Pith. sign in

REVIEW 3 major objections 5 minor 56 references

DarkDiff: Advancing Low-Light Raw Enhancement by Retasking Diffusion Models for Camera ISP

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 DarkDiff retasks a pretrained diffusion model as the camera ISP and reports the best perceptual quality on three low-light raw benchmarks.

desk verdict Well-built retasking of Stable Diffusion for low-light ISP, but perceptual-SOTA claim rests on LPIPS alone and needs human validation. read the letter →

arxiv 2505.23743 v1 pith:DKDXMDHB submitted 2025-05-29 cs.CV eess.IV

Amber Yijia Zheng , Yu Zhang , Jun Hu , Raymond A. Yeh , Chen Chen This is my paper
classification cs.CVeess.IV
keywords low-lightimageenhancementrawprocessingdiffusionmodelscameraISPperceptualqualityLPIPSlatentdenoising
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that extreme low-light raw enhancement is best done by retasking a pretrained generative diffusion model to act as the camera ISP, rather than by training a regression network or a diffusion model from scratch. DarkDiff first converts noisy Bayer raw data into brightened linear RGB, then uses a diffusion U-Net whose conditioning is a region-based cross-attention over the noisy image, with a residual VAE and a decoder-space loss to preserve structure and color. On the SID, ELD, and LRD benchmarks, the paper reports the lowest LPIPS value among all baselines, i.e., outputs judged by the metric most similar to well-exposed reference images, while PSNR and SSIM remain competitive. The practical stake is that large paired raw datasets are scarce, so reusing the natural-image prior already stored in a pretrained diffusion model could make high-quality low-light photography achievable without such data.

What carries the argument

The argument is carried by three components. Region-based cross-attention divides the noisy latent into non-overlapping patches and computes attention within each region, so the pretrained model can borrow local structure from the noisy input without hallucinating global content. A residual VAE adds skip-like connections from encoder blocks to decoder blocks so the conditioning image's content survives the latent bottleneck. A one-step decoder reconstruction loss, derived from DDPM's posterior mean estimate, supervises the output in sRGB pixel space to suppress color shift. The data pipeline itself is also load-bearing: converting Bayer raw to brightened linear RGB before the network closes the domain gap with the pretrained sRGB model.

What would settle it

A paired human preference experiment on SID, ELD, and LRD outputs comparing DarkDiff against ExposureDiffusion and LRD would settle the claim; if raters prefer the smoother higher-PSNR baselines more often than not, the perceptual-superiority claim fails. A cheaper check is to recompute rankings with a second perceptual metric, such as DISTS or a no-reference sharpness and color metric, and see whether DarkDiff still leads.

Watch

Extended reading notes

Core claim

DarkDiff's central claim is that a pretrained latent diffusion model can be repurposed as a learned camera ISP for extreme low-light raw images, and that the correct recipe is to keep classical ISP steps (white balance, demosaicing, digital gain) outside the network, feed the model linear RGB, and condition it with region-based cross-attention instead of global cross-attention or concatenation. The paper reports that this setup beats state-of-the-art baselines in LPIPS on all three test subsets—SID, ELD, and LRD—while producing sharper textures, better colors, and fewer hallucinations than ExposureDiffusion, LRD, and simple Stable Diffusion conditioning baselines. The authors argue that lower PSNR and SSIM for DarkDiff are expected because those metrics reward smoothing on ill-posed restoration tasks, and they point to a perception-distortion tradeoff rather than treating reconstruction fidelity as the objective.

Load-bearing premise

The strongest claim depends on LPIPS correctly ranking human perceptual quality in the extreme-low-light regime, because DarkDiff has lower PSNR and SSIM than several baselines yet is declared best purely on LPIPS.

Editorial extensions

If this is right

  • If DarkDiff's claim holds, diffusion-based ISP becomes a viable alternative to regression ISP for extreme low light, with the pretrained model supplying texture and detail that pixel-error training washes out.
  • The method's success on LRD, a sensor unseen in training, suggests the retasking approach generalizes across camera sensors better than from-scratch diffusion.
  • Because DarkDiff uses classifier-free guidance, perceptual quality becomes a tunable operating point; increasing guidance trades smoothing for generated detail, so practitioners can dial quality per use case.
  • The best-LPIPS results come at the cost of lower PSNR and SSIM, so DarkDiff is positioned for perceptual use such as display and editing rather than for applications where pixel fidelity is the contract.

Reading between the lines

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

  • The paper does not include a human study; a natural extension would be a forced-choice perceptual experiment on the three test sets to verify that the LPIPS ordering matches what people prefer in extreme low light.
  • Because the reported gains are largest on the darkest, most texture-heavy regions, the method could be stress-tested on non-English text and fine repeating patterns, which the paper itself flags as a limitation.
  • If LPIPS is later shown to reward fabricated texture, the architecture's value may shift from 'more faithful' to 'more generative'; a second perceptual metric or calibration on this domain would settle which.
  • The per-dataset guidance weights (2.0, 2.0, 2.5) are a potential source of hidden tuning; reporting LPIPS across a sweep of guidance values would show how much of the advantage comes from choosing that operating point.
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. The paper proposes DarkDiff, a framework that retasks a pre-trained Stable Diffusion model for extreme low-light raw image enhancement. The input Bayer raw data is converted to linear RGB through an ISP preprocessing pipeline, then encoded by a fine-tuned residual VAE; a region-based cross-attention module conditions the denoising U-Net on local structure, and a decoder-based reconstruction loss reduces color shift. The method is evaluated on the SID, ELD, and LRD datasets, where it reports the lowest LPIPS among the compared methods, while its PSNR and SSIM are below several baselines. The authors claim state-of-the-art perceptual quality based on these LPIPS results.

Significance. If the perceptual-quality claim is accepted, DarkDiff is a meaningful demonstration that large pre-trained text-to-image diffusion models can be adapted to raw low-light ISP tasks, and the region-based cross-attention and residual VAE are transferable design elements. The ablations in Tables 4-6 provide internal support for each component, and the paper is honest about limitations in Sec. A3. However, the central claim is currently supported only by LPIPS, with no human study or second perceptual metric, and the closest prior diffusion-based ISP method, LDM-ISP [50], is absent from the comparisons. These gaps make the SOTA claim conditional rather than fully established.

major comments (3)
  1. [Sec. 5, Tables 1-3, Fig. 9] The paper's central claim of state-of-the-art perceptual quality rests entirely on the LPIPS metric. DarkDiff is below several baselines in PSNR and SSIM on all three datasets (e.g., SID: PSNR 26.78 vs 27.74 for ExposureDiffusion; SSIM 0.644 vs 0.679), so the LPIPS advantage is the only quantitative evidence for the headline claim. LPIPS is a learned deep-feature distance that may reward plausible hallucinated texture over faithful rendering in extreme low-light conditions, where the model must invent content in heavily corrupted regions. Fig. 9 is anecdotal and not a substitute for systematic validation. Please add a human preference study or at least a second perceptual metric (e.g., DISTS or FID), and report confidence intervals, before the claim of perceptual superiority can be considered supported.
  2. [Sec. 2, Tables 1-3] The related work identifies LDM-ISP [50] as the closest prior diffusion-based ISP method, but this method is not included in any quantitative or qualitative comparison. Since the abstract claims state-of-the-art performance against diffusion-based approaches, omitting the most directly comparable prior method is a significant gap. Please add LDM-ISP to the baseline tables and qualitative figures, or explain concretely why a fair comparison is not possible.
  3. [Sec. A2, inference] The classifier-free guidance weights are reported as 2.0 for SID and LRD and 2.5 for ELD, but the paper does not describe how these values were selected. If they were tuned on the test sets, the reported LPIPS numbers are optimized operating points and the comparison to baselines, which likely use fixed hyperparameters from their original papers, would be unfair. Please state the selection procedure (e.g., a validation split) and include a sensitivity analysis of the results with respect to the guidance weight.
minor comments (5)
  1. [Table 4] The checkmarks in the ablation table are visually ambiguous because each row appears as '✓ ✓' with no explicit indication of which component is removed. Since the text says the first row replaces region-based cross-attention with concatenation, please use explicit ✓/× symbols under each column label so that the reader can see which components are ablated in each row.
  2. [Table 1] The SSIM value for LED is listed as '0.5200', which appears to have an extra digit; it should likely be 0.520.
  3. [Sec. 4.5, Eq. (10)] The sentence following Eq. (10) states 'where ϵ is the guidance weight,' but the equation uses ω as the guidance weight; this should be corrected to avoid confusion with the noise variable in Eq. (1).
  4. [Sec. A2, inference] The sentence 'The guidance weights is set to 2.0 for the SID [5] and LRD datasets [5], and 2.5 for ELD datasets [49], respectively' has a subject-verb agreement error and should be rephrased as 'The guidance weights are set to ...'.
  5. [Sec. 5, datasets] The paper states that the same ISP process (rawpy) is used for final evaluation of all methods and ground truth, but it does not specify the exact rawpy settings (e.g., gamma, color correction, white balance parameters). Please provide these details to ensure the comparison is reproducible and fair.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: DarkDiff’s LPIPS results are independent held-out measurements on external benchmarks, not by-construction consequences of fitted parameters.

full rationale

DarkDiff is an empirical pipeline: noisy linear RGB images are encoded, denoised by a fine-tuned Stable Diffusion U-Net with region-based cross-attention, and decoded to sRGB. The claimed predictions (LPIPS numbers in Tables 1-3) are evaluated on held-out test splits of SID, ELD, and LRD, with all baselines processed through the same rawpy ISP pipeline for fairness. No equation in the paper defines the reported LPIPS in terms of fitted constants or training losses; the stage-1 VAE uses the perceptual loss of [56] as one training term, and the guidance weights (2.0/2.0/2.5) are inference-time operating points, but neither is equivalent to the test-set LPIPS values by construction. The only self-citation (SID [5] by co-author Chen Chen) appears as dataset and baseline, not as a load-bearing justification for the central claim. The concern that LPIPS may not fully capture human perceptual preference, or that guidance weights might have been tuned on the test set, is an empirical validity risk rather than a circularity, and would not constitute circularity even if confirmed.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim depends on five assumptions: the LRGB pipeline bridges the domain gap, LPIPS equals perception, fabricated detail is acceptable, the DDPM one-step estimate is valid for pixel-space supervision, and the pretrained prior survives fine-tuning. Free parameters are the per-dataset guidance weight, the unreported decoder-loss weight lambda, and the unreported region-grouping geometry. No invented entities. The heaviest load is the LPIPS-as-perception assumption, because on all three datasets PSNR and SSIM favor baselines.

free parameters (3)
  • guidance weight omega (classifier-free) = 2.0 (SID, LRD), 2.5 (ELD)
    Inference-time knob balancing generated detail against smoothing (Fig. 8). Set per dataset with no reported validation split; LPIPS and PSNR results are sensitive to it.
  • decoder-loss weight lambda = not reported
    In L = L_LDM + lambda * L_image (Sec. 4.4); without a value the exact training objective is unreproducible and its influence on the color-shift results is unknown.
  • region size and token grouping for region-based cross-attention = not reported
    The grouping of the latent into K_l non-overlapping regions with N' tokens (Sec. 4.2) determines locality and cost; no value or ablation is given.
assumptions (5)
  • domain assumption Brightened linear-RGB preprocessing (black-level subtract, digital gain, white balance, simple demosaic) is sufficient to bridge the raw-to-sRGB domain gap so that Stable Diffusion's prior transfers.
    Core premise of the Sec. 4.1 pipeline; validated only indirectly by the LRGB-sRGB row of Table 5 outperforming Raw-sRGB and Raw-Raw formats.
  • domain assumption LPIPS is a faithful proxy for human perceptual quality in extreme-low-light generation.
    The entire 'SOTA perceptual quality' claim depends on this; the paper cites the perception-distortion tradeoff ([2,13,14]) instead of running a human study or a second perceptual metric (Sec. 5).
  • domain assumption Generating plausible but possibly fabricated detail in extremely dark regions is the correct objective for enhancement.
    The qualitative results celebrate reconstructed leaves, book patterns, and rails that are below the noise floor (Sec. 5.1, Fig. 10-12); the paper accepts hallucination risk (Fig. 8) but does not quantify how often generated details match ground truth.
  • standard math The DDPM one-step denoising approximation z0_hat = (zt - sqrt(1 - alpha_bar_t) * eps_hat) / sqrt(alpha_bar_t) is valid for the pixel-space loss.
    Eq. (8) is the standard posterior-mean estimate from Ho et al. [22] and is used to define the decoder reconstruction loss (Sec. 4.4).
  • domain assumption After fine-tuning on 280 paired SID images, the pre-trained Stable Diffusion U-Net retains enough generative prior to outperform training from scratch.
    Design premise of 'retasking' (Sec. 1, 4); supported by the failure of naive conditioning (SD Concat PSNR 17.33 in Table 1) but the contribution of the prior vs the fine-tuning is not isolated experimentally.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DarkDiff: Advancing Low-Light Raw Enhancement by Retasking Diffusion Models for Camera ISP." pith.science (2026). https://pith.science/paper/DKDXMDHB

@misc{pith2026250523743,
  author       = {Pith},
  title        = {Pith review of: DarkDiff: Advancing Low-Light Raw Enhancement by Retasking Diffusion Models for Camera ISP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKDXMDHB}},
  note         = {Machine review of arXiv:2505.23743}
}
read the original abstract

High-quality photography in extreme low-light conditions is challenging but impactful for digital cameras. With advanced computing hardware, traditional camera image signal processor (ISP) algorithms are gradually being replaced by efficient deep networks that enhance noisy raw images more intelligently. However, existing regression-based models often minimize pixel errors and result in oversmoothing of low-light photos or deep shadows. Recent work has attempted to address this limitation by training a diffusion model from scratch, yet those models still struggle to recover sharp image details and accurate colors. We introduce a novel framework to enhance low-light raw images by retasking pre-trained generative diffusion models with the camera ISP. Extensive experiments demonstrate that our method outperforms the state-of-the-art in perceptual quality across three challenging low-light raw image benchmarks.

Figures

Figures reproduced from arXiv: 2505.23743 by the authors.

Figure 1
Figure 1. Comparisons of low-light raw image enhancement results. The two input raw images were [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed DarkDiff pipeline. The noisy linear RBG (LRGB) image y is processed by the encoder E to generate latent representations zy. These representations and the Gaussian noise are fed into the Denoising U-Net, which integrates a region-based cross-attention between the noisy image and context from the pre-trained model to refine the latent variables zt. The decoder D reconstructs the final clean SR… view at source ↗
Figure 4
Figure 4. VAE reconstruction results with and without our proposed residual architecture. We observe a loss in input details when not using a residual connection. Given the latent features (tokens) z l t ∈ R N×d at the l-th layer of U-Net and the condition c ∈ RM×d , a cross-attention computes the queries, keys, and values as follows: Ql = z l t (Wl Q) ⊤, Kl = c(Wl K) ⊤, Vl = c(Wl V ) ⊤, (3) where Wl Q,Wl K,Wl V ∈ R d×d l k a… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Our data processing pipeline converts Bayer raw input into a linear RGB format by applying Attention [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Region-based cross-attention. 128 256 512 512 512 512 512 256 Conv layer Embedding [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Results of classifier-free guidance at different strengths. A small amount of guidance smooths patterns; strong guidance adds details. to take in an additional input, i.e., ϵθ(zt, c, t). The adjusted noise prediction ϵˆ is controlled by linearly combining the condition…
Figure 9
Figure 9. Figure 9: LPIPS better reflects perceptual quality [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Low-light image denoising results on SID [ [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Low-light image denoising results on ELD [ [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: Low-light image denoising results on LRD [ [PITH_FULL_IMAGE:figures/full_fig_p009_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 48 canonical work pages

  1. [50]

    LDM-ISP: Enhancing neural isp for low light with latent diffusion models.arXiv preprint arXiv:2312.01027, 2023

    Qiang Wen, Yazhou Xing, Zhefan Rao, and Qifeng Chen. LDM-ISP: Enhancing neural isp for low light with latent diffusion models.arXiv preprint arXiv:2312.01027, 2023. 3

    arXiv 2023

  2. [1]

    Noise flow: Noise modeling with conditional normalizing flows

    Abdelrahman Abdelhamed, Marcus A Brubaker, and Michael S Brown. Noise flow: Noise modeling with conditional normalizing flows. InProc. ICCV, 2019. 2

    work page 2019

  3. [2]

    The perception-distortion tradeoff

    Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. InProc. CVPR, 2018. 7, 8

    work page 2018

  4. [3]

    Unpro- cessing images for learned raw denoising

    Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unpro- cessing images for learned raw denoising. InProc. CVPR, 2019. 2

    work page 2019

  5. [4]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. InProc. CVPR, 2023. 8, 14

    work page 2023

  6. [5]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. InProc. CVPR, 2018. 1, 2, 3, 5, 6, 7, 8, 9, 14, 15, 16, 17

    work page 2018

  7. [6]

    CrossViT: Cross-attention multi-scale vision transformer for image classification

    Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. CrossViT: Cross-attention multi-scale vision transformer for image classification. InProc. CVPR, 2021. 3

    work page 2021

  8. [7]

    Training-free layout control with cross-attention guidance

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. InProc. WACV, 2024. 3

    work page 2024

Show all 56 references
  1. [8]

    DiffEdit: Diffusion-based semantic image editing with mask guidance

    Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. DiffEdit: Diffusion-based semantic image editing with mask guidance. InProc. ICLR, 2023. 7

    2023

  2. [9]

    PPformer: Using pixel-wise and patch-wise cross-attention for low-light image enhancement.Computer Vision and Image Understanding, 241:103930, 2024

    Jiachen Dang, Yong Zhong, and Xiaolin Qin. PPformer: Using pixel-wise and patch-wise cross-attention for low-light image enhancement.Computer Vision and Image Understanding, 241:103930, 2024. 3

    2024

  3. [10]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InProc. CVPR, 2021. 6

    2021

  4. [11]

    Generative diffusion prior for unified image restoration and enhancement

    Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. InProc. CVPR, 2023. 3

    2023

  5. [12]

    Implicit diffusion models for continuous super-resolution

    Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. InProc. CVPR,

  6. [13]

    Pipal: a large-scale image quality assessment dataset for perceptual image restoration

    Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. InProc. ECCV, 2020. 7, 8

    2020

  7. [14]

    Ren, Radu Timofte, Yuan Gong, Shanshan Lao, and et al Shi, Shuwei

    Jinjin Gu, Haoming Cai, Chao Dong, Jimmy S. Ren, Radu Timofte, Yuan Gong, Shanshan Lao, and et al Shi, Shuwei. NTIRE 2022 challenge on perceptual image quality assessment. InProc. CVPR Workshops,

    2022

  8. [15]

    Low-light image enhancement via breaking down the darkness.IJCV, 2023

    Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness.IJCV, 2023. 2

    2023

  9. [16]

    Lime: Low-light image enhancement via illumination map estimation

    Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE TIP, 2016. 2

    2016

  10. [17]

    Burst photography for high dynamic range and low-light imaging on mobile cameras.ACM TOG, 2016

    Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras.ACM TOG, 2016. 1

    2016

  11. [18]

    Reti-Diff: Illumination degradation image restoration with retinex-based latent diffusion model.arXiv preprint arXiv:2311.11638, 2023

    Chunming He, Chengyu Fang, Yulun Zhang, Tian Ye, Kai Li, Longxiang Tang, Zhenhua Guo, Xiu Li, and Sina Farsiu. Reti-Diff: Illumination degradation image restoration with retinex-based latent diffusion model.arXiv preprint arXiv:2311.11638, 2023. 3

    2023 arXiv

  12. [19]

    Diffusion models in low-level vision: A survey.arXiv preprint arXiv:2406.11138,

    Chunming He, Yuqi Shen, Chengyu Fang, Fengyang Xiao, Longxiang Tang, Yulun Zhang, Wangmeng Zuo, Zhenhua Guo, and Xiu Li. Diffusion models in low-level vision: A survey.arXiv preprint arXiv:2406.11138,

    arXiv

  13. [20]

    Prompt-to- prompt image editing with cross-attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to- prompt image editing with cross-attention control. InProc. ICLR, 2023. 3

    2023

  14. [21]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 7

    2022 arXiv

  15. [22]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 3, 6 10

    2020

  16. [23]

    Global structure-aware diffusion process for low-light image enhancement

    Jinhui Hou, Zhiyu Zhu, Junhui Hou, Hui Liu, Huanqiang Zeng, and Hui Yuan. Global structure-aware diffusion process for low-light image enhancement. InProc. CVPR, 2024. 3

    2024

  17. [24]

    Low-light image enhancement with wavelet-based diffusion models.ACM TOG, 2023

    Hai Jiang, Ao Luo, Haoqiang Fan, Songchen Han, and Shuaicheng Liu. Low-light image enhancement with wavelet-based diffusion models.ACM TOG, 2023

    2023

  18. [25]

    Lightendiffusion: Unsupervised low-light image enhancement with latent-retinex diffusion models

    Hai Jiang, Ao Luo, Xiaohong Liu, Songchen Han, and Shuaicheng Liu. Lightendiffusion: Unsupervised low-light image enhancement with latent-retinex diffusion models. InProc. ECCV, 2024. 3

    2024

  19. [26]

    Dnf: Decouple and feedback network for seeing in the dark

    Xin Jin, Linghao Han, Zhen Li, Zhi Chai, Chunle Guo, and Chongyi Li. Dnf: Decouple and feedback network for seeing in the dark. InProc. CVPR, 2023. 3

    2023

  20. [27]

    Lighting every darkness in two pairs: A calibration-free pipeline for raw denoising

    Xin Jin, Jia-Wen Xiao, Ling-Hao Han, Chunle Guo, Ruixun Zhang, Xialei Liu, and Chongyi Li. Lighting every darkness in two pairs: A calibration-free pipeline for raw denoising. InProc. ICCV, 2023. 2, 7, 15, 16, 17

    2023

  21. [28]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. InProc. CVPR, 2020. 7

    2020

  22. [29]

    Denoising diffusion restoration models

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. InProc. NeurIPS, 2022. 3

    2022

  23. [30]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InProc. CVPR, 2023. 7

    2023

  24. [31]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. 13, 14

    2014 arXiv

  25. [32]

    xFormers: A modular and hackable transformer modelling library

    Benjamin Lefaudeux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xFormers: A modular and hackable transformer ...

    2022

  26. [33]

    Handheld mobile photography in very low light.ACM TOG,

    Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light.ACM TOG,

  27. [34]

    DiffBIR: Towards blind image restoration with generative diffusion prior.arXiv preprint arXiv:2308.15070, 2023

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. DiffBIR: Towards blind image restoration with generative diffusion prior.arXiv preprint arXiv:2308.15070, 2023. 3, 14

    2023 arXiv

  28. [35]

    Llnet: A deep autoencoder approach to natural low-light image enhancement.Pattern Recognition, 2017

    Kin Gwn Lore, Adedotun Akintayo, and Soumik Sarkar. Llnet: A deep autoencoder approach to natural low-light image enhancement.Pattern Recognition, 2017. 2

    2017

  29. [36]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InProc. CVPR, 2022. 3

    2022

  30. [37]

    Diffusion in the dark: A diffusion model for low-light text recognition

    Cindy M Nguyen, Eric R Chan, Alexander W Bergman, and Gordon Wetzstein. Diffusion in the dark: A diffusion model for low-light text recognition. InProc. WACV, 2024. 3

    2024

  31. [38]

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models.IEEE TPAMI, 2023

    Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models.IEEE TPAMI, 2023. 3

    2023

  32. [39]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProc. CVPR, 2022. 2, 3, 6

    2022

  33. [40]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InProc. MICCAI, 2015. 3, 9

    2015

  34. [41]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InProc. ICML, 2015. 3

    2015

  35. [42]

    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. 14

    2010 arXiv

  36. [43]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Proc. NeurIPS, 2019. 3

    2019

  37. [44]

    Improved techniques for training score-based generative models

    Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. InProc. NeurIPS, 2020. 3 11

    2020

  38. [45]

    Zero-shot image restoration using denoising diffusion null-space model.arXiv preprint arXiv:2212.00490, 2022

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model.arXiv preprint arXiv:2212.00490, 2022. 3

    2022 arXiv

  39. [46]

    Low-light image enhance- ment with illumination-aware gamma correction and complete image modelling network

    Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhance- ment with illumination-aware gamma correction and complete image modelling network. InProc. ICCV,

  40. [47]

    ExposureD- iffusion: Learning to expose for low-light image enhancement

    Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. ExposureD- iffusion: Learning to expose for low-light image enhancement. InProc. ICCV, 2023. 1, 2, 3, 8, 15, 16, 17

    2023

  41. [48]

    Restoreformer: High-quality blind face restoration from undegraded key-value pairs

    Zhouxia Wang, Jiawei Zhang, Runjian Chen, Wenping Wang, and Ping Luo. Restoreformer: High-quality blind face restoration from undegraded key-value pairs. InProc. CVPR, 2022. 3

    2022

  42. [49]

    A physics-based noise formation model for extreme low-light raw denoising

    Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. InProc. CVPR, 2020. 2, 3, 6, 7, 8, 14, 15, 16, 17

    2020

  43. [51]

    Deblurring via stochastic refinement

    Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. InProc. CVPR, 2022. 3

    2022

  44. [52]

    Low-light image enhancement via structure modeling and guidance

    Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. InProc. CVPR, 2023. 2

    2023

  45. [53]

    Diff-Retinex: Rethinking low-light image enhancement with a generative diffusion model

    Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-Retinex: Rethinking low-light image enhancement with a generative diffusion model. InProc. ICCV, 2023. 2, 3

    2023

  46. [54]

    Towards general low-light raw noise synthesis and modeling

    Feng Zhang, Bin Xu, Zhiqiang Li, Xinran Liu, Qingbo Lu, Changxin Gao, and Nong Sang. Towards general low-light raw noise synthesis and modeling. InProc. ICCV, 2023. 2, 3, 6, 7, 8, 9, 14, 15, 16, 17

    2023

  47. [55]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InProc. ICCV, 2023. 8, 14

    2023

  48. [56]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProc. CVPR, 2018. 2, 6, 7 12 Appendix The appendix is organized as follows: • In Sec. A1, we provide additional qualitative r...

    2018

Pith tools

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