Pith. sign in

REVIEW 3 major objections 6 minor 62 references

A diffusion model that gates its conditioning with reflection-severity and high-frequency priors removes reflections more cleanly while keeping scene structure.

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

2026-07-13 22:14 UTC pith:OIUMPIXX

load-bearing objection Competent SIRR systems paper: VLM+HF gated ControlNet plus refiner gives strong perceptual numbers after fair same-data finetuning; novelty is incremental and the VLM severity oracle is under-specified. the 3 major comments →

arxiv 2603.19036 v2 pith:OIUMPIXX submitted 2026-03-19 cs.CV

FUMO: Prior-Modulated Diffusion for Single Image Reflection Removal

Telang Xu , Chaoyang Zhang , Guangtao Zhai , Xiaohong Liu This is my paper
classification cs.CV
keywords reflection removalconditional generationdiffusion modelspatial priorsgated conditioningsingle-image restorationControlNet
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.

Photos taken through glass mix a wanted scene with unwanted reflections whose strength changes from place to place and whose edges often look like real scene edges. This paper claims that two maps taken from the mixed photo itself—one scoring where reflections are strong, one marking where fine detail lives—can steer a diffusion restorer so it suppresses reflections hardest only where both signals agree. The first stage is a one-step conditional diffusion pass whose residual injections are multiplied by that combined gate; the second stage is a small image-space network that fixes local drift and sharpens detail. On standard real benchmarks the method is competitive on PSNR and SSIM and leads the compared methods on average perceptual scores, and on hard internet photos it leaves fewer residual reflections and fewer broken edges. A sympathetic reader cares because the usual trade-off—wipe the reflection hard and you blur or warp the scene—is treated as a spatial control problem rather than a global one.

Core claim

The paper establishes that explicit priors for spatial reflection severity and multi-scale high-frequency response, combined into a gate on ControlNet residual injections inside a one-step diffusion restorer and followed by a fine-grained image-space refiner, improve spatially adaptive reflection suppression and structural fidelity for single-image reflection removal, producing competitive fidelity metrics and stronger average perceptual quality on Nature, Real, and SIR2 plus cleaner results on in-the-wild mixtures.

What carries the argument

Prior-modulated gated residual injection: the gate g = 1 + β Pint ⊙ Phf multiplies multi-scale ControlNet residuals so conditioning is strengthened only where reflections are both severe and structure-sensitive; a subsequent fine-grained refinement module then corrects geometry and detail in image space.

Load-bearing premise

The load-bearing premise is that a vision-language model’s patch scores and reflection boxes, filtered into an intensity map, are accurate enough that multiplying them with a high-frequency map correctly tells the network where to push harder.

What would settle it

Replace the VLM intensity map with a deliberately scrambled or inverted severity field (or with human-drawn severity maps) while holding the high-frequency prior and all training fixed; if gated FUMO no longer improves LPIPS/MUSIQ and wild-image cleanliness over the ungated baseline, the claim that these priors drive the gain fails.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes FUMO, a coarse-to-fine single-image reflection removal framework built on a ControlNet-conditioned latent diffusion backbone. From the mixed input it extracts two explicit priors: a VLM-based reflection intensity map Pint (patch-level ordinal scoring plus image-level bounding-box boost, densified by guided filtering; Sec. 3.1, Eqs. 1–3) and a multi-scale residual high-frequency map Phf (Eq. 4). These are combined into a spatial gate g=1+β Pint⊙Phf (Eq. 5) that modulates ControlNet residual injections during a one-step coarse restoration (Eqs. 6, 9). A subsequent Fine-Grained Refinement Module (FGRM) with SimpleGate activations corrects geometry and details under a pixel/perceptual/gradient loss (Eqs. 7, 10–13). On Nature, Real, and SIR2 the method is competitive on PSNR/SSIM and best on average LPIPS (0.069) and MUSIQ (59.88) among the compared methods (Table 1); ablations isolate gate composition and the refinement stage (Table 2, Figs. 6–7). Code is released.

Significance. Spatially varying reflection strength and structure–reflection entanglement remain central obstacles in real-world SIRR. Framing the problem as prior-modulated conditional diffusion with an explicit severity×detail gate is a clear systems contribution, and the coarse-to-fine split is well motivated by the observed trade-off between aggressive suppression and geometric fidelity. Strengths include: (i) fairer comparison by finetuning public competitors on the same training mix (Sec. 4.2); (ii) ablations that isolate gate composition and refinement (Table 2); (iii) both full-reference and no-reference metrics plus in-the-wild qualitative checks (Figs. 1, 5); and (iv) public code. If the gated-prior story holds under better-specified and validated priors, the work is a useful incremental advance for diffusion-based restoration and SIRR practice, not a foundational theoretical result.

major comments (3)
  1. [Sec. 3.1, Eqs. (1)–(3)] Sec. 3.1, Eqs. (1)–(3): the intensity prior Pint is load-bearing for the paper’s central claim of spatially adaptive, prior-modulated conditioning (Abstract; Eqs. 5–6), yet the VLM pipeline is neither named nor validated. The manuscript never states which VLM is used, the exact prompts, adaptive patch size a, temperature τ, ordinal weights w(c), box-boost factor, or guided-filter (r, ε). There is also no calibration or correlation study against reflection masks, human severity labels, or synthetic mixtures with known γ2. Table 2 shows that some spatial gate helps under a fixed (unspecified) VLM, but that does not establish that Pint ranks or localizes reflection severity correctly. Without this, the claimed spatial controllability of g=1+β Pint⊙Phf remains an untested assumption, and the headline perceptual gains (Table 1 avg LPIPS/MUSIQ) cannot be attributed to accurate severity guidanc
  2. [Eqs. (5)–(6); Table 2; Fig. 6] Eqs. (5)–(6) and Table 2: the gate is defined as a multiplicative interaction of severity and high-frequency content, then applied as a clipped residual scale on ControlNet features. The ablations compare w/o-gate, intensity-only, HF-only, and full gate on aggregate metrics, but provide no spatial diagnostic that the gate is high where reflections are strong and low where transmission structure should be preserved (e.g., gate–mask IoU, severity-stratified error, or failure cases on multi-layer/night/glass scenes highlighted in Fig. 1). Given that Phf is extracted from the mixture and therefore contains reflection high frequencies as well as transmission detail, the product Pint⊙Phf can amplify residual injection on reflection edges. A load-bearing check is needed that the gate does not systematically over-modulate reflection structure; otherwise the mechanistic story of “reflection-domin
  3. [Table 1; Sec. 4.2] Table 1 and Sec. 4.2: on SIR2 (500 images), DAI reports higher PSNR (27.35 vs 27.22) while FUMO leads on SSIM/LPIPS/CLIPIQA/MUSIQ; average PSNR is likewise slightly below DAI (27.15 vs 27.24). The abstract and conclusion emphasize “competitive quantitative results and consistently improved perceptual quality,” which is broadly fair, but the manuscript should explicitly discuss this fidelity–perception trade-off and whether one-step coarse restoration plus FGRM systematically trades PSNR for LPIPS/MUSIQ. Without that discussion, the claim of overall superiority over the closest diffusion baseline (DAI) is overstated relative to the table.
minor comments (6)
  1. [Abstract; Sec. 1] Abstract and Sec. 1 state that priors are “extracted directly from the mixed image.” For Pint this is misleading: extraction depends on an external VLM and its prompts/logits. Rephrase to “derived from the mixed image via a VLM and multi-scale residual analysis.”
  2. [Sec. 4.1; Abstract] Sec. 4.1: “fine-graine refinement module” is a typo (should be “fine-grained”). Several abstract/intro sentences also show missing spaces (e.g., “multi-scaleresidualaggregation,” “Weproposeacoarse-to-fine”), likely PDF extraction or typesetting issues that should be cleaned.
  3. [Sec. 3.3, Eq. (9)] Eq. (9) and surrounding text: the one-step training samples target timestep t uniformly with N=1000 and at inference sets t=0. Briefly justify why uniform t (vs fixed small t or noise-prediction multi-step) is preferred for SIRR, and report sensitivity if available.
  4. [Figs. 1, 5] Fig. 1 and Fig. 5: in-the-wild examples are persuasive but lack any quantitative proxy (e.g., no-reference scores on the wild set, or user study). Even a small table of CLIPIQA/MUSIQ on the wild set would strengthen the qualitative claim.
  5. [Sec. 2.2] Related work (Sec. 2.2) cites L-DiffER, DAI, and PolarFree appropriately; a short explicit contrast of FUMO’s gate vs DAI’s ControlNet+refined decoder would help readers place the novelty beyond “add priors and refine.”
  6. [Secs. 3.2–3.3] Hyperparameters β_max=0.25, warmup 0.1, and λ_pix/λ_perc/λ_grad=0.5/0.25/0.25 are stated as empirical (Secs. 3.2–3.3) with no sensitivity. A brief appendix sweep would aid reproducibility.

Circularity Check

0 steps flagged

No circular derivation: FUMO is an empirical systems method whose claims rest on held-out reconstruction metrics, not on identities forced by definition or self-citation.

full rationale

The paper proposes a practical SIRR pipeline: VLM-based intensity prior and multi-scale residual high-frequency prior (Sec. 3.1, Eqs. 1–4), gated ControlNet residual injection into a one-step diffusion coarse stage (Eqs. 5–6, 9), and a separate image-space refinement network trained with L1/LPIPS/gradient losses against ground-truth transmission (Eqs. 7, 10–13). Quantitative claims (Table 1) are PSNR/SSIM/LPIPS/CLIPIQA/MUSIQ on held-out Nature/Real/SIR2 splits and qualitative wild images; ablations (Table 2, Figs. 6–7) compare gate and refinement variants under the same training protocol. Nothing in the chain is algebraically equivalent to its inputs by construction: the gate is a design choice, not a fitted identity renamed as prediction; losses optimize against external targets; priors are extracted from the mixture and used as conditioning, not as the quantity being “predicted.” Self-citations (e.g., related diffusion/SIRR work) are background, not load-bearing uniqueness theorems. Weaknesses of the VLM oracle are validation/correctness risks, not circularity. Score 0 with empty steps is the honest finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing content is mostly engineering assumptions and free hyperparameters, not physical postulates. The central empirical claim rests on (i) VLM severity/localization being a valid reflection prior, (ii) multiplicative gating of ControlNet residuals being a useful control, (iii) one-step latent regression being adequate for coarse SIRR, and (iv) several hand-set scalars (β_max, loss weights, warmup). No new physical entity is claimed; invented constructs are method modules.

free parameters (5)
  • β_max (gate strength)
    Maximum modulation amplitude in Eq. 6; set empirically to 0.25 with warmup ratio 0.1 (Sec. 3.2). Directly scales how much priors alter conditioning.
  • refinement loss weights λ_pix, λ_perc, λ_grad
    Set to 0.5 / 0.25 / 0.25 empirically (Eq. 13, Sec. 3.3); control the trade-off among pixel, LPIPS, and edge terms that define the refined output.
  • VLM temperature τ and ordinal weights w(c)
    Softmax temperature and fixed weights {1..5} in Eqs. 1–2 convert logits into continuous severity; choices affect Pint and thus the gate.
  • synthetic blend coefficients γ1, γ2
    Per-channel sampling in M=γ1T+γ2R−γ1γ2T⊙R (Sec. 4.1, following DSRNet) shapes a large fraction of training data and therefore the learned prior.
  • one-step target timestep sampling / N=1000
    Training samples t uniformly with maximally noised z_N (Eq. 9); inference uses t=0. This design choice defines the coarse restoration objective.
axioms (5)
  • domain assumption A pretrained VLM’s restricted next-token logits over {None, Minor, Mid, Major, Critical} plus reflection bounding boxes yield a usable spatial reflection-severity field after guided filtering.
    Sec. 3.1 Branch I; entire intensity prior and gate depend on this without independent calibration against human severity maps.
  • domain assumption Multi-scale dilated-convolution residual aggregation from the mixture is a valid detail-sensitive guidance signal even though it mixes transmission and reflection high frequencies.
    Sec. 3.1 Branch II, Eq. 4; paper states Phf is guidance-only because it contains both layers.
  • domain assumption One-step latent regression with ControlNet residual injection (following cited one-step restoration practice) is sufficient for coarse reflection suppression under gated modulation.
    Sec. 3.2–3.3, Eq. 9; freezes most SD U-Net and VAE, trains ControlNet + upsampling blocks only.
  • domain assumption Standard image-formation / evaluation assumptions for SIRR: paired mixed/transmission data, PSNR/SSIM/LPIPS/CLIPIQA/MUSIQ as quality proxies.
    Sec. 4.1–4.2; usual CV restoration evaluation regime.
  • ad hoc to paper Elementwise residual gating ecs = clip(Is(g),1,1+β_max) ⊙ cs is a stable and meaningful way to spatially control ControlNet features.
    Eqs. 5–6; design choice specific to this framework, supported by ablation but not derived.
invented entities (3)
  • Dual prior extraction pipeline (VLM intensity prior Pint + multi-scale residual Phf) no independent evidence
    purpose: Provide explicit spatial severity and detail cues from the mixed image alone for gated conditioning.
    Defined in Sec. 3.1 and Fig. 2; central methodological construct. No external physical referent beyond the maps themselves.
  • Prior-modulated residual gate g = 1 + β Pint ⊙ Phf no independent evidence
    purpose: Spatially strengthen ControlNet residual injections where reflections are severe and structures are detail-rich.
    Eqs. 5–6; paper-specific control mechanism validated only by internal ablations.
  • Fine-Grained Refinement Module (FGRM) with SimpleGate no independent evidence
    purpose: Correct geometric drift and restore sharp details after aggressive coarse diffusion restoration.
    Sec. 3.2 Eq. 7; architectural module trained with composite image-space losses.

pith-pipeline@v1.1.0-grok45 · 18540 in / 3988 out tokens · 39463 ms · 2026-07-13T22:14:02.801381+00:00 · methodology

0 comments
read the original abstract

Single image reflection removal (SIRR) is challenging in real scenes, where reflection strength varies spatially and reflection patterns are tightly entangled with transmission structures. This paper presents a diffusion model with prior modulation framework (FUMO) that introduces explicit priors for spatially adaptive conditioning and structurally faithful restoration. Two priors are extracted directly from the mixed image, an intensity prior that estimates spatial reflection severity and a high-frequency prior that captures detail-sensitive responses via multi-scale residual aggregation. We propose a coarse-to-fine training paradigm. In the first stage, these cues are combined to gate the conditional residual injections, focusing the conditioning on regions that are both reflection-dominant and structure-sensitive. In the second stage, a fine-grained refinement network corrects local misalignment and sharpens fine details in the image space. Experiments conducted on both standard benchmarks and challenging images in the wild demonstrate competitive quantitative results and consistently improved perceptual quality. The code is released at https://github.com/Lucious-Desmon/FUMO.

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

62 extracted references · 11 linked inside Pith

  1. [1]

    In: ACM SIGGRAPH 2005 Papers, pp

    Agrawal, A., Raskar, R., Nayar, S.K., Li, Y.: Removing photography artifacts using gradient projection and flash-exposure sampling. In: ACM SIGGRAPH 2005 Papers, pp. 828–835 (2005)

    2005

  2. [2]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    Arvanitopoulos, N., Achanta, R., Susstrunk, S.: Single image reflection suppres- sion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (July 2017)

    2017

  3. [3]

    arXiv preprint arXiv:2502.13923 (2025)

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    IEEETransactionsonCircuitsandSystemsforVideoTechnology35(5),4397–4408 (2024)

    Chen, G.Y., Zheng, C.W., Fan, G., Su, J.N., Gan, M., Chen, C.P.: Real-world image reflection removal: An ultra-high-definition dataset and an efficient baseline. IEEETransactionsonCircuitsandSystemsforVideoTechnology35(5),4397–4408 (2024)

    2024

  5. [5]

    arXiv preprint arXiv:2502.20172 (2025)

    Chen, L., Bai, S., Chai, W., Xie, W., Zhao, H., Vinci, L., Lin, J., Chang, B.: Multimodal representation alignment for image generation: Text-image interleaved control is easier than you think. arXiv preprint arXiv:2502.20172 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    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

  7. [7]

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

    Dong, Z., Xu, K., Yang, Y., Bao, H., Xu, W., Lau, R.W.: Location-aware single im- age reflection removal. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5017–5026 (2021)

    2021

  8. [8]

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

    Han, B.J., Sim, J.Y.: Reflection removal using low-rank matrix completion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5438–5446 (2017)

    2017

  9. [9]

    IEEE transactions on pattern analysis and machine intelligence35(6), 1397–1409 (2012)

    He, K., Sun, J., Tang, X.: Guided image filtering. IEEE transactions on pattern analysis and machine intelligence35(6), 1397–1409 (2012)

    2012

  10. [10]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  11. [11]

    In: European Conference on Computer Vision

    Hong, Y., Zhong, H., Weng, S., Liang, J., Shi, B.: L-differ: Single image reflec- tion removal with language-based diffusion model. In: European Conference on Computer Vision. pp. 58–76. Springer (2024) 16 T. Xu et al

    2024

  12. [12]

    arXiv preprint arXiv:2503.17347 (2025)

    Hu, J., Yang, C., Zhou, Z., Fang, J., Yang, X., Tian, Q., Shen, W.: Dereflection any image with diffusion priors and diversified data. arXiv preprint arXiv:2503.17347 (2025)

    arXiv 2025

  13. [13]

    Advances in Neural Information Processing Systems 34, 24683–24694 (2021)

    Hu, Q., Guo, X.: Trash or treasure? an interactive dual-stream strategy for single image reflection separation. Advances in Neural Information Processing Systems 34, 24683–24694 (2021)

    2021

  14. [14]

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

    Hu, Q., Guo, X.: Single image reflection separation via component synergy. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13138–13147 (2023)

    2023

  15. [15]

    Electronics letters44(13), 800–801 (2008)

    Huynh-Thu, Q., Ghanbari, M.: Scope of validity of psnr in image/video quality assessment. Electronics letters44(13), 800–801 (2008)

    2008

  16. [16]

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

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5148–5157 (2021)

    2021

  17. [17]

    arXiv preprint arXiv:1412.6980 (2014)

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

    Pith/arXiv arXiv 2014

  18. [18]

    Advances in neural information processing systems25 (2012)

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con- volutional neural networks. Advances in neural information processing systems25 (2012)

    2012

  19. [19]

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

    Lei, C., Chen, Q.: Robust reflection removal with reflection-free flash-only cues. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14811–14820 (2021)

    2021

  20. [20]

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

    Lei, C., Huang, X., Zhang, M., Yan, Q., Sun, W., Chen, Q.: Polarized reflection removal with perfect alignment in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1750–1758 (2020)

    2020

  21. [21]

    IEEE Transactions on Pattern Analysis and Machine Intel- ligence29(9), 1647–1654 (2007)

    Levin, A., Weiss, Y.: User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intel- ligence29(9), 1647–1654 (2007)

    2007

  22. [22]

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

    Li, C., Yang, Y., He, K., Lin, S., Hopcroft, J.E.: Single image reflection removal through cascaded refinement. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3565–3574 (2020)

    2020

  23. [23]

    Applied Intelligence53(16), 19433–19448 (2023)

    Li, Y., Liu, M., Yi, Y., Li, Q., Ren, D., Zuo, W.: Two-stage single image reflection removal with reflection-aware guidance. Applied Intelligence53(16), 19433–19448 (2023)

    2023

  24. [24]

    arXiv preprint arXiv:2501.02189 (2025)

    Li, Z., Wu, X., Du, H., Liu, F., Nghiem, H., Shi, G.: A survey of state of the art large vision language models: Alignment, benchmark, evaluations and challenges. arXiv preprint arXiv:2501.02189 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014)

    2014

  26. [26]

    arXiv preprint arXiv:2507.20590 (2025)

    Lin, X., Yu, F., Hu, J., You, Z., Shi, W., Ren, J.S., Gu, J., Dong, C.: Harnessing diffusion-yielded score priors for image restoration. arXiv preprint arXiv:2507.20590 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

    arXiv preprint arXiv:2504.17761 (2025)

    Liu, S., Han, Y., Xing, P., Yin, F., Wang, R., Cheng, W., Liao, J., Wang, Y., Fu, H., Han, C., et al.: Step1x-edit: A practical framework for general image editing. arXiv preprint arXiv:2504.17761 (2025)

    Pith/arXiv arXiv 2025

  28. [28]

    In: European Conference on Computer Vision

    Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: European Conference on Computer Vision. pp. 182–

  29. [29]

    In: International conference on machine learning

    Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International conference on machine learning. pp. 8162–8171. PMLR (2021) FUMO 17

    2021

  30. [30]

    Advances in neural information processing sys- tems32(2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing sys- tems32(2019)

    2019

  31. [31]

    Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

    2023

  32. [32]

    arXiv preprint arXiv:2307.01952 (2023)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    Pith/arXiv arXiv 2023

  33. [33]

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

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  34. [34]

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

    Shih,Y.,Krishnan,D.,Durand,F.,Freeman,W.T.:Reflectionremovalusingghost- ing cues. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3193–3201 (2015)

    2015

  35. [35]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2015)

    Simon, C., Kyu Park, I.: Reflection removal for in-vehicle black box videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2015)

    2015

  36. [36]

    In: Proceedings IEEE Conference on Com- puter Vision and Pattern Recognition

    Szeliski, R., Avidan, S., Anandan, P.: Layer extraction from multiple images con- taining reflections and transparency. In: Proceedings IEEE Conference on Com- puter Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662). vol. 1, pp. 246–253. IEEE (2000)

    2000

  37. [37]

    In: Proceedings of the IEEE international conference on computer vision

    Wan, R., Shi, B., Duan, L.Y., Tan, A.H., Kot, A.C.: Benchmarking single-image reflection removal algorithms. In: Proceedings of the IEEE international conference on computer vision. pp. 3922–3930 (2017)

    2017

  38. [38]

    International Journal of Computer Vision129(2), 385–399 (2021)

    Wan, R., Shi, B., Li, H., Duan, L.Y., Kot, A.C.: Face image reflection removal. International Journal of Computer Vision129(2), 385–399 (2021)

    2021

  39. [39]

    In: Proceedings of the AAAI conference on artificial intelligence

    Wang, J., Chan, K.C., Loy, C.C.: Exploring clip for assessing the look and feel of images. In: Proceedings of the AAAI conference on artificial intelligence. vol. 37, pp. 2555–2563 (2023)

    2023

  40. [40]

    International Journal of Computer Vision 132(12), 5929–5949 (2024)

    Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision 132(12), 5929–5949 (2024)

    2024

  41. [41]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, R., Zheng, Y., Zhang, Z., Li, C., Liu, S., Zhai, G., Liu, X.: Learning haz- ing to dehazing: Towards realistic haze generation for real-world image dehazing. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 23091–23100 (2025)

    2025

  42. [42]

    5: Advancing open-source multimodal models in versatility, reasoning, and efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    Pith/arXiv arXiv 2025

  43. [43]

    In: The thrity-seventh Asilomar conference on signals, systems & computers, 2003

    Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The thrity-seventh Asilomar conference on signals, systems & computers, 2003. vol. 2, pp. 1398–1402. IEEE (2003)

    2003

  44. [44]

    arXiv preprint arXiv:2509.04548 (2025)

    Wei, H., Xu, B., Liu, H., Wu, S., Liu, J., Peng, Y., Wang, P., Liu, Z., He, J., Xietian, Y., et al.: Skywork unipic 2.0: Building kontext model with online rl for unified multimodal model. arXiv preprint arXiv:2509.04548 (2025)

    arXiv 2025

  45. [45]

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

    Wei, K., Yang, J., Fu, Y., Wipf, D., Huang, H.: Single image reflection removal exploiting misaligned training data and network enhancements. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8178–8187 (2019) 18 T. Xu et al

    2019

  46. [46]

    arXiv preprint arXiv:2312.17090 (2023)

    Wu, H., Zhang, Z., Zhang, W., Chen, C., Liao, L., Li, C., Gao, Y., Wang, A., Zhang, E., Sun, W., et al.: Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090 (2023)

    Pith/arXiv arXiv 2023

  47. [47]

    Advances in Neural Information Processing Systems 37, 92529–92553 (2024)

    Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real- world image super-resolution. Advances in Neural Information Processing Systems 37, 92529–92553 (2024)

    2024

  48. [48]

    Xu, G., Ge, Y., Liu, M., Fan, C., Xie, K., Zhao, Z., Chen, H., Shen, C.: What matters when repurposing diffusion models for general dense perception tasks? arXiv preprint arXiv:2403.06090 (2024)

    Pith/arXiv arXiv 2024

  49. [49]

    ACM Transactions on Graphics (TOG)34(4), 1–11 (2015)

    Xue, T., Rubinstein, M., Liu, C., Freeman, W.T.: A computational approach for obstruction-free photography. ACM Transactions on Graphics (TOG)34(4), 1–11 (2015)

    2015

  50. [50]

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

    Yang, J., Li, H., Dai, Y., Tan, R.T.: Robust optical flow estimation of double-layer images under transparency or reflection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1410–1419 (2016)

    2016

  51. [51]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    Yang, K., Cai, J., Ouyang, L., Vasluianu, F.A., Timofte, R., Ding, J., Sun, H., Fu, L., Li, J., Ho, C.M., Meng, Z., Li, M., Wang, H., Hu, Q., Wang, J., Zhao, H., Hu, J., Guo, X., Yang, M., He, J., Wang, Y., Chen, Z., Fang, H., Zhang, W., Cong, R., Hegde, D.D., Kalal, J., Akalwadi, N., Tabib, R.A., Mudenagudi, U., Lin, Y.F., Lee, C.M., Hsu, C.C., Zhang, M....

    2025

  52. [52]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yao, M., Wang, M., Tam, K.M., Li, L., Xue, T., Gu, J.: Polarfree: Polarization- based reflection-free imaging. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10890–10899 (2025)

    2025

  53. [53]

    ACM Trans- actions on Graphics (TOG)43(6), 1–18 (2024)

    Ye, C., Qiu, L., Gu, X., Zuo, Q., Wu, Y., Dong, Z., Bo, L., Xiu, Y., Han, X.: Sta- blenormal: Reducing diffusion variance for stable and sharp normal. ACM Trans- actions on Graphics (TOG)43(6), 1–18 (2024)

    2024

  54. [54]

    arXiv preprint arXiv:1511.07122 (2015)

    Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2015)

    Pith/arXiv arXiv 2015

  55. [55]

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

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3836–3847 (2023)

    2023

  56. [56]

    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

  57. [57]

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

    Zhang, X., Ng, R., Chen, Q.: Single image reflection separation with perceptual losses. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4786–4794 (2018)

    2018

  58. [58]

    In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference

    Zhao, H., Li, M., Hu, Q., Guo, X.: Reversible decoupling network for single image reflection removal. In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference. pp. 26430–26439 (2025)

    2025

  59. [59]

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

    Zheng, Q., Shi, B., Chen, J., Jiang, X., Duan, L.Y., Kot, A.C.: Single image reflec- tion removal with absorption effect. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13395–13404 (2021)

    2021

  60. [60]

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

    Zhong, H., Hong, Y., Weng, S., Liang, J., Shi, B.: Language-guided image reflection separation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24913–24922 (2024) FUMO 19

    2024

  61. [61]

    In: Proceedings of the AAAI Confer- ence on Artificial Intelligence

    Zhou, H., Dong, W., Liu, X., Zhang, Y., Zhai, G., Chen, J.: Low-light image en- hancement via generative perceptual priors. In: Proceedings of the AAAI Confer- ence on Artificial Intelligence. vol. 39, pp. 10752–10760 (2025)

    2025

  62. [62]

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

    Zhu, Y., Fu, X., Jiang, P.T., Zhang, H., Sun, Q., Chen, J., Zha, Z.J., Li, B.: Revis- iting single image reflection removal in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25468–25478 (2024)

    2024