{"id":"02d35bfb-63e4-43ae-9e12-8bee4db77eac","arxiv_id":"2412.13655","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VIIS is a diffusion-based method that jointly enhances and fuses visible and infrared images to restore severely dark scenes, trained with an augmentation-based pretext task.","lead":"This paper proposes a method that combines visible and infrared images to enhance photos taken in very dark conditions, restoring details that ordinary low-light enhancement cannot. It introduces a new task, VIIS, and a diffusion model with a sparse attention mechanism to fuse and enhance both modalities at once.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central SOTA claim rests on a circular fidelity experiment (Sec. 9) and no-reference metrics on a subjectively chosen real test set; no non-circular evidence shows ISPT training transfers to real severe low-light imagery.","rationale":"The reader's conditional verdict is appropriate. I agree with the reader's weakest assumption: the ISPT degradation model is the linchpin of training, and Section 9 does not validate it because the test set uses the same augmentation. My stress test sharpens this: the failure is not merely that the augmentation may be unrealistic, but that the only quantitative experiment with a reference target is exactly on the training distribution, so it cannot discriminate between learning to enhance real darkness and learning to invert a known transform. The real-data tables depend on no-reference metrics with no significance testing, and the user study is small and on a selected subset. These limitations are visible in the manuscript itself in Sections 4.1, 5, 8, and 9, so the concern is grounded in the text rather than imposed from outside. I also note that the newly designed baselines are not trained end-to-end for the joint task, which weakens the 'also outperforms new baselines' portion of the claim. The paper has positive elements worth crediting: it formulates a genuinely new task, the SADMR ablation supports the design's contribution, code is released, and the qualitative results are plausible. Those are real evidence, but they do not overcome the lack of non-circular real-data validation. Conditional acceptance with a concrete real-data evaluation requirement remains the right calibration, so the reader's verdict should stand unchanged.","tokens_in":15200,"tokens_out":4390,"duration_ms":42728,"concrete_test":"Use the released code to evaluate on a real VIIS test set constructed from static-scene visible+IR pairs: capture nighttime visible and IR images together with a well-lit visible ground truth, or register nearest temporally adjacent daytime frames from KAIST video for static backgrounds. Compute PSNR, SSIM, and LPIPS against the reference, with paired bootstrap confidence intervals, comparing VIIS to the strongest baseline from Tables 2-3 and to the best new pipeline. If VIIS does not beat that baseline significantly on real, non-ISPT-degraded inputs, the claimed superiority on realistic severe low-light images is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the claim of outperforming prior methods on realistic severe low-light images, the evaluation must show gains on real data, not only on the training pretext. Section 9's fidelity experiment is circular: test inputs are generated by the exact ISPT augmentation (gamma=6, alpha=0.5, lambda=10, sigma=5) inside the training ranges (gamma 3-10, alpha 0.1-1, lambda 0-20, sigma 0-10). The model is therefore being tested on the same synthetic degradation it was trained to invert; the +3 PSNR and +0.07 SSIM margin in Table 8 measures inversion of the augmentation, not restoration of real severe low-light scenes. On actual MSRS and KAIST-MS night images (Tabs. 2-3), there is no ground truth: only NIQE, BRISQUE, SD, and EN on a subjectively selected subset (62 and 68 pairs), with no error bars or significance testing. NIQE and BRISQUE are no-reference and can be improved by increased contrast or colorfulness rather than fidelity, while SD and EN reward high-contrast or noisy outputs. The user study covers 20 images and 15 participants, which is not enough to establish state-of-the-art. Moreover, the newly designed baselines are pipelines of independently pretrained modules, not jointly optimized for the synthesis task, so beating them is weak evidence. Unless real paired data or a much stronger perceptual protocol is used, the paper's central quantitative claim is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new task, visible and infrared information synthesis (VIIS), which aims to jointly enhance severely low-light visible images and colorize infrared images into a single high-quality output. To overcome the absence of ground truth, the authors propose an information synthesis pretext task (ISPT) based on gamma/contrast/noise augmentation, and a sparse attention-based dual-modalities residual (SADMR) conditioning mechanism for a latent diffusion model. Experiments on MSRS and KAIST-MS, including no-reference metric comparisons, ablations, a user study, and a synthetic fidelity test, are presented to support the claim that the method outperforms state-of-the-art low-light enhancement, infrared colorization, and visible-infrared fusion methods as well as newly constructed pipeline baselines.","tokens_in":15485,"tokens_out":5313,"duration_ms":43613,"significance":"If the claimed performance is substantiated, VIIS would be a useful contribution to severe low-light imaging, combining intra-modal enhancement and inter-modal fusion in a single diffusion-based model. The paper offers a novel task formulation, a publicly available codebase, and extensive ablations of the conditioning mechanism and the pretext task components. The methodological novelty (SADMR) is clearly described. However, the current evaluation does not convincingly establish the central claim: the fidelity experiment is circular, and the real-image comparison relies on no-reference metrics without statistical validation. The contribution is therefore promising but not yet demonstrated to the standard required for the claimed state-of-the-art results.","major_comments":[{"comment":"The fidelity evaluation is circular: the synthetic test inputs are generated by the same ISPT augmentation (γ=6, α=0.5, λ=10, σ=5) used in training, and these parameters lie inside the training ranges given in Sec. 4.1 (γ∈[3,10], α∈[0.1,1], σ∈[0,10], λ∈[0,20]). Consequently the PSNR/SSIM gains (+3 dB and +0.07 in Table 8) may reflect the model's ability to invert its own training degradation rather than genuine generalization to real severe low-light conditions. The authors should provide non-circular evidence, e.g., testing with a different degradation model or with real paired low-light/high-light data.","section":"Sec. 9, Table 8"},{"comment":"The real-image evaluation relies on no-reference metrics (NIQE, BRISQUE, SD, EN) computed on small, subjectively selected subsets (62 and 68 pairs) without error bars or significance tests. NIQE and BRISQUE are known to be imperfect proxies for perceptual quality and can be improved by increased contrast or colorfulness; SD and EN reward high contrast and high entropy, which may favor noisy outputs. This evidence is not sufficient to support the claim that the method outperforms baselines on real severe low-light images.","section":"Sec. 4.2, Tables 2 and 3"},{"comment":"The user study uses only 20 images and 15 participants, and the voting percentages show large margins but no statistical analysis (e.g., confidence intervals or significance testing). The small sample size and the fact that most participants have computer vision knowledge limit the strength of the conclusions drawn from this study.","section":"Sec. 8, Table 6"},{"comment":"The newly designed baselines (EMD, PSMD, RMD, PDMD, UMD) are pipelines that chain independently pretrained enhancement, colorization, and fusion networks without any joint training for the VIIS objective. Beating these baselines is weak evidence of superiority over methods that are actually optimized for the combined task; the comparison against individual task-specific SOTA methods, while informative, does not directly address the joint VIIS setting.","section":"Sec. 4.1, Tables 1 and 7"},{"comment":"The authors acknowledge that their pretext task may weaken visible-specific information (e.g., texture or color unique to the visible image), but this trade-off is not quantified. Given that the abstract claims perceptual quality superiority, a quantitative analysis of this failure mode (e.g., how often and where visible-only cues are lost) would help the reader assess the practical applicability of the method.","section":"Sec. 5"}],"minor_comments":[{"comment":"The ablation table in the main text is labeled Table 4 but is referenced as 'Tab. 5' in the text; please fix the cross-references.","section":"Sec. 4.3"},{"comment":"The phrase 'data argumentation' should be 'data augmentation'.","section":"Sec. 9.1"},{"comment":"The Gaussian noise term is written as N(0, σ) and the Poisson noise as P(λ); clarify whether σ is the standard deviation or variance, and specify the parameterization and domain of λ.","section":"Eq. (7)"},{"comment":"The caption 'The overall of our model' should be 'Overview of our model'.","section":"Fig. 2"},{"comment":"Please state the sampling distribution for the augmentation parameters (e.g., uniform over the stated ranges) and whether the same random draw is used for both modalities; this is important for reproducibility.","section":"Sec. 4.1"},{"comment":"The paper would benefit from reporting inference runtime and model complexity, especially since diffusion models are typically computationally expensive.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-written and the proposed method is interesting, but the evaluation falls short of supporting the strong state-of-the-art claims. The circularity of the fidelity experiment in Sec. 9 is a serious issue that undermines the main quantitative evidence. I believe the paper can be revised with additional experiments on non-circular benchmarks or real paired data, along with a more rigorous real-image evaluation protocol. Given the scope of the revision, I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper defines a genuinely new task—joint visible-infrared enhancement and fusion under severe low light—and builds a latent diffusion model with a sparse deformable cross-attention conditioning mechanism to do it. That part is solid: the task framing is useful, the method is a sensible combination of known pieces, and the code is public. The authors also explicitly admit that visible-unique information may be weakened, which is more honest than most.\n\nThe problem is the evaluation. The main fidelity experiment (Sec. 9) is circular: the test inputs are generated with exactly the same ISPT augmentation used in training, with parameters inside the training range. So the +3 PSNR / +0.07 SSIM margin measures how well the model inverts its own augmentation, not how well it restores real severe low-light scenes. On real MSRS and KAIST-MS night images, there is no ground truth, so the paper falls back on NIQE/BRISQUE/SD/EN over a subjectively selected 62+68 pairs, with no error bars or significance testing. Those no-reference metrics can be gamed by contrast and colorfulness rather than fidelity. The user study (20 images, 15 participants) is too small to settle anything. The newly designed baselines are pipelines of independently pretrained modules rather than jointly optimized end-to-end models, so beating them is weak evidence.\n\nNone of this kills the idea. The task is real, the method is plausible, and the failure mode is common in this subfield—no paired real data for severe low light. But the paper's central claim of state-of-the-art is not supported by the current evidence.\n\nMy take: it deserves a serious referee, but with a clear request for major revision. The authors should either find real paired low-light visible-infrared data (simulated or captured) or present a much stronger no-reference evaluation with statistical testing and a larger user study. If they can show the ISPT augmentation transfers to real imagery, the paper would be a useful contribution. As it stands, I'd read it as a new task proposal with a promising method, not as a proven SOTA.\n\nRecommendation: send to peer review, but be explicit that the evaluation needs to be substantially strengthened.","headline":"A novel task and a plausible diffusion method, but the circular fidelity test and weak no-reference metrics don't support the state-of-the-art claim.","tokens_in":16050,"tokens_out":3547,"would_cite":true,"duration_ms":27413,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a new task, visible and infrared information synthesis (VIIS), in which a severely dark visible image and its paired infrared image are jointly enhanced and fused into a single high-quality color image.","keywords":["low-light image enhancement","infrared and visible image fusion","diffusion model","information synthesis","sparse cross-attention","pretext task","multi-modal conditioning","image restoration"],"falsifier":"A direct test is to collect a small set of real visible–infrared pairs captured in genuinely severe low light, with human-verified ground-truth content for objects hidden in darkness, and compare the model's output against the original visible image and the infrared image. If the model does not reliably reveal objects that are invisible in the visible input but present in the infrared input—or if its outputs on real data are markedly worse than on pseudo-low-light images from the same daytime scenes—the claim that the pretext task transfers would be falsified.","tokens_in":14950,"feed_emoji":"🌃","tokens_out":4211,"duration_ms":36605,"temperature":0.7,"pith_summary":"The paper proposes a new task, visible and infrared information synthesis (VIIS), in which a pair of severely dark visible and infrared images is turned into a single bright, colored, detail-rich image. Existing low-light enhancement methods fail where visible information is completely absent, and existing fusion methods only combine modalities without enhancing them. The authors argue that doing both at once—enhancing each modality and fusing them—solves the ill-posed problem, because the infrared image supplies the missing structure while the visible image supplies color and texture. To make training possible without ground truth, they design an augmentation-based pretext task that degrades daytime images into pseudo-low-light pairs. The paper reports that their diffusion-based model outperforms state-of-the-art low-light enhancement, infrared colorization, and fusion methods on standard datasets.","feed_headline":"Dark visible plus infrared yields one bright, colored image","feed_subtitle":"A diffusion model that enhances and fuses both modalities in one pass outperforms existing low-light and fusion methods.","key_machinery":"The key machinery is the sparse attention-based dual-modalities residual (SADMR) conditioning mechanism inside a latent diffusion model (LDM). The visible and infrared images are encoded by a pre-trained VAE, concatenated with the noise, and also encoded into multi-scale features that are injected in parallel into the encoder and decoder of the denoising U-Net. In place of simple addition or concatenation, a sparse cross-attention module based on deformable attention lets each element of the intermediate features attend only to sampled points around its reference location in both modality feature maps, enabling adaptive, iterative cross-modal interaction during denoising.","core_discovery":"The central claim is that simultaneous intra-modal enhancement and inter-modal fusion of visible and infrared images is achievable in a single diffusion-based model, and that this joint formulation outperforms both singular-modality enhancement and conventional fusion. The model, VIIS, is trained with an information synthesis pretext task: daytime visible images are degraded by gamma darkening, contrast reduction, and Gaussian and Poisson noise to create pseudo-low-light inputs, and the model must recover the original image using the corresponding infrared image as guidance. During denoising, a sparse attention-based dual-modalities residual (SADMR) conditioning mechanism lets intermediate features of the U-Net attend to each modality's features iteratively. On the MSRS and KAIST-MS datasets, the authors find the model produces higher perceptual quality and better information recovery than ten existing methods and four newly constructed pipeline baselines.","pith_inferences":["A natural extension the authors do not explore is whether the same pretext-task design transfers to other modality pairs, such as visible and depth or visible and event cameras, where one modality is robust to darkness.","The fidelity experiment only tests synthetic pseudo-low-light degradation; a real-world benchmark with manual annotations for objects hidden in darkness would be a stronger test of the transfer claim.","The sparse attention mechanism's sampling offsets could be visualized to check whether the model is genuinely using infrared structure in the darkest regions, which would make the fusion mechanism more interpretable."],"forward_implications":["A model trained on pseudo-low-light pairs can restore regions where the visible image carries no information, by borrowing structure from infrared and color from visible cues.","The same diffusion backbone with the SADMR conditioning mechanism can be applied to other multi-modal synthesis tasks that require iterative cross-modal interaction.","Because the pretext task requires no paired ground truth, the training procedure can be adapted to any aligned visible–infrared dataset with daytime images.","The reported results imply that joint enhancement and fusion is preferable to chaining separate enhancement, colorization, and fusion modules."],"supporting_citations":[{"why":"Supplies the latent diffusion model framework that VIIS builds on for efficient image generation.","marker":"[39]"},{"why":"Provides the deformable attention operation that the sparse cross-attention module adapts for multi-modal feature injection.","marker":"[61]"},{"why":"Provides the MSRS aligned visible–infrared dataset used for training and testing.","marker":"[47]"},{"why":"Provides the KAIST-MS aligned visible–infrared dataset used for training and testing.","marker":"[17]"},{"why":"Introduces denoising diffusion probabilistic models, the generative foundation the method's diffusion process relies on.","marker":"[11]"},{"why":"The self-supervised decomposition fusion model used both as a comparison baseline and as the fusion component of the constructed pipeline baselines.","marker":"[29]"}],"fun_headline_variants":["One diffusion pass fuses visible and infrared for dark scenes","VIIS: one diffusion model brightens and fuses visible+IR","Enhance and fuse visible+IR in a single diffusion step","Dark scenes: one model does both enhancement and fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic degradation used in the pretext task—intense gamma darkening, contrast reduction, and Gaussian and Poisson noise—sufficiently mimics real severe low-light conditions, so that a model trained on pseudo-low-light daytime images transfers to real nighttime and subterranean imagery.","fun_headline_variants_meta":{"raw":{"variants":["One diffusion pass fuses visible and infrared for dark scenes","VIIS: one diffusion model brightens and fuses visible+IR","Enhance and fuse visible+IR in a single diffusion step","Dark scenes: one model does both enhancement and fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00061,"raw_usage":{"total_tokens":2837,"prompt_tokens":940,"completion_tokens":1897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1838}},"tokens_in":556,"tokens_out":1897,"duration_ms":12972,"temperature":1.0,"reasoning_tokens":1838,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:54:50.027625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to collect a small set of real visible–infrared pairs captured in genuinely severe low light, with human-verified ground-truth content for objects hidden in darkness, and compare the model's output against the original visible image and the infrared image. If the model does not reliably reveal objects that are invisible in the visible input but present in the infrared input—or if its outputs on real data are markedly worse than on pseudo-low-light images from the same daytime scenes—the claim that the pretext task transfers would be falsified.","supporting_citations":[{"cited_title":"Piafusion: A progressive infrared and visible im- age fusion network based on illumination aware.Information Fusion, 83:79–92, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the MSRS aligned visible–infrared dataset used for training and testing."},{"cited_title":"High-resolution image syn- thesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion model framework that VIIS builds on for efficient image generation."},{"cited_title":"Deformable DETR: Deformable Transform- ers for End-to-End Object Detection","cited_arxiv_id":null,"evidence_quote":"Provides the deformable attention operation that the sparse cross-attention module adapts for multi-modal feature injection."},{"cited_title":"Multispectral pedestrian detection: Benchmark dataset and baseline","cited_arxiv_id":null,"evidence_quote":"Provides the KAIST-MS aligned visible–infrared dataset used for training and testing."},{"cited_title":"Denoising diffu- sion probabilistic models","cited_arxiv_id":null,"evidence_quote":"Introduces denoising diffusion probabilistic models, the generative foundation the method's diffusion process relies on."},{"cited_title":"Fusion from decomposition: A self-supervised decomposi- tion approach for image fusion","cited_arxiv_id":null,"evidence_quote":"The self-supervised decomposition fusion model used both as a comparison baseline and as the fusion component of the constructed pipeline baselines."}],"review_version":1}