of CT acquisition physics, the large reported gains (e.g., +2.0056 dB on Pancreas ×2) could be an artifact of the test protocol. Please add a real low-dose/high-dose validation set, or at minimum report patient-level confidence intervals and a statistical test across the 17 and 4 test patients, and discuss the domain gap explicitly.
  3. [§3.3, Table 1 and code statement] The comparison set and reproducibility are insufficient to support the claim of outperforming existing approaches. The text says twelve state-of-the-art methods but lists thirteen, and no recent CT-specific SR methods (e.g., the dual-stream diffusion method [1] or GAN-CIRCLE [3] cited in the introduction) or Stable-Diffusion/ControlNet-based SR baselines are included. No code, pretrained weights, or detailed hyperparameters (e.g., number of diffusion timesteps in training, prompt template, VAE channel handling) are provided, so the results cannot be checked. Please release the code and add the missing baselines or justify their omission.
minor comments (5)
  1. [Eq. (7)] There is a typo in the expectation: E_{z0,flr,t,fp,ϵ∼N}[...] should specify the distribution of ϵ, and the norm expression should read ||ϵ − ϵ_θ(z_t, t, f_lr, f_p)||_2^2 with balanced parentheses.
  2. [§1, contribution statement] The claim of being 'the first attempt to adapt Stable Diffusion for medical image super-resolution' is too strong and should be revised in light of the diffusion-based medical imaging works already cited ([1], [24]–[27]) and the wider literature; please add a more careful related-work discussion.
  3. [§2.2, Eq. (3)] It is not stated how single-channel CT slices are converted to the three-channel input expected by Stable Diffusion's VAE and by CLIP; please specify the replication/windowing procedure and whether the same conversion is applied at train and test time.
  4. [Fig. 3] The grouped bar chart is difficult to read and does not show error bars or numeric labels; please provide exact values in the caption or a companion table, and clarify which configurations correspond to the colors shown.
  5. [§3.1, dataset citations] 3D-IRCADb is cited to TCIA [18] rather than the original dataset publication; please cite the original source and specify the preprocessing and filtering steps used for the 2,823 slices.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method trains a conditional diffusion model and evaluates on held-out data; no claim reduces to its own inputs by construction.

full rationale

The paper is an empirical deep-learning system paper, not a derivation-from-first-principles argument. The training objective in Eq. (7) is a standard conditional noise-prediction loss on low-resolution images, text embeddings, and noisy latents; the reported PSNR/SSIM improvements in Table 1 are measured on held-out test slices, so the central quantitative claim is not a fitted parameter renamed as a prediction. The LLaVA-Med text prompt is generated from the low-resolution image at test time and is a legitimate conditioning input, not a hidden fit of the evaluation metric. The Real-ESRGAN degradation pipeline is a data-synthesis choice; using the same synthetic degradation for training and evaluation creates an external-validity or benchmark-fairness concern about real low-dose CT, but it does not make the method's output equivalent to its input by construction. Any self-citation via reference [21] (LLaVA-Med) is not load-bearing: the vision-language model is an external pretrained component that could in principle be replaced by any comparable VLM, and the paper does not rely on a self-cited uniqueness theorem or ansatz to force its architectural choices. The tensor-shape inconsistency between Eq. (3) (LR latent at 32×32 or 16×16 for ×2/×4 from a 256×256 or 128×128 input) and Eqs. (4)-(5) (64×64 HR latent) is a real implementability and reproducibility gap that must be fixed, but it is a correctness issue, not a circularity: nothing is defined in terms of the target result. Overall, the claimed contributions have independent empirical content and no load-bearing step reduces to its own inputs.

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

The paper introduces no new physical or conceptual entities; it relies on pretrained models (Stable Diffusion, LLaVA-Med, CLIP) as black boxes. The only 'free' choices are the training hyperparameters and the side-network architecture, which are not fitted constants. The main unverified premises are the transferability of natural-image priors to CT and the fidelity of the synthetic degradation model.

assumptions (4)
  • domain assumption The Real-ESRGAN degradation pipeline is an accurate model of low-dose CT degradation.
    Used in Sec. 3.1 to generate all low-resolution inputs; the clinical claim rests on this.
  • domain assumption Stable Diffusion's natural-image prior can be transferred to CT via fine-tuning with the side-controlling network.
    The whole method assumes the pretrained U-Net and VAE provide useful features for CT; this is stated but not proven.
  • domain assumption LLaVA-Med produces semantically correct and useful text descriptions from low-resolution CT images.
    Text prompts are generated from low-resolution inputs in Sec. 2.1; if descriptions are wrong, guidance may be harmful.
  • ad hoc to paper The additive skip-connection condition injection in Eq. 5 preserves the pretrained diffusion model's denoising capability.
    This architectural choice is introduced for this paper and not derived from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Taming Stable Diffusion for Computed Tomography Blind Super-Resolution." pith.science (2026). https://pith.science/paper/F65EZQ36

@misc{pith2026250611496,
  author       = {Pith},
  title        = {Pith review of: Taming Stable Diffusion for Computed Tomography Blind Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F65EZQ36}},
  note         = {Machine review of arXiv:2506.11496}
}
read the original abstract

High-resolution computed tomography (CT) imaging is essential for medical diagnosis but requires increased radiation exposure, creating a critical trade-off between image quality and patient safety. While deep learning methods have shown promise in CT super-resolution, they face challenges with complex degradations and limited medical training data. Meanwhile, large-scale pre-trained diffusion models, particularly Stable Diffusion, have demonstrated remarkable capabilities in synthesizing fine details across various vision tasks. Motivated by this, we propose a novel framework that adapts Stable Diffusion for CT blind super-resolution. We employ a practical degradation model to synthesize realistic low-quality images and leverage a pre-trained vision-language model to generate corresponding descriptions. Subsequently, we perform super-resolution using Stable Diffusion with a specialized controlling strategy, conditioned on both low-resolution inputs and the generated text descriptions. Extensive experiments show that our method outperforms existing approaches, demonstrating its potential for achieving high-quality CT imaging at reduced radiation doses. Our code will be made publicly available.

Figures

Figures reproduced from arXiv: 2506.11496 by the authors.

Figure 1
Figure 1. Overview of our proposed framework. The pipeline consists of two stages: (1) text description generation from low-resolution images using LLaVA-Med, and (2) im￾age super-resolution through the proposed side-controlling strategy, which integrates both the generated text prompts and low-resolution images into the Stable Diffusion model. – We propose a two-stage framework combining LLaVA-Med with Stable Dif￾fusion, ena… view at source ↗
Figure 2
Figure 2. Qualitative results on the 3D-IRCADb dataset with a scale factor of 4. The top-left image shows a zoomed-in region of a high-resolution image, focusing on the liver and its internal structures for better visualization. Subsequent images display super-resolution results from various methods. The bottom-right image presents our proposed method’s output, with red arrows highlighting regions where our approach generates… view at source ↗
Figure 3
Figure 3. Ablation study results on the Pancreas dataset. We evaluate the following configurations: random initialization (yellow) vs. pre-trained initialization (green), and frozen visual encoder (pink) vs. learnable visual encoder (green), for both ×2 and ×4 upscaling factors. Performance is measured using PSNR and SSIM. Our method consistently achieves the best results across all experimental settings. initialization [PIT… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. MedDiT4SR: Tri-Stream Joint Adaptation of Pre-Trained Diffusion Transformers for Medical Image Super-Resolution

    eess.IV 2026-07 conditional novelty 6.0 of 10

    A tri-stream joint-attention adaptation of SD3 diffusion transformers with local and semantic adapters improves medical image super-resolution across five modalities.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Medical Image Computing and Computer Assisted Intervention, pp

    Chu, Y., Zhou, L., Luo, G., Qiu, Z., Gao, X.: Topology-preserving computed to- mography super-resolution based on dual-stream diffusion model. In: Medical Image Computing and Computer Assisted Intervention, pp. 260–270. (2023)

    work page 2023

  2. [3]

    IEEE Transactions on Medical Imaging39(1), 188–203 (2019)

    You, C., Cong, W., Vannier, M.W., Saha, P.K., Hoffman, E.A., Wang, G., Li, G., Zhang, Y., Zhang, X., Shan, H., Li, M., Ju, S., Zhao, Z., Zhang, Z.: CT super- resolution GAN constrained by the identical, residual, and cycle learning ensemble (GAN-CIRCLE). IEEE Transactions on Medical Imaging39(1), 188–203 (2019)

    work page 2019

  3. [2]

    European Radiology29, 6163–6171 (2019)

    Akagi, M., Nakamura, Y., Higaki, T., Narita, K., Honda, Y., Zhou, J., Yu, Z., Akino, N., Awai, K.: Deep learning reconstruction improves image quality of abdominal ultra-high-resolution CT. European Radiology29, 6163–6171 (2019)

    work page 2019

  4. [4]

    Biomedical Signal Pro- cessing and Control73, 103412 (2022)

    Chen, Y., Zheng, Q., Chen, J.: Double paths network with residual information distillation for improving lung CT image super resolution. Biomedical Signal Pro- cessing and Control73, 103412 (2022)

    work page 2022

  5. [5]

    Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A. P., Tejani, A., Totz, J., Wang, Z., Shi, W.: Photo-realistic single image super- resolutionusingagenerativeadversarialnetwork.In:IEEEConferenceonComputer Vision and Pattern Recognition, pp. 4681–4690. (2017)

    work page 2017

  6. [6]

    In: European Conference on Computer Vision Workshops, pp

    Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Loy, C.C.: ES- RGAN: Enhanced super-resolution generative adversarial networks. In: European Conference on Computer Vision Workshops, pp. 63–79. (2018)

    work page 2018

  7. [7]

    In: IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp

    Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 136–144. (2017)

    work page 2017

  8. [8]

    In: European Conference on Computer Vision, pp

    Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: European Conference on Computer Vision, pp. 286–301. (2018)

    work page 2018

Show all 31 references
  1. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pp

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using Swin Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pp. 1833–1844. (2021)

    2021

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Gu, J., Lu, H., Zuo, W., Dong, C.: Blind super-resolution with iterative kernel correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1604–1613. (2019)

    2019

  3. [11]

    In: Advances in Neural Information Processing Systems, pp

    Huang, Y., Li, S., Wang, L., Tan, T.: Unfolding the alternating optimization for blind super resolution. In: Advances in Neural Information Processing Systems, pp. 5632–5643. (2020)

    2020

  4. [12]

    In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision Workshops, pp

    Wang, X., Xie, L., Dong, C., Shan, Y.: Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision Workshops, pp. 1905–1914. (2021)

    2021

  5. [13]

    J., Norouzi, M.: Image super-resolution via iterative refinement

    Saharia, C., Ho, J., Chan, W., Salimans, T., Fleet, D. J., Norouzi, M.: Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence45(4), 4713–4726 (2022) 10 C. Li et al

    2022

  6. [14]

    In: Medical Image Computing and Computer Assisted Intervention, pp

    Chi, J., Sun, Z., Zhao, T., Wang, H., Yu, X., Wu, C.: Low-dose CT image super- resolution network with dual-guidance feature distillation and dual-path content communication. In: Medical Image Computing and Computer Assisted Intervention, pp. 98–108. (2023)

    2023

  7. [15]

    In: Medical Image Computing and Computer Assisted Intervention, pp

    Liu, T., Chen, Z., Li, Q., Wang, Y., Zhou, K., Xie, W., Fang, Y., Zheng, K., Zhao, Z., Liu, S., Yang, W.: MDA-SR: Multi-level domain adaptation super-resolution for wireless capsule endoscopy images. In: Medical Image Computing and Computer Assisted Intervention, pp. 518–527. (2023)

    2023

  8. [16]

    In: International Conference on Learning Repre- sentations, (2023)

    Chen, Z., Zhang, Y., Gu, J., Kong, L., Yang, X.: Recursive generalization trans- former for image super-resolution. In: International Conference on Learning Repre- sentations, (2023)

    2023

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10684–10695. (2019)

    2019

  10. [18]

    Journal of Digital Imaging26(6), 1045–1057 (2013)

    Clark, K.W., Vendt, B.A., Smith, K.E., Freymann, J.B., Kirby, J.S., Koppel, P., Moore, S.M., Phillips, S.R., Maffitt, D.R., Pringle, M., Tarbox, L., Prior, F.W.: The cancer imaging archive (TCIA): Maintaining and operating a public information repository. Journal of Digital Im...

    2013

  11. [19]

    P., Ba, J.: Adam: A method for stochastic optimization

    Kingma, D. P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  12. [20]

    IEEE Transactions on Image Processing13(4), 600–612 (2004)

    Zhou W., Alan C.B., Hamid R.S., Eero P.S.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing13(4), 600–612 (2004)

    2004

  13. [21]

    In: Advances in Neural Information Processing Systems, pp

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: LLaVA-Med: Training a large language-and-vision assistant for biomedicine in one day. In: Advances in Neural Information Processing Systems, pp. 28541–28564. (2023)

    2023

  14. [22]

    Huang, Y., Wang, Q., Omachi, S.: Rethinking degradation: Radiograph super- resolutionviaaid-srgan.In:InternationalWorkshoponMachineLearninginMedical Imaging, pp. 43–52. (2022)

    2022

  15. [23]

    IEEE Journal of Biomedical and Health Informatics27(3), 1409–1418 (2022)

    Zhou, H., Huang, Y., Li, Y., Zhou, Y., Zheng, Y.: Blind super-resolution of 3D MRI via unsupervised domain transformation. IEEE Journal of Biomedical and Health Informatics27(3), 1409–1418 (2022)

    2022

  16. [24]

    In: Medical Image Computing and Computer Assisted Intervention, pp

    Lin, T., Chen, Z., Yan, Z., Yu, W., Zheng, F.: Stable diffusion segmentation for biomedical images with single-step reverse process. In: Medical Image Computing and Computer Assisted Intervention, pp. 656–666. (2024)

    2024

  17. [25]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp

    Tian, J., Aggarwal, L., Colaco, A., Kira, Z., Gonzalez-Franco, M.: Diffuse attend and segment: Unsupervised zero-shot segmentation using Stable Diffusion. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp. 3554–3563. (2024)

    2024

  18. [26]

    A.: Zero-shot low light image enhancement with diffusion prior

    Cho, J., Aghajanzadeh, S., Zhu, Z., Forsyth, D. A.: Zero-shot low light image enhancement with diffusion prior. arXiv preprint arXiv:2412.13401 (2024)

    2024 arXiv

  19. [27]

    ACM Transactions on Graphics,42(6), 1–14 (2023)

    Jiang, H., Luo, A., Fan, H., Han, S., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. ACM Transactions on Graphics,42(6), 1–14 (2023)

    2023

  20. [28]

    In: International Conference on Machine Learning, pp

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, pp. 8748–...

    2021

  21. [29]

    IEEE Transactions on Nuclear Science62(5), 2226–2233 (2015) Taming Stable Diffusion for Computed Tomography Blind Super-Resolution 11

    Zeng, D., Huang, J., Bian, Z., Niu, S., Zhang, H., Feng, Q., Liang, Z., Ma, J.: A simple low-dose X-ray CT simulation from high-dose scan. IEEE Transactions on Nuclear Science62(5), 2226–2233 (2015) Taming Stable Diffusion for Computed Tomography Blind Super-Resolution 11

    2015

  22. [30]

    In: International Conference on Machine Learning, pp

    Alexander Q.N., Prafulla D.: Improved denoising diffusion probabilistic models. In: International Conference on Machine Learning, pp. 8162–8171. (2021)

    2021

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp

    Ji, X., Cao, Y., Tai, Y., Wang, C., Li, J., Huang, F.: Real-world super-resolution via kernel estimation and noise injection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 466–467 (2020)

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.