acceptodds
Withdrawn from ICLR 2027

VPRP: Visual Prompt Restoration Pipeline for High-Quality Nighttime Flare Removal

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Lens flare removal is essential for restoring images degraded by unwanted light scattering or reflection, yet existing methods frequently introduce artifacts in restored outputs. End-to-end approaches struggle with severe degradations where flare completely occludes underlying content, while conventional coarse-to-fine strategies yield marginal improvements when applied directly to flare removal. We identify that the core limitation lies in the lack of explicit guidance for identifying and restoring flare-corrupted regions. To address this, we propose the Visual Prompt Restoration Pipeline (VPRP), a model-agnostic framework that redefines coarse-to-fine processing through a plug-and-play Prompt Restoration Network (PRN). VPRP introduces two synergistic mechanisms: (1) a mask block that leverages predicted flare from the coarse stage to explicitly exclude corrupted pixels, enabling context-aware restoration, and (2) a prompt calibration block that repurposes high-quality decoder features as visual prompts to guide the refinement network, ensuring structural consistency. VPRP seamlessly integrates with arbitrary restoration architectures while introducing negligible computational overhead (minimum 1% parameter increase). Experiments demonstrate that VPRP significantly reduces artifacts and achieves state-of-the-art performance across multiple benchmarks, with PSNR improvements of up to 0.92 dB over strong baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Nighttime Flare Removal

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_JztUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VPRP, a two-stage framework for nighttime lens flare removal. A pretrained coarse network estimates a flare-free image and a flare component. A Prompt Restoration Network then refines the result using a mask-derived input and multiscale decoder features from the frozen coarse network. A prompt calibration block uses these features to modulate the refinement features. Experiments report improvements across several U-shaped backbones on Flare7K++, with additional evaluations on BracketFlare, FlareX, and FlareReal600. On Flare7K++, VPRP improves UNet from 27.189 to 28.109 dB and FF-Former from 27.900 to 28.443 dB.

Critical strengths

The motivation is clear and addresses a practical limitation of existing flare removal methods: residual artifacts after coarse restoration. The framework is complete, combining mask-guided refinement with feature guidance from the coarse network. The authors evaluate several backbones, including UNet, Uformer, HINet, Restormer, and FF-Former, and provide additional results on BracketFlare, FlareX, and FlareReal600 in the appendix. This makes the empirical evidence broader than a single-backbone improvement. Freezing the coarse network also makes the method potentially useful for improving existing pretrained models without retraining the entire pipeline.

Critical weaknesses

The main concern I have is about the threshold selection procedure. Appendix A.0.2 states that the authors evaluate mask IoU against ground-truth masks on the test set and then choose a threshold around 0.1 based on the observed optimum. This seems to imply that test-set annotations are used to select a method hyperparameter. The paper should clarify whether the threshold was independently fixed before this analysis or actually selected using these results. If the latter is the case, the affected evaluations should be repeated with a validation-selected threshold. Another issue is that the source of the performance gain is not fully clear. Table 2 shows that adding PCB improves the refinement network, but it does not compare PCB with ordinary feature concatenation or standard affine modulation using the same coarse-stage features. It also omits the configuration with the mask block but without PCB. Therefore, the results support the usefulness of transferring coarse features, but do not fully establish whether the proposed visual prompting mechanism is necessary or whether the two components work together as claimed. The explanation of masking is also somewhat confusing. Equation 2 supplies the complete coarse flare-free image to the refinement network alongside the masked image and predicted flare. Consequently, information from the masked regions remains available through another input branch. This does not mean the design cannot work, but describing it as explicitly excluding corrupted pixels seems stronger than what the architecture actually enforces. The paper should distinguish between providing a spatial cue and removing access to potentially unreliable information. In terms of novelty, the method currently appears incremental. The PCB uses conditional multiplicative and additive feature modulation, while transferring features between restoration stages is also an established idea. The paper would benefit from a clearer explanation of what is technically new beyond combining these operations with flare-derived masking. A controlled comparison with a conventional feature-conditioning module would make this contribution easier to assess. Finally, the efficiency claims are stronger than the reported evidence. The minimum parameter increase of around 1% corresponds to a lightweight configuration, whereas the UNet configuration achieving the approximately 0.92 dB gain increases parameters from 9.0M to 11.8M and FLOPs from 62.4G to 84.9G. These may be reasonable trade-offs, but they are not uniformly negligible. The paper should present the performance and overhead of each configuration together rather than emphasizing the lowest overhead separately from the largest improvement.

Questions for authors

1.Was the threshold around 0.1 selected using an independent validation set or the test-set ground-truth masks described in Appendix A.0.2? If test annotations informed the choice, could the authors reevaluate the affected results with a validation-selected threshold? 2.Could the authors compare PCB with ordinary concatenation-plus-convolution or standard affine feature modulation, using the same coarse features and comparable parameter and training budgets? 3.Could the authors provide the missing mask-only configuration in Table 2? This would help separate the individual contributions of masking and PCB. 4.Since the complete coarse image is still supplied in Equation 2, what does “excluding corrupted pixels” mean operationally? What happens if the complete-image input branch is removed? 5.Could the authors explicitly pair each configuration’s restoration gain with its parameter, FLOPs, and runtime overhead, and clarify the inference timing protocol?

Reviewer_6ceZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles the night time flare removal problem, they introduce Visual Prompt Restoration Pipeline (VPRP), which is a two stage pipeline: Coarse Stage: Can be any pretrained U-shaped flare removal model that estimates a flare free image and a flare component Refinement Stage: Prompt Restoration Network (PRN), refines the flare free image using two components: Mask Block: Binarizes the predicted flare component to mask flare corrupted pixels, they claim to frame the problem as a guided inpainting Prompt Calibration Block: uses the coarse stage's decoder features as visual prompts, it operates at each decoder level.

The coarse network is frozen as the PRN is trained, Experiments on the Flare7K++ are done using multiple backbones, the appendix adds results on multiple other datasets.

Critical strengths

PRN is trained on top of a frozen, pretrained flare removal network, so it can be added to existing models without retraining them Applying VPRP improves every backbone tested on Flare7K++ Table 3 reports parameters, FLOPs and runtime for each channel setting; despite a noticeable increase in parameters and FLOPs, the runtime increase is modest.

Critical weaknesses

The paper claims that the mask block explicitly excludes flare corrupted pixels, but the refinement stage input is the concatenation of F_DF, I_mask and I_coarse, and I_coarse still contains the information for the flare-corrupted pixels. The visual prompting is very similar in concept and formulation to the SFT layer of Wang et al. (CVPR 2018 https://openaccess.thecvf.com/content_cvpr_2018/papers/Wang_Recovering_Realistic_Texture_CVPR_2018_paper.pdf). The minimal overhead claim is contradicted by Table3 where for C=16 which is what the main method uses the overhead in parameters is more than 1%, also the usage of minimum in the abstract is unclear at Section A.0.2 it states that the threshold for the mask binarization process was tuned by evaluating the Intersection over Union (IoU) between the predicted flare masks and the ground truth masks on the test set, which shows a form of data leakage. Equation 2 reuses I_mask for both the mask-selected image and its complement, leaving the mathematical definition ambiguous.

Questions for authors

What does “minimum 1% parameter increase” mean, given that Table 3 shows the added parameters exceed 1% at C=16 which is the setting used in the paper?

Since I_coarse, which still contains the flare-corrupted pixels, is concatenated into the refinement input, in what sense does the mask block exclude those pixels?

How does the Prompt Calibration Block (Eq. 3) differ from the SFT layer of Wang et al. (CVPR 2018)?

Why was the binarization threshold tuned on the test set rather than a held-out validation set?

Which of the two definitions of I_mask in Equation 2 is intended?

Reviewer_6K2GOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VPRP for nighttime image flare removal. The method first uses a coarse flare removal network to predict preliminary restoration results and flare information, then employs a binary mask to obscure the contaminated regions, and feeds the decoder features from the coarse stage as guidance into a refinement network. The authors conducted experiments on Flare7K, BracKetFlare, FlareX, and FlareReal600, and integrated this module into various U-shaped restoration networks.

Critical strengths

1.The two components of the method have some experimental support. In Table 2, after adding PCB, the PSNR increased from 27.65 dB to 28.06 dB; after further adding Mask, it reached 28.24 dB. It can be seen that these two designs are effective under the current settings.

2.The authors did not test solely on a single network. Tables 1 and 3 provide results for UNet, Uformer, HINet, Restormer, FF-Former, etc., and the appendix also supplements testing on FlareX and FlareReal600, making the scope of the experiments relatively comprehensive.

Critical weaknesses

1.The method for determining the mask threshold is not rigorous enough. Appendix A.0.2 explains that the author selects the threshold based on the IoU between the predicted mask and ground truth mask on the test set, ultimately using a threshold of about 0.1. It is recommended that the author clarify whether an independent validation set was available for threshold selection, or provide results under a fixed threshold to make the experimental procedure more complete.

  1. The construction of the flare residuals for FlareReal600 needs to be clearly explained. The main text states that after gamma correction, the flare component is obtained by subtracting the flared image from the clean image; however, in Algorithm 1, the output_tensor is written as the flare-free image predicted by the coarse network. The supervisory information corresponding to the two approaches is different, and the authors need to clarify which one was actually used during training.
Questions for authors

1.Table 2 does not show results for "using only Mask without PCB." At present, it can only be stated that Mask can still provide benefits when PCB is present, making it difficult to assess its standalone effect.

2.The text refers to VPRP as model-agnostic, but all experiments are based on U-shaped networks. Does this module require the backbone network to provide encoder-decoder features at specific scales? It is recommended to make the applicable scope more specific.

3.The runtime overhead in Table 3 is moderate, but the increase in parameters and FLOPs varies across backbone networks. For example, the parameters for UNet increase from 9.0M to 11.8M. Please clarify whether the main results in Table 1 all use (C=16), and provide a more precise description of the computational overhead.

4.Table 2 demonstrates that PCB is effective, but it has not yet been compared with simpler fusion methods such as concatenation with ordinary features. Therefore, the statement 'visual prompt is non-substitutable' is somewhat strong and could be supplemented with comparisons or appropriately tempered.

Reviewer_HC2DOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes VPRP, a model-agnostic two-stage pipeline for nighttime lens flare removal that reformulates the refinement stage as guided inpainting. A plug-and-play Prompt Restoration Network (PRN) is appended to any pretrained U-shape flare removal network. It uses (1) a mask block that binarizes the predicted flare to exclude corrupted pixels, and (2) a prompt calibration block (PCB) that modulates refinement features with coarse-stage decoder features as "visual prompts". Experiments on Flare7K++ and, in the appendix, BracketFlare, FlareX, and FlareReal600 report consistent gains across five backbones with modest parameter overhead.

Critical strengths

This paper reports flare-specific metrics (G-PSNR, S-PSNR) beyond standard PSNR, SSIM, and LPIPS, demonstrating outperformance on BracketFlare, FlareX, and FlareReal600.

Critical weaknesses
  1. The core idea of the proposed method is to first mask out the corrupted region using a predicted mask, then inpaint it with a refinement network guided by coarse-stage features. This has been widely used in related layer-separation tasks:
  • visible watermark removal, e.g., SplitNet (Cun & Pun, Split then Refine: Stacked Attention-guided ResUNets for Blind Single Image Visible Watermark Removal, AAAI 2021) and SLBR (Liang et al., Visible Watermark Removal via Self-calibrated Localization and Background Refinement, ACM MM 2021);
  • reflection removal, e.g., RAGNet (Li et al., Two-stage Single Image Reflection Removal with Reflection-aware Guidance, Applied Intelligence 2023), which the paper itself cites. The paper neither discusses nor compares against these mask-then-refine designs, which limits both its novelty and the evidence of its advantage over them.
  1. The qualitative results in Fig. 3 raise concerns about reliability. On zoom-in, the PIP column, which the paper never defines and which I assume is the proposed method, is noticeably sharper than all other columns. Examples include the blinds in the second row, the text on the house in the third row, and the ceiling grid in the last row. It even appears sharper than both the input and the GT, which should not happen for a flare removal method. It is therefore impossible to tell from this figure whether PIP looks better because of a sharpness difference or because the network actually removes flare better. This also weakens the credibility of the reported quantitative results.

  2. The ablation does not establish what each component contributes, and the "visual prompt" claim is not substantiated.

  • There are no visual ablation results, so it is unclear how the mask block and the PCB each improve the final output.
  • Table 2 is incomplete: it reports PSNR only, on a single backbone (Uformer), and lacks the mask-only configuration.
  • "Visual Prompt" appears in the title, but no experiment shows that the coarse-stage decoder features actually act as a prompt. Functionally, the PCB (Eq. 3) is a standard feature modulation or skip-fusion. It is never compared against plain concatenation or addition of the same features. The authors may need to reconsider the title so that it does not mislead readers and better reflects the paper's main contribution.
Questions for authors
  1. Please discuss and compare with existing mask-then-refine methods in watermark and reflection removal (e.g., SplitNet, SLBR, RAGNet), and clarify what is new beyond applying this paradigm to flare removal.

  2. Please define "PIP" in Fig. 3 and explain why its outputs appear sharper than both the input and the GT. Confirm that all methods are visualized at identical resolution and with identical post-processing, and provide uncompressed full-resolution comparisons.

  3. Please complete the ablation:

  • Add the mask-only configuration.
  • Provide visual results for each variant.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.