Pith. sign in

REVIEW 3 major objections 6 minor 46 references

A frozen diffusion model can remove shadows without training by transferring light from lit regions and selectively keeping only the content features that shadows barely change.

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 ↗

A frozen Stable Diffusion model can remove shadows at test time by reweighting self-attention for illumination transfer and selectively reinjecting shadow-insensitive structure and high-frequency detail.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Solid training-free shadow remover with real OOD numbers; the SI layer picks are a mild soft spot, not a collapse of the claim. the 3 major comments →

arxiv 2607.26715 v1 pith:CHS4PVKD submitted 2026-07-29 cs.CV

FreeShadow: Training-Free Shadow Removal via Illumination Transfer and Selective Content Preservation in Diffusion Models

Yinan Wang , Yan Huang , Yong Xu , Patrick Le Callet This is my paper
classification cs.CV
keywords shadow removaldiffusion modelstraining-freeillumination transfer attentionselective content preservationlocal texture-preserving relightingStable Diffusion
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 reading

Shadow removal has long been limited by scarce paired data: supervised and unsupervised networks overfit the few available training sets, and zero-shot optimizers are slow and artifact-prone. FreeShadow shows that a pretrained text-to-image diffusion model already contains enough natural-image knowledge to do the job with no fine-tuning and no test-time training. During denoising it re-weights self-attention so shadow tokens borrow illumination from non-shadow tokens, then reinjects only those attention maps and latent high-frequency bands that an analysis shows are least distorted by shadows, and finally stitches local texture statistics back from the original photograph to undo VAE compression. The result is realistic, detail-preserving shadow-free images that generalize across outdoor scenes and portraits better than methods trained on the same limited data.

Core claim

A completely training-free pipeline on frozen Stable Diffusion—illumination-transfer attention plus selective reinjection of the least shadow-sensitive self-attention maps and latent high-frequency components, followed by local texture-preserving relighting—restores illumination while preserving structure and fine detail, and yields stronger cross-dataset generalization than supervised, unsupervised, and zero-shot baselines trained on ISTD+.

What carries the argument

Illumination Transfer Attention (ITA): a mask-guided re-weighting of self-attention that amplifies attention from shadow queries to non-shadow keys, so lighting cues flow into shadowed regions; paired with Selective Attention Map and High-Frequency Injection that keep only the decoder layers and wavelet bands least altered by shadows.

Load-bearing premise

Non-shadow areas supply usable lighting references, the particular attention layers and high-frequency bands judged least shadow-sensitive stay the right content carriers on new images, and a usable shadow mask (plus its boundary band) is available at test time.

What would settle it

Run the identical frozen pipeline, without any hyper-parameter change, on a large held-out collection of hard cast and self-shadows whose masks are either noisy or replaced by loose bounding boxes; if residual shadows, boundary artifacts, or large drops in MAE/SSIM appear relative to the reported UIUC+/SRD/PSM numbers, the claim fails.

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

If this is right

  • Shadow removal no longer requires collecting paired shadow/shadow-free training sets for every new domain.
  • The same selective-injection recipe can be dropped onto other frozen diffusion backbones (e.g., SDXL) with only mask and layer choices adjusted.
  • Cross-dataset scores on UIUC+, SRD and the portrait set PSM become the practical benchmark for generalization rather than in-distribution ISTD+ numbers.
  • Runtime can be traded for quality simply by reducing DDIM steps (e.g., 20 steps already competitive) without retraining.

Where Pith is reading between the lines

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

  • The same SI analysis that selects illumination-invariant attention maps could guide training-free removal of other spatially localized degradations such as specular highlights or localized haze.
  • Because the method already works from coarse bounding-box masks when SHFI and LTPR are disabled, it could be paired with an off-the-shelf shadow detector for fully automatic pipelines.
  • Self-shadows remain harder precisely because geometry and appearance are entangled; an explicit surface-normal or depth cue injected into the same attention path is a natural next test.
Share X Bluesky LinkedIn Reddit HN

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. FreeShadow proposes a fully training-free shadow removal pipeline on frozen Stable Diffusion 2.1. After DDIM inversion of a masked shadow image, sampling is steered by (i) illumination transfer attention (ITA), which re-weights self-attention so shadow queries attend more to non-shadow keys (Eqs. 1–2), (ii) selective reinjection of decoder self-attention maps and latent high-frequency bands judged least shadow-sensitive (SAMI/SHFI; SI metric Eq. 4; Figs. 3–5), (iii) local texture-preserving relighting (LTPR) that transfers local mean/variance from the SD output onto the input texture (Eqs. 8–9), and (iv) shadow-removal redirect guidance (SRRG, Eq. 10). Non-shadow latents are locked to the inversion trajectory (Eq. 3). On ISTD+, UIUC+, SRD, and PSM, with supervised/unsupervised baselines trained only on ISTD+, the method reports competitive or superior MAE/SSIM/PSNR, especially under cross-dataset and portrait evaluation (Tables I–II), with component ablations (Tables IV–V) and qualitative comparisons.

Significance. If the empirical claims hold, the work is a clear contribution to training-free image restoration with large diffusion priors. It shows that a frozen T2I backbone, with carefully designed attention re-weighting and selective feature reinjection rather than fine-tuning or test-time optimization, can outperform both dataset-trained shadow removers under distribution shift and existing zero-shot methods that are slow or artifact-prone. Strengths include a consistent cross-dataset protocol, multi-component ablations, fixed hyperparameters across four datasets, mask-robustness checks (Fig. 13), guidance-scale and α sweeps, and an explicit failure-mode discussion (self-shadows). The SI analysis and selective-injection design are concrete and falsifiable engineering contributions, not only prompt engineering.

major comments (3)
  1. [§III.C–D, Eq. (4), Figs. 4–5; Tables I–II] §III.C–D, Eq. (4), Figs. 4–5: The SAMI layer policy (all three 16×16 maps plus the first map at 32×32 and 64×64) and the SHFI boundary-only keep-out are justified by SI boxplots and HF error maps computed on ISTD+ and SRD, then applied unchanged to UIUC+ and PSM (Tables I–II). The central OOD-generalization claim therefore partly rests on the premise that these illumination-invariant carriers transfer. Please either (a) report SI / HF-discrepancy statistics on UIUC+ and PSM under the same protocol, or (b) ablate alternative fixed layer sets on the held-out sets and show that the reported margins are not driven by this ISTD+/SRD-informed choice. Without that, the “pure prior / training-free” framing overstates independence from shadow-data statistics.
  2. [Tables I–II; Table III] Tables I–II and implementation: All quantitative scores are single-run point estimates with no error bars, multi-seed variation, or sensitivity to DDIM stochasticity / mask jitter beyond the qualitative bounding-box check in Fig. 13. Given that margins over strong SL baselines on UIUC+/SRD/PSM are a primary claim, at least seed- or step-schedule variation (or bootstrap over images) on one OOD set is needed to establish that the ranking is stable. Table III already varies T; extending that discipline to the main metrics would substantially strengthen the result.
  3. [§III.B–E; Fig. 13] Eqs. (1)–(3), (5)–(9) and Fig. 13: The pipeline assumes a usable shadow mask and a morphology-derived boundary band (kernel 19). Fig. 13 shows that coarse boxes still work if SHFI/LTPR are dropped, but the full model can produce boundary artifacts under mask error, and no automatic masker is evaluated end-to-end. For a method positioned as practical and training-free, please quantify full-pipeline performance with an off-the-shelf shadow detector (or report degraded-mask MAE on ISTD+/SRD) and state clearly in the abstract/intro that a mask is required, with the ITA+SAMI-only fallback as the recommended operating mode under unreliable masks.
minor comments (6)
  1. [Table I] Table I header colors supervised/unsupervised/zero-shot winners, but FreeShadow is listed under ZS-style comparison without a distinct TF highlight; a fourth category or explicit “training-free” row group would avoid implying it is optimized like Self-SGAN.
  2. [Eq. (4)] Eq. (4): SI normalizes by the global mean of the PCA map; a brief note on stability under sign flips of PCA components (or use of absolute PCA loadings) would help reproducibility.
  3. [Fig. 5; §III.D] Fig. 5 shows HF error maps at t=T, 0.5T, 0 but does not state whether the same interior-vs-boundary split holds across timesteps used in sampling; one sentence would clarify why a single boundary mask is applied at every t.
  4. [§II.B] Related work: BCDiff and StableShadow are discussed; a short contrast with other training-free diffusion editing controls (e.g., PnP / FreeControl-style attention injection) would situate ITA/SAMI more clearly for the diffusion-editing audience.
  5. [throughout] Typos/notation: “V AE” spacing is inconsistent (VAE); “timestept” missing space in §III.B; arXiv IDs in references for concurrent editing models are fine but venue years should be double-checked before camera-ready.
  6. [Fig. 12] Fig. 12 guidance-scale curve: y-axis metric is not labeled in the caption; state whether it is MAE on ISTD+.

Circularity Check

0 steps flagged

No significant circularity: external GT benchmarks and design choices that do not algebraically force the reported metrics.

full rationale

FreeShadow’s load-bearing claims are empirical shadow-removal quality (MAE/SSIM/PSNR) against held-out paired ground-truth images on ISTD+, UIUC+, SRD, and PSM, plus cross-dataset generalization when SL/UL baselines are trained only on ISTD+. The method is a fixed, training-free recipe on frozen SD 2.1: ITA re-weights attention (Eqs. 1–3), SAMI/SHFI selectively reinject decoder self-attention maps and non-boundary latent high-frequency bands (Eqs. 4–7, Figs. 3–5), LTPR matches local mean/variance (Eqs. 8–9), and SRRG steers sampling (Eq. 10). Layer/band selection is justified by an SI statistic and error maps computed on ISTD+/SRD attention and latents; that is ordinary empirical design/ablation (Tables IV–V), not a fitted parameter renamed as a prediction of a quantity that equals the fit by construction. Reported scores are not algebraically identical to SI, α, s, or the mask. No self-citation uniqueness theorem, no ansatz smuggled in as external fact, and no renaming of a known closed-form result. Mild dependence of design on related data is a generalization/robustness concern, not circularity. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 3 invented entities

The claim rests on pretrained diffusion priors, mask-conditioned attention surgery, and an empirical layer-selection rule—not on new physics. Load-bearing inputs are the frozen SD prior, availability of a shadow mask, the hand-chosen scalars (α, s, morphology, LTPR window), and the assumption that SI-minimizing maps generalize as content carriers. No new particles or forces; invented pieces are methodological modules and the SI statistic.

free parameters (6)
  • ITA re-weighting coefficient α = 1.5
    Controls how strongly shadow queries attend to non-shadow keys (Eq. 1–2). Set to 1.5 after a sensitivity sweep (Fig. 14); performance depends on this choice in [1,3].
  • Shadow removal guidance scale s (SRRG) = 7
    Amplifies the ITA+SAMI noise direction versus the plain denoiser (Eq. 10). Empirically set to 7 after a sweep (Fig. 12).
  • DDIM steps T = 50 (main)
    Inversion/sampling length; main results use 50, with a runtime–quality table for 10–50. Affects both metrics and latency.
  • Shadow-boundary morphology kernel = 19
    Mb = dilation(Ms)−erosion(Ms) with kernel size 19; defines where SHFI/LTPR trust the generative model vs the input.
  • LTPR window size k and stride d = k=7, d=1
    Local mean/std matching patch size (Eq. 8); set to k=7, d=1 without a full public sensitivity study in the main text.
  • SAMI layer-selection policy = 16×16:{1,2,3}; 32×16 first; 64×64 first
    Which decoder attentions to inject is chosen from SI boxplots on ISTD+/SRD (all 16×16 maps; first map only at 32 and 64). Discrete design choice fitted to observed SI, then frozen for all datasets.
axioms (7)
  • domain assumption Large-scale pretrained T2I diffusion models encode natural-image priors sufficient to synthesize plausible shadow-free appearance when sampling is suitably steered.
    Stated throughout §I and §III.A; the entire training-free premise depends on it.
  • domain assumption Non-shadow regions of the same image provide valid illumination references for shadow regions via self-attention reweighting.
    Motivation for ITA in §III.B; fails for complex multi-light or self-shadow geometry (limitation in §IV.D.3).
  • domain assumption A binary (or coarse) shadow mask Ms is available at test time and can be downsampled by max-pooling to latent resolution.
    Used in Eqs. 1–3 and throughout; robustness to boxes is shown but mask-free operation is not claimed.
  • ad hoc to paper Lower-resolution decoder self-attention and first-layer medium/high-resolution maps are sufficiently illumination-invariant to safely reinject for structure preservation.
    Derived from SI analysis in §III.C / Fig. 4 and ablated in Table V; not a standard theorem.
  • ad hoc to paper High-frequency latent discrepancy between shadow and shadow-free images concentrates on shadow boundaries, so interior HF from the inversion is trustworthy.
    Justified by Fig. 5 and operationalized in Eqs. 5–7 (SHFI).
  • domain assumption DDIM inversion is approximately reversible enough that the unguided sampling direction reconstructs the shadow input, so steering away from it yields deshadowing.
    Explicit in §III.F explaining SRRG.
  • standard math Standard linear algebra / softmax attention and DWT/IDWT identities hold as used in Eqs. 2 and 5–7.
    Routine operators; no novel math claimed.
invented entities (3)
  • Shadow influence (SI) metric on PCA-reduced self-attention maps no independent evidence
    purpose: Quantify per-layer shadow contamination to decide which attention maps to reinject (Eq. 4).
    Paper-defined statistic; not an external physical quantity. Guides SAMI design; independent_evidence is only the internal boxplots and ablations.
  • Illumination transfer attention (ITA) no independent evidence
    purpose: Mask-conditioned additive reweighting of QKᵀ so shadow queries favor non-shadow keys.
    Named mechanism; engineering construct rather than a postulated natural entity. Validated only via this paper’s ablations.
  • Local texture-preserving relighting (LTPR) no independent evidence
    purpose: Transfer local mean/variance from SD output onto input patches to undo VAE texture drift (Eqs. 8–9).
    Method module combining known local statistics matching with a boundary exception; no external falsifiable handle beyond reported metrics.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FreeShadow: Training-Free Shadow Removal via Illumination Transfer and Selective Content Preservation in Diffusion Models." pith.science (2026). https://pith.science/paper/CHS4PVKD

@misc{pith2026260726715,
  author       = {Pith},
  title        = {Pith review of: FreeShadow: Training-Free Shadow Removal via Illumination Transfer and Selective Content Preservation in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CHS4PVKD}},
  note         = {Machine review of arXiv:2607.26715}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing supervised and unsupervised shadow removal methods often suffer from limited generalization due to the insufficient diversity of available training datasets, while zero-shot methods tend to produce artifacts and require time-consuming test-time optimization. To address these issues, we propose FreeShadow, a training-free shadow removal method built upon pretrained diffusion models, which exploits diffusion priors for shadow removal without any training or optimization. For illumination recovery, we propose an illumination transfer attention (ITA), which re-weights the self-attention maps in diffusion model to transfer illumination cues from non-shadow to shadow regions. For content preservation, we analyze the effects of illumination variations on self-attention maps and latent high-frequency features in diffusion model, and selectively preserve illumination-invariant components to maintain content fidelity while suppressing residual shadows. We further propose local texture-preserving relighting (LTPR) to mitigate local texture misalignment caused by VAE compression. Extensive experiments demonstrate that our method achieves strong generalization and produces realistic shadow-free images.

Figures

Figures reproduced from arXiv: 2607.26715 by Patrick Le Callet, Yan Huang, Yinan Wang, Yong Xu.

Figure 1
Figure 1. Figure 1: Comparison of shadow removal results on out-of-distribution images. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of FreeShadow. The shadow image is first inverted into a latent sequence via DDIM inversion. During DDIM sampling, ITA [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of PCA-reduced self-attention maps across different layers in SD2.1 U-net decoder. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Boxplot of shadow influence (SI) on SD2.1 self-attention maps of shadow images in the ISTD+ and SRD datasets. Results on more datasets and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Error map of the high-frequency components in DDIM inversion latent [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparisons with state-of-the-art methods on the ISTD+ dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparisons with state-of-the-art methods on the UIUC+ dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual comparisons with state-of-the-art methods on the SRD dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparisons with state-of-the-art methods on the PSM dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual effect of different components. Input (c) (e) (g) Full [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual effect of the selective injection strategy. (c), (e), and (g) [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Shadow removal performance under inaccurate shadow masks. [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Failure cases of our method. transitions along shadow boundaries often degrade local details in these regions. Since SHFI and LTPR rely on the shadow boundary mask to restore fine details in degraded boundary regions, inaccurate shadow masks may cause the full model to produce boundary artifacts. By retaining only ITA and SAMI (excluding SHFI and LTPR), these artifacts can be effectively suppressed, provi… 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]

    Generative portrait shadow removal,

    J. S. Yoon, Z. Shu, M. Ren, C. Zhang, Y . Hold-Geoffroy, K. K. Singh, and H. Zhang, “Generative portrait shadow removal,”ACM TOG, vol. 43, no. 6, pp. 1–13, 2024

    2024

  2. [2]

    Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon,

    Y . Zhong, X. Liu, D. Zhai, J. Jiang, and X. Ji, “Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon,” inCVPR, 2022, pp. 15 345–15 354

    2022

  3. [3]

    Unsupervised moving object detection in complex scenes using adversarial regularizations,

    M. Sultana, A. Mahmood, and S. K. Jung, “Unsupervised moving object detection in complex scenes using adversarial regularizations,”IEEE TMM, vol. 23, pp. 2005–2018, 2020

    2005

  4. [4]

    Hyper rpca: joint maximum correntropy criterion and laplacian scale mixture modeling on-the-fly for moving object detection,

    Z. Shao, Y . Pu, J. Zhou, B. Wen, and Y . Zhang, “Hyper rpca: joint maximum correntropy criterion and laplacian scale mixture modeling on-the-fly for moving object detection,”IEEE TMM, vol. 25, pp. 112– 125, 2021

    2021

  5. [5]

    Decoupled spatial neural attention for weakly supervised semantic segmentation,

    T. Zhang, G. Lin, J. Cai, T. Shen, C. Shen, and A. C. Kot, “Decoupled spatial neural attention for weakly supervised semantic segmentation,” IEEE TMM, vol. 21, no. 11, pp. 2930–2941, 2019

    2019

  6. [6]

    Consistency preservation and feature entropy regularization for gan based face editing,

    W. Xie, W. Lu, Z. Peng, and L. Shen, “Consistency preservation and feature entropy regularization for gan based face editing,”IEEE TMM, vol. 25, pp. 8892–8905, 2023

    2023

  7. [7]

    A boundary-aware network for shadow removal,

    K. Niu, Y . Liu, E. Wu, and G. Xing, “A boundary-aware network for shadow removal,”IEEE TMM, vol. 25, pp. 6782–6793, 2022

    2022

  8. [8]

    A decoupled multi- task network for shadow removal,

    J. Liu, Q. Wang, H. Fan, W. Li, L. Qu, and Y . Tang, “A decoupled multi- task network for shadow removal,”IEEE TMM, vol. 25, pp. 9449–9463, 2023

    2023

  9. [9]

    Shadowformer: Global context helps shadow removal,

    L. Guo, S. Huang, D. Liu, H. Cheng, and B. Wen, “Shadowformer: Global context helps shadow removal,”AAAI, vol. 37, no. 1, pp. 710– 718, 2023

    2023

  10. [10]

    Recasting regional lighting for shadow removal,

    Y . Liu, Z. Ke, K. Xu, F. Liu, Z. Wang, and R. W. Lau, “Recasting regional lighting for shadow removal,” inAAAI, vol. 38, no. 4, 2024, pp. 3810–3818

    2024

  11. [11]

    Omnisr: Shadow removal under direct and indirect lighting,

    J. Xu, Z. Li, Y . Zheng, C. Huang, R. Gu, W. Xu, and G. Xu, “Omnisr: Shadow removal under direct and indirect lighting,” inAAAI, vol. 39, no. 8, 2025, pp. 8887–8895

    2025

  12. [12]

    Shadow removal via shadow image decompo- sition,

    H. Le and D. Samaras, “Shadow removal via shadow image decompo- sition,” inICCV, 2019, pp. 8578–8587

    2019

  13. [13]

    Deshadownet: A multi- context embedding deep network for shadow removal,

    L. Qu, J. Tian, S. He, Y . Tang, and R. W. Lau, “Deshadownet: A multi- context embedding deep network for shadow removal,” inCVPR, 2017, pp. 4067–4075

    2017

  14. [14]

    Mask-shadowgan: Learning to remove shadows from unpaired data,

    X. Hu, Y . Jiang, C.-W. Fu, and P.-A. Heng, “Mask-shadowgan: Learning to remove shadows from unpaired data,” inICCV, 2019, pp. 2472–2481

    2019

  15. [15]

    Learning to remove shadows from a single image,

    H. Jiang, Q. Zhang, Y . Nie, L. Zhu, and W.-S. Zheng, “Learning to remove shadows from a single image,”IJCV, vol. 131, no. 9, pp. 2471– 2488, 2023

    2023

  16. [16]

    From shadow generation to shadow removal,

    Z. Liu, H. Yin, X. Wu, Z. Wu, Y . Mi, and S. Wang, “From shadow generation to shadow removal,” inCVPR, 2021, pp. 4927–4936

    2021

  17. [17]

    When shadow removal meets intrinsic image decomposition: A joint learning framework using unpaired data,

    R. Zheng, Q. Zhang, Y . Nie, and W.-S. Zheng, “When shadow removal meets intrinsic image decomposition: A joint learning framework using unpaired data,” inAAAI, vol. 39, no. 10, 2025, pp. 10 599–10 607

    2025

  18. [18]

    Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network,

    Y . Jin, A. Sharma, and R. T. Tan, “Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network,” inICCV, 2021, pp. 5027–5036

    2021

  19. [19]

    From shadow segmentation to shadow removal,

    H. Le and D. Samaras, “From shadow segmentation to shadow removal,” inECCV. Springer, 2020, pp. 264–281

    2020

  20. [20]

    Shadowdiffusion: When degradation prior meets diffusion model for shadow removal,

    L. Guo, C. Wang, W. Yang, S. Huang, Y . Wang, H. Pfister, and B. Wen, “Shadowdiffusion: When degradation prior meets diffusion model for shadow removal,” inCVPR, 2023, pp. 14 049–14 058

    2023

  21. [21]

    Detail- preserving latent diffusion for stable shadow removal,

    J. Xu, Y . Zheng, Z. Li, C. Wang, R. Gu, W. Xu, and G. Xu, “Detail- preserving latent diffusion for stable shadow removal,” inCVPR, 2025, pp. 7592–7602

    2025

  22. [22]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” NeurIPS, vol. 33, pp. 6840–6851, 2020

    2020

  23. [23]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inCVPR, 2022, pp. 10 684–10 695

    2022

  24. [24]

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

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

    2024

  25. [25]

    Zero-shot low- light image enhancement via latent diffusion models,

    Y . Huang, X. Liao, J. Liang, Y . Quan, B. Shi, and Y . Xu, “Zero-shot low- light image enhancement via latent diffusion models,” inAAAI, vol. 39, no. 4, 2025, pp. 3815–3823

    2025

  26. [26]

    Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal,

    L. Guo, C. Wang, W. Yang, Y . Wang, and B. Wen, “Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal,” inICCV, 2023, pp. 13 045–13 054

    2023

  27. [27]

    Homoformer: Homogenized transformer for image shadow removal,

    J. Xiao, X. Fu, Y . Zhu, D. Li, J. Huang, K. Zhu, and Z.-J. Zha, “Homoformer: Homogenized transformer for image shadow removal,” inCVPR, 2024, pp. 25 617–25 626

    2024

  28. [28]

    Icdsr: Integrated conditional diffusion model for single image super-resolution,

    C. Hu, X.-Z. Wei, and X.-J. Wu, “Icdsr: Integrated conditional diffusion model for single image super-resolution,”IEEE TMM, vol. 28, pp. 1302– 1313, 2026

    2026

  29. [29]

    Deblurring via stochastic refinement,

    J. Whang, M. Delbracio, H. Talebi, C. Saharia, A. G. Dimakis, and P. Milanfar, “Deblurring via stochastic refinement,” inCVPR, 2022, pp. 16 293–16 303

    2022

  30. [30]

    Mmginpainting: Multi- modality guided image inpainting based on diffusion models,

    C. Zhang, W. Yang, X. Li, and H. Han, “Mmginpainting: Multi- modality guided image inpainting based on diffusion models,”IEEE TMM, vol. 26, pp. 8811–8823, 2024

    2024

  31. [31]

    Latent feature-guided diffusion models for shadow removal,

    K. Mei, L. Figueroa, Z. Lin, Z. Ding, S. Cohen, and V . M. Patel, “Latent feature-guided diffusion models for shadow removal,” inWACV, 2024, pp. 4313–4322

    2024

  32. [32]

    Des3: Adaptive attention-driven self and soft shadow removal using vit similarity,

    Y . Jin, W. Ye, W. Yang, Y . Yuan, and R. T. Tan, “Des3: Adaptive attention-driven self and soft shadow removal using vit similarity,” in AAAI, vol. 38, no. 3, 2024, pp. 2634–2642

    2024

  33. [33]

    Diff-shadow: Global-guided diffusion model for shadow removal,

    J. Luo, R. Li, C. Jiang, X. Zhang, M. Han, T. Jiang, H. Fan, and S. Liu, “Diff-shadow: Global-guided diffusion model for shadow removal,” in AAAI, vol. 39, no. 6, 2025, pp. 5856–5864

    2025

  34. [34]

    Zero-shot text-to-image generation,

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. V oss, A. Radford, M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,” inICML. Pmlr, 2021, pp. 8821–8831

    2021

  35. [35]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inICLR, 2021

    2021

  36. [36]

    Plug-and-play diffusion features for text-driven image-to-image translation,

    N. Tumanyan, M. Geyer, S. Bagon, and T. Dekel, “Plug-and-play diffusion features for text-driven image-to-image translation,” inCVPR, 2023, pp. 1921–1930

    2023

  37. [37]

    Training-free style and content transfer by leveraging u-net skip connections in stable diffusion 2

    L. Schaerf, A. Alfarano, F. Silvestri, and L. Impett, “Training-free style and content transfer by leveraging u-net skip connections in stable diffusion 2.”arXiv preprint arXiv:2501.14524, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Stylessp: Sampling startpoint enhancement for training-free diffusion-based method for style transfer,

    R. Xu, W. Xi, X. Wang, Y . Mao, and Z. Cheng, “Stylessp: Sampling startpoint enhancement for training-free diffusion-based method for style transfer,” inCVPR, 2025, pp. 18 260–18 269

    2025

  39. [39]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis,

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, and R. Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,”arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Paired regions for shadow detection and removal,

    R. Guo, Q. Dai, and D. Hoiem, “Paired regions for shadow detection and removal,”IEEE TPAMI, vol. 35, no. 12, pp. 2956–2967, 2012

    2012

  41. [41]

    Portrait shadow manipulation,

    X. Zhang, J. T. Barron, Y .-T. Tsai, R. Pandey, X. Zhang, R. Ng, and D. E. Jacobs, “Portrait shadow manipulation,”ACM TOG, vol. 39, no. 4, pp. 78–1, 2020

    2020

  42. [42]

    Single-image shadow removal using deep learning: A comprehensive survey,

    L. Guo, C. Wang, Y . Wang, Y . Yu, S. Huang, W. Yang, A. C. Kot, and B. Wen, “Single-image shadow removal using deep learning: A comprehensive survey,”arXiv preprint arXiv:2407.08865, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Image shadow removal via multi-scale deep retinex decomposition,

    Y . Huang, X. Lu, Y . Quan, Y . Xu, and H. Ji, “Image shadow removal via multi-scale deep retinex decomposition,”PR, vol. 159, p. 111126, 2025

    2025

  44. [44]

    Shadow removal using bilateral filtering,

    Q. Yang, K.-H. Tan, and N. Ahuja, “Shadow removal using bilateral filtering,”IEEE TIP, vol. 21, no. 10, pp. 4361–4368, 2012

    2012

  45. [45]

    Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space,

    B. F. Labs, S. Batifol, A. Blattmann, F. Boesel, S. Consul, C. Diagne, T. Dockhorn, J. English, Z. English, P. Esseret al., “Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space,”arXiv preprint arXiv:2506.15742, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Qwen-image technical report,

    C. Wu, J. Li, J. Zhou, J. Lin, K. Gao, K. Yan, S.-m. Yin, S. Bai, X. Xu, Y . Chenet al., “Qwen-image technical report,”arXiv preprint arXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

This paper was first reviewed by grok-4.5 on July 30, 2026.