Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Multi-exposure HDR reconstruction succeeds in one diffusion step by refining an 8-bit gain map instead of generating full HDR content.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 15:31 UTC pith:65JLQ6WY

load-bearing objection Solid engineering contribution with a load-bearing decoder fix, but the 'outperforms SOTA' claim is not yet statistically supported. the 3 major comments →

arxiv 2512.16357 v3 pith:65JLQ6WY submitted 2025-12-18 cs.CV

GMODiff: One-Step Gain Map Refinement with Diffusion Priors for HDR Reconstruction

Tao Hu , Weiyu Zhou , Yanjie Tu , Peng Wu , Wei Dong , Qingsen Yan , Yanning Zhang This is my paper
classification cs.CV
keywords HDR reconstructionmulti-exposure imaginggain maplatent diffusion modelone-step denoisingregression priordeghostingimage restoration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that diffusion models fail at HDR reconstruction because they try to generate full HDR radiance from a latent space built for 8-bit LDR images. Its fix is to split the HDR image into an LDR base layer and an 8-bit gain map that encodes the extra dynamic range, then let a pre-trained latent diffusion model refine that gain map. A regression network supplies the starting estimate and a degradation-aware embedding, so the diffusion process needs only a single denoising step. The authors show that this setup outperforms prior HDR reconstruction methods on perceptual metrics while running roughly 100x faster than earlier diffusion-based approaches, making perceptual-quality HDR reconstruction practical.

Core claim

The central claim is that gain maps—not HDR images—are the right surface for diffusion-based HDR reconstruction. Because a gain map has the same 8-bit depth as an LDR image, it can pass through a pre-trained LDM's latent space without retraining the autoencoder. GMODiff first trains a degradation-aware regressor to produce an initial gain map and an embedding that marks unreliable regions, then fine-tunes the LDM with low-rank adapters to perform one-step denoising seeded from that estimate. A degradation-aware decoder injects encoder features under the guidance of the regression embedding to recover fine detail that the compressed latent space would otherwise lose. The paper reports that th

What carries the argument

The gain map is the load-bearing object: an 8-bit per-pixel multiplier (with logarithmic compression) that converts an LDR base layer into an HDR image. Because it is 8-bit, it can be encoded, denoised, and decoded by a pre-trained latent diffusion model without domain mismatch. The framework's second essential piece is the degradation-aware regressor, which supplies the initial gain map and a spatial embedding of where that estimate is unreliable; that embedding conditions both the one-step denoising and the decoder's feature injection.

Load-bearing premise

The load-bearing premise is that the pre-trained LDM's heavily compressed 4-channel latent space, patched by the degradation-aware decoder's injected encoder features, can faithfully represent and refine gain-map structure on unseen scenes.

What would settle it

Compute the average absolute difference between the refined gain map and the ground-truth gain map on a held-out set of high-motion scenes; if the one-step diffusion output is not closer to the ground truth than the initial regression estimate, the diffusion refinement is not genuinely refining the gain map.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Pre-trained latent diffusion models can be adapted to HDR imaging without retraining their autoencoders, avoiding the need for large HDR training sets.
  • One-step denoising from a regression estimate cuts inference time by about two orders of magnitude relative to earlier diffusion-based HDR methods, making the perceptual benefits of diffusion practical for real-time use.
  • Regression-guided denoising suppresses hallucination while preserving structural accuracy, suggesting a general recipe for applying diffusion priors to ill-posed inverse problems.
  • The gain-map representation ties HDR reconstruction to any task expressed as a per-pixel gain, so the same pipeline can be transferred to exposure correction and inverse tone mapping.
  • Perceptual quality metrics improve over prior state-of-the-art methods without sacrificing distortion-based fidelity.

Where Pith is reading between the lines

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

  • A likely extension is HDR video: since the gain map is a compact 8-bit representation, the same one-step refinement could be run per-frame with temporal consistency enforced by conditioning on neighboring frames.
  • The paper's own ablation shows that removing the decoder's feature injection drops PU-PSNR from 41.95 to 35.78; a fair inference is that the compressed latent space alone cannot carry the fine detail, and any latent-diffusion restoration method with similar compression will need an analogous detail-injection mechanism.
  • The group-training strategy—diversifying the regressor's training to make the diffusion model robust to its errors—could be applied to other regression-to-generative pipelines as a cheap way to improve robustness.
  • One can test whether the gain-map approach generalizes to higher dynamic ranges than the 8-bit map can represent; if the range exceeds roughly 10 stops, the representation may saturate and the reconstruction would need a two-scale gain map.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes GMODiff, a two-stage HDR reconstruction framework. Stage 1 trains a NAFNet-based degradation-aware regressor (DaReg) to predict an initial 8-bit gain map and degradation-aware embeddings/mask from three LDR exposures. Stage 2 fine-tunes a pre-trained Stable Diffusion 2.1 latent diffusion model with LoRA for one-step refinement of the gain-map latent, followed by a proposed degradation-aware decoder that fuses VAE encoder features. Experiments on a combined 78-scene test set report improved LPIPS/DISTS/MANIQA/MUSIQ/CLIPIQA relative to nine SOTA baselines, at 0.89s per 1080×1920 image versus 13.52s for DiffHDR and 93.4s for UltraFusion.

Significance. The idea of reformulating HDR reconstruction as gain-map refinement is well motivated and technically coherent, and the use of an 8-bit gain map to avoid VAE retraining is a clever practical choice. The two-stage design (degradation-aware regression prior + one-step latent diffusion + detail-preserving decoder) is a plausible way to trade off fidelity and perceptual quality. The authors provide a code link and report many training details, which supports reproducibility. If the claimed margins are statistically robust, the speed advantage over prior diffusion-based HDR methods is practically valuable. However, the current experimental validation is not strong enough to establish the headline quantitative claim.

major comments (3)
  1. [4.2 / Table 1] The central 'outperforms SOTA' claim rests on a single pooled test set of 78 scenes with no per-dataset breakdown, no confidence intervals, and no significance tests. Several reported margins are small: MUSIQ 60.81 vs 60.28 for DiffHDR, CLIPIQA 0.4211 vs 0.4089 for SAFNet, LPIPS 0.0346 vs 0.0361 for SAFNet. With n=78 these differences could easily be sampling noise. Moreover, the method is slightly below AFUNet on PU-PSNR (41.95 vs 42.08) and PU-SSIM (0.9913 vs 0.9914), so the claim should be scoped to perceptual metrics. Please provide per-dataset results, error bars/paired significance tests, and details on how the baselines were retrained/tuned.
  2. [4.1 / Evaluation Metrics] MANIQA, MUSIQ, and CLIPIQA are no-reference SDR metrics computed in a μ-law tone-mapped domain. These metrics are known to be sensitive to tone-curve/contrast style and are not validated for HDR radiance fidelity. Since the main quantitative advantages are exactly on these metrics, the paper should either (a) report an HDR-aware perceptual metric or a user study, or (b) show that the gains persist when tone-mapping style is controlled. Without this, the 'perceptual superiority' claim is vulnerable to the objection that the method wins by producing a particular tone-mapping rather than better HDR content.
  3. [Eq. (6) / Sec. 3.3] The one-step denoising rule is not clearly derived. Z_L is defined as the clean VAE encoding of the initial gain map, but Eq. (6) has the form of the standard DDPM inversion formula for a noised sample x_T at timestep T. If Z_L is clean, the formula is not the usual denoising step; if Z_L is meant to be a noised sample, then the noise-adding process is not described. Please specify the forward noising convention, whether T is fixed or sampled, how \bar{\alpha}_T is determined, and the exact training objective for \epsilon_\theta. As written, this part of the method is not reproducible.
minor comments (6)
  1. [Figure 1] The caption and labels are confusing. The times (0.89s, 93.4s, etc.) appear to be inference latencies, but their placement in the figure relative to method names is hard to parse. Please redraw or annotate clearly.
  2. [Eq. (1)] Q_max is not defined numerically; specify its value/range used in the experiments. Also state explicitly how the gain-map formula in Eq. (1) relates to the μ-law formulation used for supervision.
  3. [Eq. (5)] The 'ground-truth' mask is computed from the current estimate \hat H, so it is a moving pseudo-label. Please state whether \hat H is detached and discuss the stability of this bootstrap procedure.
  4. [Sec. 3.3 / Eq. (10)] The number of groups N in the group training strategy is not specified in the implementation details. Please report N and how the N+1 DaRegs are initialized.
  5. [Table 3] Inference time is reported for a single image. Please report standard deviation over multiple runs and specify the exact GPU model for all methods, including the baselines, so the speed comparison is reproducible.
  6. [Throughout] Minor typographical issues: an extra bracket in 'References [33]]' in Related Work, 'funtion' in Eq. (4), and inconsistent capitalization 'LoRa/LoRA'.

Circularity Check

0 steps flagged

No significant circularity found: GMODiff's claims are supported by supervised training and held-out evaluation rather than by definitional reduction or self-citation chains.

full rationale

GMODiff's derivation chain is an empirical, supervised pipeline rather than a self-referential one. Eq. (1) is a standard dual-layer HDR/gain-map decomposition (cited to Adobe/Apple/Google), used to construct training targets; the DaReg regressor is trained with explicit L1 and reconstruction losses (Eqs. 3-4); the diffusion stage is fine-tuned with LoRA and trained with Eq. (9), including an adversarial loss; and final claims are evaluated on 78 held-out test scenes against retrained baselines. No 'prediction' is defined from the model's own outputs in a load-bearing way. The degradation-aware mask in Eq. (5) is a heuristic pseudo-label derived from the first-stage reconstruction error; although self-referential, it is an auxiliary training target, not a derived result offered as evidence, and the paper does not treat it as independent ground truth. The ablation showing that removing the DA Decoder drops PU-PSNR from 41.95 to 35.78 honestly demonstrates the component's role rather than reducing the central claim to an input. Self-citations ([17], [41], [44]) appear as baselines or as an architectural choice for implicit alignment, not as justification of the main result. The statistical-significance concern about Table 1 is a correctness/robustness issue, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 1 invented entities

The method rests on domain assumptions about the LDM latent space and data representativeness, not on new free parameters. No new physics or entities are introduced; the degradation-aware mask is an algorithmic invention that cannot be independently verified.

axioms (4)
  • domain assumption Pre-trained Stable Diffusion 2.1 latent space is a sufficient representation for gain maps after LoRA fine-tuning.
    Required for the whole method; the VAE is not retrained and gain maps are encoded directly into the LDM latent space.
  • domain assumption The collected set of 836 pairs from [18, 20, 34] is representative enough to compare methods trained from scratch on it.
    All compared methods are retrained on the same collected dataset, but the dataset is a mix of sources and no per-source quantitative breakdown is given.
  • domain assumption Perceptual metrics (MANIQA, MUSIQ, CLIPIQA) computed in the mu-law domain are reliable indicators of HDR reconstruction quality.
    Used as primary evidence of improvement, but these metrics were not designed for HDR and the paper applies them after tone-mapping.
  • domain assumption The one-step denoising formula Eq. (6) is a valid approximation of the multi-step denoising process.
    Adopted directly from prior one-step diffusion work [40]; the paper provides no new derivation.
invented entities (1)
  • Degradation-aware mask M and embeddings c_L no independent evidence
    purpose: Spatially weight the diffusion denoising and DA decoder to focus on unreliable regions of the initial gain map.
    Introduced here as an internal component; no external falsifiable prediction.

pith-pipeline@v1.3.0-alltime-deepseek · 12546 in / 6341 out tokens · 49931 ms · 2026-08-03T15:31:19.663215+00:00 · methodology

0 comments
read the original abstract

Pre-trained Latent Diffusion Models (LDMs) have recently shown strong perceptual priors for low-level vision tasks, making them a promising direction for multi-exposure High Dynamic Range (HDR) reconstruction. However, directly applying LDMs to HDR remains challenging due to: (1) limited dynamic-range representation caused by 8-bit latent compression, (2) high inference cost from multi-step denoising, and (3) content hallucination inherent to their generative nature. To address these challenges, we introduce GMODiff, a gain map-driven one-step diffusion framework for multi-exposure HDR reconstruction. Instead of reconstructing full HDR content, we reformulate HDR reconstruction as a degradation-aware Gain Map (GM) refinement problem, where the GM encodes the extended dynamic range while retaining the same bit depth as LDR images. We initialize the denoising process from an informative regression-based estimate rather than pure noise, allowing the model to generate high-quality GMs in a single denoising step. Furthermore, recognizing that regression-based models excel in content fidelity while LDMs favor perceptual quality, we leverage regression priors to guide both the denoising process and latent decoding of the LDM, suppressing hallucinations while preserving structural accuracy. Extensive experiments demonstrate that GMODiff performs favorably against several state-of-the-art methods and is 100x faster than previous LDM-based methods. Code is available at https://github.com/gbymat/GMODiff.

Figures

Figures reproduced from arXiv: 2512.16357 by Peng Wu, Qingsen Yan, Tao Hu, Wei Dong, Weiyu Zhou, Yanjie Tu, Yanning Zhang.

Figure 1
Figure 1. Figure 1: Compared to DNN-based methods, LDM-based method [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Limitations of the vanilla LDMs in generating HDR [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparisons are conducted on testing data, focusing on zoomed-in local areas of the HDR images estimated by our [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparisons are conducted on testing data , focusing on zoomed-in local areas of the HDR images estimated by our [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of our ablation study. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

46 extracted references · 5 linked inside Pith

  1. [1]

    Gain map specification.https : / / helpx

    Adobe. Gain map specification.https : / / helpx . adobe . com / camera - raw / using / gain - map . html, 2024. Accessed: 2025-11-13. 2

    2024

  2. [2]

    Explore hdr rendering with edr.https : / / developer

    Apple. Explore hdr rendering with edr.https : / / developer . apple . com / videos / play / wwdc2021 / 10161/, 2021. Accessed: 2025-11-13. 2, 3

    2021

  3. [3]

    Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr

    Maryam Azimi et al. Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr. In 2021 Picture Coding Symposium (PCS), pages 1–5. IEEE,

    2021

  4. [4]

    Extending dynamic range of monochrome and color images through fusion

    Luca Bogoni. Extending dynamic range of monochrome and color images through fusion. InProceedings 15th In- ternational Conference on Pattern Recognition. ICPR-2000, pages 7–12. IEEE, 2000. 2

    2000

  5. [5]

    Adversarial diffusion compression for real-world image super-resolution

    Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 28208–28220, 2025. 2

    2025

  6. [6]

    Attention-guided progressive neural tex- ture fusion for high dynamic range image restoration.IEEE Transactions on Image Processing, 31:2661–2672, 2022

    Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural tex- ture fusion for high dynamic range image restoration.IEEE Transactions on Image Processing, 31:2661–2672, 2022. 2

    2022

  7. [7]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InEuropean confer- ence on computer vision, pages 17–33. Springer, 2022. 4, 6

    2022

  8. [8]

    Ultrafusion: Ul- tra high dynamic imaging using exposure fusion

    Zixuan Chen, Yujin Wang, Xin Cai, Zhiyuan You, Zheming Lu, Fan Zhang, Shi Guo, and Tianfan Xue. Ultrafusion: Ul- tra high dynamic imaging using exposure fusion. InProceed- ings of the Computer Vision and Pattern Recognition Confer- ence, pages 16111–16121, 2025. 2, 8

    2025

  9. [9]

    Diffusion models beat gans on image synthesis.NeurIPS, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.NeurIPS, 2021. 1

    2021

  10. [10]

    Superpc: a single diffusion model for point cloud completion, upsampling, de- noising, and colorization

    Yi Du, Zhipeng Zhao, Shaoshu Su, Sharath Golluri, Haoze Zheng, Runmao Yao, and Chen Wang. Superpc: a single diffusion model for point cloud completion, upsampling, de- noising, and colorization. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 16953– 16964, 2025. 2

    2025

  11. [11]

    Artifact-free high dynamic range imag- ing

    Orazio Gallo, Natasha Gelfandz, Wei-Chao Chen, Marius Tico, and Kari Pulli. Artifact-free high dynamic range imag- ing. In2009 IEEE International conference on computa- tional photography (ICCP), pages 1–7. IEEE, 2009. 2

    2009

  12. [12]

    Ultra hdr image format.https://developer

    Google. Ultra hdr image format.https://developer. android . com / media / platform / hdr - image - format, 2024. Accessed: 2025-11-13. 2, 3, 5

    2024

  13. [13]

    Hdr image generation via gain map decomposed diffusion

    Yuanshen Guan, Ruikang Xu, Yinuo Liao, Mingde Yao, Lizhi Wang, and Zhiwei Xiong. Hdr image generation via gain map decomposed diffusion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17536–17545, 2025. 1, 2

    2025

  14. [14]

    Denoising diffu- sion probabilistic models.NeurIPS, 2020

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

    2020

  15. [15]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 5

    2022

  16. [16]

    Gallo, K

    Jun Hu, O. Gallo, K. Pulli, and Xiaobai Sun. HDR deghost- ing: How to deal with saturation? InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1163–1170, 2013. 2

    2013

  17. [17]

    Gen- erating content for hdr deghosting from frequency view

    Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Gen- erating content for hdr deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 25732–25741, 2024. 2

    2024

  18. [18]

    Deep high dynamic range imaging of dynamic scenes.ACM Trans

    Nima Khademi Kalantari, Ravi Ramamoorthi, et al. Deep high dynamic range imaging of dynamic scenes.ACM Trans. Graph., 36(4):144–1, 2017. 1, 2, 5

    2017

  19. [19]

    High dynamic range video.ACM Transac- tions on Graphics (TOG), 22(3):319–325, 2003

    Sing Bing Kang, Matthew Uyttendaele, Simon Winder, and Richard Szeliski. High dynamic range video.ACM Transac- tions on Graphics (TOG), 22(3):319–325, 2003. 2

    2003

  20. [20]

    Safnet: Selective alignment fusion net- work for efficient hdr imaging

    Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. Safnet: Selective alignment fusion net- work for efficient hdr imaging. InEuropean Conference on Computer Vision, pages 256–273. Springer, 2024. 2, 5, 6, 7

    2024

  21. [21]

    Efficient diffusion as low light enhancer

    Guanzhou Lan, Qianli Ma, Yuqi Yang, Zhigang Wang, Dong Wang, Xuelong Li, and Bin Zhao. Efficient diffusion as low light enhancer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21277–21286, 2025. 2

    2025

  22. [22]

    Afunet: Cross- iterative alignment-fusion synergy for hdr reconstruction via deep unfolding paradigm.arXiv preprint arXiv:2506.23537,

    Xinyue Li, Zhangkai Ni, and Wenhan Yang. Afunet: Cross- iterative alignment-fusion synergy for hdr reconstruction via deep unfolding paradigm.arXiv preprint arXiv:2506.23537,

    Pith/arXiv arXiv

  23. [23]

    Learning gain map for inverse tone mapping

    Yinuo Liao, Yuanshen Guan, Ruikang Xu, Jiacheng Li, Shida Sun, and Zhiwei Xiong. Learning gain map for inverse tone mapping. InThe Thirteenth International Conference on Learning Representations. 2

  24. [24]

    Ghost-free high dynamic range imaging with context-aware transformer

    Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. InEuropean Conference on computer vision, pages 344–360. Springer, 2022. 1, 2, 4, 6, 7

    2022

  25. [25]

    Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps.Ad- vances in neural information processing systems, 35:5775– 5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps.Ad- vances in neural information processing systems, 35:5775– 5787, 2022. 2

    2022

  26. [26]

    Semantic masking with curriculum learning for robust hdr image reconstruction: Z

    Zhangkai Ni, Yang Zhang, Kerui Ren, Wenhan Yang, Hanli Wang, and Sam Kwong. Semantic masking with curriculum learning for robust hdr image reconstruction: Z. ni et al.In- ternational Journal of Computer Vision, pages 1–16, 2025. 2

    2025

  27. [27]

    Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions.IEEE Trans- actions on Image Processing, 30:3885–3896, 2021

    Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson WH Lau. Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions.IEEE Trans- actions on Image Processing, 30:3885–3896, 2021. 2, 6, 7

    2021

  28. [28]

    One-step image translation with text-to-image models.arXiv preprint arXiv:2403.12036, 2024

    Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models.arXiv preprint arXiv:2403.12036, 2024. 3

    Pith/arXiv arXiv 2024

  29. [29]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023. 1

    Pith/arXiv arXiv 2023

  30. [30]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 6

    2022

  31. [31]

    Photorealistic text-to-image diffusion models with deep language understanding.NeurIPS, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.NeurIPS, 2022. 1

    2022

  32. [32]

    Ro- bust patch-based hdr reconstruction of dynamic scenes.ACM Trans

    Pradeep Sen, Nima Khademi Kalantari, Maziar Yaesoubi, Soheil Darabi, Dan B Goldman, and Eli Shechtman. Ro- bust patch-based hdr reconstruction of dynamic scenes.ACM Trans. Graph., 31(6):203–1, 2012. 2

    2012

  33. [33]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1, 2

    Pith/arXiv arXiv 2010

  34. [34]

    Alignment-free hdr deghosting with semantics con- sistent transformer.arXiv preprint arXiv:2305.18135, 2023

    Steven Tel, Zongwei Wu, Yulun Zhang, Barth ´el´emy Heyr- man, C ´edric Demonceaux, Radu Timofte, and Dominique Ginhac. Alignment-free hdr deghosting with semantics con- sistent transformer.arXiv preprint arXiv:2305.18135, 2023. 1, 2, 5, 6, 7, 8

    Pith/arXiv arXiv 2023

  35. [35]

    Lediff: Latent exposure diffusion for hdr generation

    Chao Wang, Zhihao Xia, Thomas Leimkuhler, Karol Myszkowski, and Xuaner Zhang. Lediff: Latent exposure diffusion for hdr generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 453–464,

  36. [36]

    Deep learning for hdr imaging: State-of-the-art and future trends.IEEE transactions on pat- tern analysis and machine intelligence, 44(12):8874–8895,

    Lin Wang and Kuk-Jin Yoon. Deep learning for hdr imaging: State-of-the-art and future trends.IEEE transactions on pat- tern analysis and machine intelligence, 44(12):8874–8895,

  37. [37]

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion.Advances in neural information processing systems, 36: 8406–8441, 2023

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion.Advances in neural information processing systems, 36: 8406–8441, 2023. 3

    2023

  38. [38]

    Deep high dynamic range imaging with large foreground motions

    Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. InProceedings of the European Conference on Computer Vision (ECCV), pages 117–132, 2018. 1, 6

    2018

  39. [39]

    Diffusion- based event generation for high-quality image deblurring

    Xinan Xie, Qing Zhang, and Wei-Shi Zheng. Diffusion- based event generation for high-quality image deblurring. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2194–2203, 2025. 2

    2025

  40. [40]

    Ufogen: You forward once large scale text-to-image gener- ation via diffusion gans

    Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image gener- ation via diffusion gans. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196–8206, 2024. 3, 5

    2024

  41. [41]

    Attention- guided network for ghost-free high dynamic range imaging

    Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hen- gel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention- guided network for ghost-free high dynamic range imaging. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751–1760, 2019. 1, 2, 4, 6, 7, 8

    2019

  42. [42]

    Deep hdr imaging via a non- local network.IEEE Transactions on Image Processing, 29: 4308–4322, 2020

    Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qin- feng Shi, and Yanning Zhang. Deep hdr imaging via a non- local network.IEEE Transactions on Image Processing, 29: 4308–4322, 2020. 6

    2020

  43. [43]

    A unified hdr imaging method with pixel and patch level

    Qingsen Yan, Weiye Chen, Song Zhang, Yu Zhu, Jinqiu Sun, and Yanning Zhang. A unified hdr imaging method with pixel and patch level. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22211–22220, 2023. 1, 2

    2023

  44. [44]

    Towards high- quality hdr deghosting with conditional diffusion models

    Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high- quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Tech- nology, pages 1–1, 2023. 4

    2023

  45. [45]

    Toward high- quality hdr deghosting with conditional diffusion models

    Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Toward high- quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Tech- nology, 34(5):4011–4026, 2023. 2, 6, 7, 8

    2023

  46. [46]

    Arbitrary-steps image super-resolution via diffusion inver- sion

    Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inver- sion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 23153–23163, 2025. 2

    2025