Pith. sign in

REVIEW 3 major objections 4 cited by

A training-free method speeds up Diffusion Transformers by focusing compute on aesthetic regions and still raises image quality.

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 · grok-4.5

2026-07-14 22:13 UTC pith:6FJCBJQA

load-bearing objection We only have the AccelAes abstract; the cached full text is a different paper (DINOLight), so the 2.11× / +11.9% ImageReward claims are unauditable. the 3 major comments →

arxiv 2603.12575 v2 pith:6FJCBJQA submitted 2026-03-13 cs.CV

AccelAes: Accelerating Diffusion Transformers for Training-Free Aesthetic-Enhanced Image Generation

Xuanhua Yin , Chuanzhi Xu , Haoxian Zhou , Boyu Wei , Weidong Cai This is my paper
classification cs.CV
keywords Diffusion Transformerstext-to-image generationtraining-free accelerationaesthetic focus maskspatio-temporal reductioncross-attentionImageReward
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.

Diffusion Transformers deliver high-fidelity text-to-image results but pay a heavy price for quadratic attention over every spatial token at every denoising step. The authors observe that this cost is uneven: prompt tokens that describe aesthetics attract concentrated cross-attention and produce large changes over time, while other regions evolve smoothly and waste computation. AccelAes turns that observation into a practical, training-free pipeline. It builds a one-shot aesthetic focus mask from prompt semantics and cross-attention, then reallocates self-attention and guidance to the masked regions and caches lightweight step-level predictions so full Transformer evaluations can be skipped. On representative DiT families the result is both faster inference and higher aesthetics-oriented scores, with a reported 2.11 imes speedup and +11.9 % ImageReward gain on Lumina-Next.

Core claim

Denoising inside Diffusion Transformers is spatially non-uniform with respect to aesthetic descriptors in the prompt; regions tied to those descriptors carry the quality-critical temporal variation, so a one-shot aesthetic mask plus sparse reallocation and a step-level cache can cut compute while improving perceptual aesthetics, all without any retraining.

What carries the argument

AesMask (a one-shot aesthetic focus mask derived from prompt semantics and cross-attention) together with SkipSparse (which reallocates computation and guidance onto the mask) and a lightweight step-level prediction cache that periodically replaces full Transformer passes.

Load-bearing premise

The method rests on the claim that aesthetic-related regions really do dominate quality-critical change while the remaining regions are safely compressible, so a single mask and sparse schedule will not harm (and can even raise) final aesthetics.

What would settle it

Run the identical AccelAes pipeline on a DiT family while deliberately randomizing or inverting the AesMask; if speedups remain but ImageReward and other aesthetics metrics fall below the dense baseline, the spatial-non-uniformity premise fails.

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

If this is right

  • High-resolution DiT serving can drop latency by roughly half without a second training stage or quality regression.
  • Aesthetic prompt tokens become an explicit computational budget that can be traded against background regions.
  • Step-level caching can be combined with other token-pruning schemes once the aesthetic mask is known.
  • The same non-uniformity observation may extend to other conditional generative backbones that mix cross-attention with dense self-attention.

Where Pith is reading between the lines

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

  • If aesthetic focus is stable early in denoising, the mask could be frozen after the first few steps and reused, further reducing overhead.
  • The same cross-attention signal might serve dual duty as both a quality prior and a free saliency map for downstream editing tasks.
  • Failure cases would likely appear on prompts whose aesthetic language is vague or globally distributed (e.g., “overall cinematic lighting”), testing how far the local-mask assumption stretches.

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 / 0 minor

Summary. The manuscript presents DINOLight, a training-based ambient light normalization (ALN) framework that injects DINOv2 features as a visual prior. Features from layers 1, 6, and 12 of a ViT-B/14 DINOv2 are fused by an Adaptive Feature Fusion Module (AFFM) that produces point-wise softmax weights (Eq. 1). The fused features are then injected into a hierarchical Transformer restoration network via Auxiliary Cross-Attention (ACA) operating in both spatial and frequency domains inside SFDINO blocks (Eq. 2). The network is trained end-to-end with an L1 + MS-SSIM loss on Ambient6K and is also evaluated on ISTD/ISTD+ shadow-removal benchmarks without using shadow masks. Reported results claim state-of-the-art PSNR/SSIM/LPIPS on Ambient6K (Table 1) and competitive mask-free RMSE on ISTD/ISTD+ (Table 4).

Significance. If the claims hold, the work supplies a practical mask-free route to ALN that exploits geometric and semantic cues already present in a frozen foundation model, together with a dual-domain attention design that is reusable for other lighting-related restoration tasks. Strengths include a clear two-stage architecture, ablations that isolate layer choice, AFFM, and sequential versus parallel ACA (Tables 2–3), and competitive transfer to conventional shadow removal without mask priors. The public Ambient6K benchmark and the promise of code release further support reproducibility. The contribution is incremental rather than foundational, but it is a concrete and useful step for real-world lighting normalization.

major comments (3)
  1. The supplied manuscript body is DINOLight (ambient light normalization), while the arXiv identifier and abstract describe AccelAes (training-free DiT acceleration). No AesMask, SkipSparse, step-level cache, Lumina-Next numbers, or ImageReward results appear. The central claims of the abstract therefore cannot be audited against any methods, equations, or tables in the body; this identity mismatch must be resolved before any technical evaluation of AccelAes is possible.
  2. Table 1 reports MACs both with and without the DINOv2 backbone (7.89 G vs 16.86 G). Because DINOv2 is an inseparable part of the inference pipeline, the efficiency comparison with IFBlend and PromptNorm is incomplete unless wall-clock latency and peak memory that include the ViT forward pass are also reported under identical hardware.
  3. Section 4.2 ablations (Tables 2–3) are trained for only 200 K iterations with batch size 2, while the main model uses 400 K iterations and batch size 4. Absolute PSNR numbers therefore cannot be compared directly to Table 1; the relative ranking of AFFM and ACA components may still be informative, but the magnitude of gains is not established under the same training budget.

Circularity Check

0 steps flagged

No circularity: DINOLight is an empirical architecture paper whose claims rest on external benchmarks and ablations, not on self-definitional or fitted-as-prediction steps.

full rationale

The provided full manuscript is DINOLight (ambient light normalization via DINOv2 priors, AFFM, and dual-domain ACA), not AccelAes. Its derivation chain is architectural and empirical: DINOv2 layer features are extracted, fused by a learned point-wise softmax mask (Eq. 1), and injected via Auxiliary Cross-Attention (Eq. 2) into a hierarchical Transformer; the network is trained end-to-end with L1 + MS-SSIM on Ambient6K and evaluated on held-out Ambient6K, ISTD, and ISTD+ test sets against independent baselines. No equation equates a claimed prediction to a fitted constant by construction; no uniqueness theorem or ansatz is imported from overlapping-author prior work as a load-bearing premise; ablations (Tables 2–3) simply remove components and re-measure PSNR/SSIM/LPIPS. Self-citations are absent from the central argument. The paper is therefore self-contained against external data, yielding score 0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

Abstract-only review of AccelAes. Load-bearing content is observational and engineering, not axiomatic math. Free parameters (mask thresholds, cache period, guidance reweighting) are implied but not quantified. No invented physical entities. Domain assumptions about DiT cross-attention and temporal redundancy are the main unproved premises.

free parameters (3)
  • AesMask construction thresholds / fusion weights
    One-shot mask from prompt semantics and cross-attention necessarily involves cutoffs or aggregation choices not specified in the abstract; quality and speed claims depend on them.
  • SkipSparse reallocation / guidance schedule
    How much compute and guidance is moved to masked vs unmasked regions is a design knob that determines both speedup and aesthetic scores.
  • Step-level prediction cache period / replacement policy
    How often full Transformer evaluations are replaced by cache is a free schedule parameter controlling the temporal reduction factor.
axioms (3)
  • domain assumption Denoising is spatially non-uniform w.r.t. aesthetic prompt descriptors: aesthetic-token regions get concentrated cross-attention and larger temporal variation; low-affinity regions evolve smoothly with redundant compute.
    Stated as the core observation motivating AesMask and SkipSparse; if false on many prompts, both acceleration safety and aesthetic gains fail.
  • ad hoc to paper A one-shot aesthetic focus mask plus sparse reallocation can improve (not only preserve) perceptual aesthetics while reducing FLOPs, without training.
    The dual claim of speedup and quality improvement rests on this design hypothesis; abstract does not derive it.
  • domain assumption Periodic replacement of full DiT evaluations by a lightweight step-level prediction cache preserves generation fidelity for the reported schedules.
    Temporal reduction assumes sufficient step-to-step redundancy outside full evaluations.
invented entities (2)
  • AesMask no independent evidence
    purpose: One-shot aesthetic focus mask from prompt semantics and cross-attention to drive localized compute.
    Named construct introduced by the paper; no independent external definition in the abstract.
  • SkipSparse no independent evidence
    purpose: Reallocate computation and guidance to AesMask regions when localized computation is feasible.
    Named module of the framework; evidence is internal experimental claim only.

pith-pipeline@v1.1.0-grok45 · 18403 in / 2693 out tokens · 30212 ms · 2026-07-14T22:13:55.101336+00:00 · methodology

0 comments
read the original abstract

Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.

discussion (0)

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

Forward citations

Cited by 4 Pith papers

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

  1. Learning Color Grading, No Photo Sharing: Federated Aesthetic Preference Learning for Personalized Image Enhancement

    cs.CV 2026-07 conditional novelty 7.0

    FedPAIE trains a shared aesthetic scorer with federated learning, calibrates it per user from sparse private ratings, and uses it to guide a small color-LUT enhancer on unpaired local photos.

  2. Aes3D: Aesthetic Assessment in 3D Gaussian Splatting

    cs.CV 2026-05 unverdicted novelty 7.0

    Aes3D creates the first dedicated dataset for 3D scene aesthetics and a model that predicts aesthetic scores straight from 3D Gaussian primitives.

  3. SAFE-DiT: Semantics-Aware Fast-path Execution for High-Resolution Diffusion Transformers

    cs.CV 2026-06 unverdicted novelty 6.0

    SAFE-DiT accelerates DiT inference at high resolutions by eliding mask-induced dispatch tax in attention while preserving semantics via selective scheduling, delivering up to 5x speedup and large memory savings with v...

  4. Aes3D: Aesthetic Assessment in 3D Gaussian Splatting

    cs.CV 2026-05 unverdicted novelty 6.0

    Aes3D creates the first 3D scene aesthetic assessment dataset and a model that regresses aesthetic scores from 3DGS representations alone.

Reference graph

Works this paper leans on

59 extracted references · 8 linked inside Pith · cited by 3 Pith papers

  1. [1]

    IEEE transactions on pattern analysis and machine intelligence33(6), 1202–1216 (2010)

    Arbel, E., Hel-Or, H.: Shadow removal using intensity surfaces and texture an- chor points. IEEE transactions on pattern analysis and machine intelligence33(6), 1202–1216 (2010)

    2010

  2. [2]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  3. [3]

    In: European conference on computer vision

    Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European conference on computer vision. pp. 17–33. Springer (2022)

    2022

  4. [4]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 182–192 (2021)

    2021

  5. [5]

    In: The eleventh international conference on learning representations (2023)

    Cui, Y., Tao, Y., Bing, Z., Ren, W., Gao, X., Cao, X., Huang, K., Knoll, A.: Selective frequency network for image restoration. In: The eleventh international conference on learning representations (2023)

    2023

  6. [6]

    In: Proceedings of the AAAI conference on artificial intelligence

    Cun, X., Pun, C.M., Shi, C.: Towards ghost-free shadow removal via dual hierar- chical aggregation network and shadow matting gan. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 10680–10687 (2020)

    2020

  7. [7]

    arXiv preprint arXiv:2309.16588 (2023)

    Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need reg- isters. arXiv preprint arXiv:2309.16588 (2023)

    Pith/arXiv arXiv 2023

  8. [8]

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

    Dong, W., Zhou, H., Tian, Y., Sun, J., Liu, X., Zhai, G., Chen, J.: Shadowrefiner: Towards mask-free shadow removal via fast fourier transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6208–6217 (2024)

    2024

  9. [9]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  10. [10]

    In: European Conference on Computer Vision

    Duan, H., Min, X., Wu, S., Shen, W., Zhai, G.: Uniprocessor: a text-induced unified low-level image processor. In: European Conference on Computer Vision. pp. 180–

  11. [11]

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

    El Banani, M., Raj, A., Maninis, K.K., Kar, A., Li, Y., Rubinstein, M., Sun, D., Guibas,L.,Johnson,J.,Jampani,V.:Probingthe3dawarenessofvisualfoundation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21795–21806 (2024) DINOLight 13

    2024

  12. [12]

    Neural networks107, 3–11 (2018)

    Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-weighted linear units for neural net- work function approximation in reinforcement learning. Neural networks107, 3–11 (2018)

    2018

  13. [13]

    International Journal of Computer Vision85(1), 35–57 (2009)

    Finlayson, G.D., Drew, M.S., Lu, C.: Entropy minimization for shadow removal. International Journal of Computer Vision85(1), 35–57 (2009)

    2009

  14. [14]

    In: Computer Vision—ECCV 2002: 7th European Conference on Computer Vision Copenhagen, Denmark, May 28–31, 2002 Proceedings, Part IV 7

    Finlayson, G.D., Hordley, S.D., Drew, M.S.: Removing shadows from images. In: Computer Vision—ECCV 2002: 7th European Conference on Computer Vision Copenhagen, Denmark, May 28–31, 2002 Proceedings, Part IV 7. pp. 823–836. Springer (2002)

    2002

  15. [15]

    Fu, L., Zhou, C., Guo, Q., Juefei-Xu, F., Yu, H., Feng, W., Liu, Y., Wang, S.: Auto- exposurefusionforsingle-imageshadowremoval.In:ProceedingsoftheIEEE/CVF conference on computer vision and pattern recognition. pp. 10571–10580 (2021)

    2021

  16. [16]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Fu, M., Liu, H., Yu, Y., Chen, J., Wang, K.: Dw-gan: A discrete wavelet transform gan for nonhomogeneous dehazing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 203–212 (2021)

    2021

  17. [17]

    In: European Conference on Computer Vision

    Gaintseva, T., Benning, M., Slabaugh, G.: Rave: Residual vector embedding for clip-guided backlit image enhancement. In: European Conference on Computer Vision. pp. 412–428. Springer (2024)

    2024

  18. [18]

    In: Proceedings of the AAAI conference on artificial intel- ligence

    Guo, L., Huang, S., Liu, D., Cheng, H., Wen, B.: Shadowformer: global context helps shadow removal. In: Proceedings of the AAAI conference on artificial intel- ligence. vol. 37, pp. 710–718 (2023)

    2023

  19. [19]

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

    Guo, L., Wang, C., Yang, W., Huang, S., Wang, Y., Pfister, H., Wen, B.: Shad- owdiffusion: When degradation prior meets diffusion model for shadow removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14049–14058 (2023)

    2023

  20. [20]

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

    Guo, L., Wang, C., Yang, W., Wang, Y., Wen, B.: Boundary-aware divide and con- quer: A diffusion-based solution for unsupervised shadow removal. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13045–13054 (2023)

    2023

  21. [21]

    In: Proceedings of the IEEE/CVF international con- ference on computer vision

    Hu, X., Jiang, Y., Fu, C.W., Heng, P.A.: Mask-shadowgan: Learning to remove shadows from unpaired data. In: Proceedings of the IEEE/CVF international con- ference on computer vision. pp. 2472–2481 (2019)

    2019

  22. [22]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Jin, Y., Sharma, A., Tan, R.T.: Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5027–5036 (2021)

    2021

  23. [23]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Jin, Y., Ye, W., Yang, W., Yuan, Y., Tan, R.T.: Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2634–2642 (2024)

    2024

  24. [24]

    arXiv preprint arXiv:1412.6980 (2014)

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

    Pith/arXiv arXiv 2014

  25. [25]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based trans- formers for high-quality image deblurring. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 5886–5895 (2023)

    2023

  26. [26]

    In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision

    Le, H., Samaras, D.: Shadow removal via shadow image decomposition. In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8578–8587 (2019)

    2019

  27. [27]

    In: Computer Vision–ECCV2020:16thEuropeanConference,Glasgow,UK,August23–28,2020, Proceedings, Part XI 16

    Le, H., Samaras, D.: From shadow segmentation to shadow removal. In: Computer Vision–ECCV2020:16thEuropeanConference,Glasgow,UK,August23–28,2020, Proceedings, Part XI 16. pp. 264–281. Springer (2020) 14 Oh et al

    2020

  28. [28]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833–1844 (2021)

    2021

  29. [29]

    In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision

    Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for unsupervised backlit image enhancement. In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision. pp. 8094–8103 (2023)

    2023

  30. [30]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lin, J., Zhang, Z., Wei, Y., Ren, D., Jiang, D., Tian, Q., Zuo, W.: Improving image restoration through removing degradations in textual representations. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2866–2878 (2024)

    2024

  31. [31]

    arXiv preprint arXiv:2312.01677 (2023)

    Lin, X., Yue, J., Chan, K.C., Qi, L., Ren, C., Pan, J., Yang, M.H.: Multi-task image restoration guided by robust dino features. arXiv preprint arXiv:2312.01677 (2023)

    Pith/arXiv arXiv 2023

  32. [32]

    In: Computer Vision– ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part IV 10

    Liu, F., Gleicher, M.: Texture-consistent shadow removal. In: Computer Vision– ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part IV 10. pp. 437–450. Springer (2008)

    2008

  33. [33]

    IEEE Transactions on Image Processing 30, 1853–1865 (2021)

    Liu, Z., Yin, H., Mi, Y., Pu, M., Wang, S.: Shadow removal by a lightness-guided network with training on unpaired data. IEEE Transactions on Image Processing 30, 1853–1865 (2021)

    2021

  34. [34]

    arXiv preprint arXiv:1608.03983 (2016)

    Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)

    Pith/arXiv arXiv 2016

  35. [35]

    arXiv preprint arXiv:2310.01018 3(8) (2023)

    Luo, Z., Gustafsson, F.K., Zhao, Z., Sjölund, J., Schön, T.B.: Controlling vision- language models for universal image restoration. arXiv preprint arXiv:2310.01018 3(8) (2023)

    Pith/arXiv arXiv 2023

  36. [36]

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

    Luo, Z., Gustafsson, F.K., Zhao, Z., Sjölund, J., Schön, T.B.: Photo-realistic image restoration in the wild with controlled vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6641–6651 (2024)

    2024

  37. [37]

    Advances in Neural Information Processing Systems37, 76819–76847 (2025)

    Man, Y., Zheng, S., Bao, Z., Hebert, M., Gui, L., Wang, Y.X.: Lexicon3d: Probing visual foundation models for complex 3d scene understanding. Advances in Neural Information Processing Systems37, 76819–76847 (2025)

    2025

  38. [38]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Mei, K., Figueroa, L., Lin, Z., Ding, Z., Cohen, S., Patel, V.M.: Latent feature- guided diffusion models for shadow removal. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4313–4322 (2024)

    2024

  39. [39]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    Pith/arXiv arXiv 2023

  40. [40]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  41. [41]

    Raghu, M., Unterthiner, T., Kornblith, S., Zhang, C., Dosovitskiy, A.: Do vision transformers see like convolutional neural networks? Advances in neural informa- tion processing systems34, 12116–12128 (2021)

    2021

  42. [42]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021)

    2021

  43. [43]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Serrano-Lozano, D., Molina-Bakhos, F.A., Xue, D., Yang, Y., Pilligua, M., Baldrich, R., Vanrell, M., Vazquez-Corral, J.: Promptnorm: Image geometry guides ambient light normalization. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 905–916 (2025) DINOLight 15

    2025

  44. [44]

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

    Vasluianu, F.A., Romero, A., Van Gool, L., Timofte, R.: Shadow removal with paired and unpaired learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 826–835 (2021)

    2021

  45. [45]

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

    Vasluianu, F.A., Seizinger, T., Timofte, R.: Wsrd: A novel benchmark for high resolution image shadow removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1826–1835 (2023)

    2023

  46. [46]

    In: European Conference on Computer Vision

    Vasluianu, F.A., Seizinger, T., Wu, Z., Ranjan, R., Timofte, R.: Towards image ambient lighting normalization. In: European Conference on Computer Vision. pp. 385–404. Springer (2024)

    2024

  47. [47]

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

    Vasluianu, F.A., Seizinger, T., Zhou, Z., Wu, Z., Chen, C., Timofte, R., Dong, W., Zhou, H., Tian, Y., Chen, J., et al.: Ntire 2024 image shadow removal challenge report. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6547–6570 (2024)

    2024

  48. [48]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Vasluianu, F.A., Seizinger, T., Zhou, Z., Wu, Z., Timofte, R., Bao, Y., Wang, X., Lu, X., Yang, J., Hu, A., et al.: Ntire 2025 ambient lighting normalization challenge report. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 1289–1300 (2025)

    2025

  49. [49]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  50. [50]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Wang, J., Li, X., Yang, J.: Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1788–1797 (2018)

    2018

  51. [51]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 17683–17693 (2022)

    2022

  52. [52]

    In: The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003

    Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for im- age quality assessment. In: The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398–1402. Ieee (2003)

    2003

  53. [53]

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

    Xiao, J., Fu, X., Zhu, Y., Li, D., Huang, J., Zhu, K., Zha, Z.J.: Homoformer: Homogenized transformer for image shadow removal. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25617– 25626 (2024)

    2024

  54. [54]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728–5739 (2022)

    2022

  55. [55]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 14821–14831 (2021)

    2021

  56. [56]

    IEEE Transactions on Image Processing 24(11), 4623–4636 (2015)

    Zhang, L., Zhang, Q., Xiao, C.: Shadow remover: Image shadow removal based on illumination recovering optimization. IEEE Transactions on Image Processing 24(11), 4623–4636 (2015)

    2015

  57. [57]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018)

    2018

  58. [58]

    arXiv preprint arXiv:2206.10910 (2022)

    Zhang,X.F.,Gu,C.C.,Zhu,S.Y.:Spa-former:Transformerimageshadowdetection and removal via spatial attention. arXiv preprint arXiv:2206.10910 (2022)

    Pith/arXiv arXiv 2022

  59. [59]

    IEEE Transactions on Image Processing (2025)

    Zhang,X.,Ma,J.,Wang,G.,Zhang,Q.,Zhang,H.,Zhang,L.:Perceive-ir:Learning to perceive degradation better for all-in-one image restoration. IEEE Transactions on Image Processing (2025)

    2025