{"id":"4b97965b-45c3-42b0-b839-f9531e94f112","arxiv_id":"2505.04526","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A joint enhancement-and-fusion network, DFVO, produces clearer and more evenly illuminated visible-infrared fusion results in dark scenes and reports state-of-the-art metrics on the LLVIP benchmark.","lead":"This paper presents a neural network that simultaneously brightens dark visible images and fuses them with infrared images, so nighttime scenes become clearer for driving systems. It reports higher fusion scores and better pedestrian detection than prior two-stage enhancement-plus-fusion methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PSNR claim is unverifiable and partly trained into the model: Eq. (20) directly penalizes L2 distance from fused to infrared, so if the reported PSNR uses the infrared reference (or includes it), the headline margin is substantially by construction.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the PSNR reference is undefined, and the content loss Lcont (Eq. 20) directly minimizes L2 distance between the fused image and the infrared image (and the enhanced visible image). Since PSNR is a monotone transform of MSE, any evaluation that includes the infrared image as reference is partially optimizing the reported metric. This is not a superficial reporting gap: it affects the central claim that DFVO is quantitatively state-of-the-art. The concern is concrete and checkable because the code is public and the ablation in Table IV shows PSNR strongly depends on Lcont. I do not escalate to rejection because the paper also reports no-reference or less circular metrics (SF, AG, CC) where DFVO leads, and the code makes the metric protocol inspectable. The appropriate verdict remains conditional, pending clarification and recomputation of the PSNR/MSE protocol. I agree with the reader's assessment; no additional load-bearing concern supersedes this one.","tokens_in":16960,"tokens_out":4861,"duration_ms":50992,"concrete_test":"Run the released code's metric script on the same 50 LLVIP validation pairs and print the reference tensor used for PSNR/MSE; then recompute Table II's PSNR column under three explicit references: (a) fused vs infrared, (b) fused vs enhanced visible, (c) average over fused-vs-visible and fused-vs-infrared. Check whether DFVO's 63.258 dB lead over DIVFusion persists under (b) or (c); if it does not, the headline metric is an artifact of the chosen reference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing empirical claim is that DFVO achieves the best quantitative fusion quality, headlined by 63.258 dB PSNR on LLVIP (Abstract, Table II). This number cannot be evaluated from the paper because Section IV-A3 lists PSNR and MSE without defining the reference image. In fusion there is no ground-truth fused image, so PSNR must be computed against some input or derived image; the choice determines the ranking. The problem is not merely a missing definition: the content loss Lcont (Eq. 20) contains omega_ir * ||If - Iir||_2 and omega_R * ||If - Ihat_R||_2. PSNR is a monotone decreasing function of MSE. If the reported PSNR is computed against Iir, or against an average that includes Iir, DFVO is evaluated on a metric it was explicitly trained to minimize. That would explain the unusually large margin over DIVFusion (63.258 vs 57.891) and would make the headline 'best performance' circular. No passage in the paper specifies the metric reference, and the ablation table (Table IV) shows PSNR tracks the presence of Lcont (dropping to 60.321 when Lcont is removed), which is consistent with the metric being directly optimized. Additionally, PSNR = 63.258 dB with MSE = 0.032 is internally consistent only on a 0-255 scale, whereas Section IV-A1 says training images are normalized to 0-1; this reinforces that the evaluation pipeline is underspecified. The empirical superiority claim therefore rests on an unstated and potentially self-fulfilling evaluation choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DFVO, a jointly trained network for visible-infrared image fusion under low-illumination conditions, replacing the conventional two-stage enhancement-then-fusion pipeline with a single cascaded multi-task model. The architecture combines a latent-common feature extractor with a details-extraction module (DEM) for high-frequency information and a hyper cross-attention module (HCAM) for low-frequency structural information, alongside task-specific losses for infrared reconstruction, illumination disentanglement, and fusion. Experiments are reported on LLVIP with additional generalization tests on MSRS, SMOD, and KAIST, including qualitative comparisons, nine quantitative metrics, ablation studies, a perceptual MOS study, and detection experiments with YOLOv5. The code is publicly released.","tokens_in":17403,"tokens_out":7289,"duration_ms":68604,"significance":"If the reported results hold, DFVO is a meaningful contribution to low-light infrared-visible fusion: a single-stage formulation that avoids error accumulation across separate enhancement and fusion stages is well motivated, and the design of DEM and HCAM is coherent. The paper provides code, extensive ablations, downstream detection experiments, and cross-dataset generalization, which are concrete strengths. However, the headline quantitative claims are currently undermined by an undefined evaluation reference and by the direct overlap between the training losses and the reported metrics, so the significance cannot be fully assessed from the manuscript as written.","major_comments":[{"comment":"The reference image for PSNR and MSE is never defined. In image fusion there is no ground-truth fused image, so PSNR/MSE must be computed against some reference (for example, a source image, an enhanced image, or a constructed ideal); the paper lists PSNR and MSE among the metrics but does not state this choice. This is load-bearing because the Abstract and Section IV-B3 claim the best LLVIP PSNR of 63.258 dB and MSE of 0.032. Moreover, Section IV-A1 states that training images are normalized to [0,1], yet PSNR=63.258 dB with MSE=0.032 is numerically consistent only if MSE is computed on a 0-255 scale. The authors must specify the reference image, the value range, and the exact computation for PSNR/MSE and for each other metric whose reference is ambiguous.","section":"Section IV-A3, Tables I-II"},{"comment":"The reported quantitative gains overlap with the training losses in a way that makes the ranking partly by construction. Lcont (Eq. 20) is a weighted L2 penalty between the fused image and the corrected-enhanced and infrared images, which directly minimizes MSE and hence maximizes PSNR if either of those images is the metric reference; Lstr (Eq. 23) penalizes differences in gradient magnitude, directly improving spatial-frequency and average-gradient scores; and Lcos (Eq. 24) maximizes cosine similarity to the enhanced visible image, which would directly improve correlation-based metrics if the reference is the visible/enhanced image. Table IV is consistent with this concern: PSNR drops from 63.268 to 60.321 when Lcont is removed, and SF/AG drop sharply when Lstr is removed. To support the \"best performance\" claim, the authors should report metrics whose references are explicitly defined and that are not simply the training objectives, or show that the advantage persists under no-reference fusion metrics and on downstream tasks.","section":"Eqs. (20), (23), (24) and Table IV"},{"comment":"The quantitative ranking is based on point estimates on a fixed test subset. The paper trains on 240 image pairs and reports validation results on 50 \"typical\" pairs from LLVIP (Section IV-A1); Tables I and II do not state sample size, variance, or significance. Given that some margins over baselines are small (for example, CC 0.724 vs 0.718 in Table I), the \"best performance\" claims need either the full test set, multiple splits, or statistical tests to rule out selection effects. The same applies to the generalization tables (Table III), which report results on 183, 331, and 25 pairs without error bars.","section":"Section IV-A1, Tables I-III"}],"minor_comments":[{"comment":"The notation in Eq. (4) is difficult to parse: the summation indices n do not appear in the summands (e.g., \\sum_{n=1}^{C/2} \\Theta_t^I), and the channel-splitting notation is unclear; please rewrite the iteration with clearly indexed channel groups.","section":"Section III-B1, Eq. (4)"},{"comment":"The caption reports a \"maximum mAP of 0.88 for vehicle recognition,\" but Table VI lists mAP@.5=0.861 and mAP@.5:.95=0.479 for pedestrian detection; please reconcile the numbers and clarify which detection task and metric are being reported.","section":"Fig. 9 caption and Table VI"},{"comment":"The PSNR for the full model is 63.258 dB in Table II but 63.268 dB in Table IV; please unify the numbers.","section":"Tables II and IV"},{"comment":"The text says DFVO \"leads in SF, PSNR, and MSR\"; MSR is not defined or used elsewhere and appears to be a typo for MSE.","section":"Section IV-C2"},{"comment":"The iteration index n in Eq. (25) is defined only as \"the n-th iteration,\" while training is reported in epochs (Section IV-A2); please clarify whether n counts epochs, iterations within an epoch, or total optimizer steps, since \\zeta_3 takes very different values in each case.","section":"Section III-C, Eq. (25)"},{"comment":"The remaining metrics (EN, SF, MI, SD, VIF, AG, CC) are named but not defined; please provide definitions or citations, since the paper's ranking claims depend on them.","section":"Section IV-A3"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the headline quantitative claim may be an artifact of an undefined metric reference and of training losses that directly optimize the reported metrics. If the authors clarify the evaluation protocol and add more independent metrics, the paper could become acceptable; otherwise the empirical contribution is not verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on DFVO (2505.04526). The genuinely new thing is the joint one-stage design: instead of the usual two-stage enhance-then-fuse pipeline, they train illumination disentanglement, infrared reconstruction, and fusion together in one network, with loss functions appropriate to each task. The architecture is coherent: invertible residual blocks extract high-frequency detail, and the hyper cross-attention uses infrared gradients as keys, which makes sense for dark scenes. The code is public and the ablations are honest, including an acknowledged limitation section.\n\nThe real soft spot is exactly what your reader flagged. PSNR and MSE are reported without defining the reference image. For fusion there is no ground-truth fused image, so some choice of reference is required. Eq. (20) is an L2 content loss against the enhanced infrared and visible images, so if PSNR is computed against either of those, the headline metric is partially trained into the model. The ablation table strengthens the suspicion: removing Lcont drops PSNR from 63.268 to 60.321. I looked for a 'reference' in Section IV-A3 and there is none. That is a genuine verifiability problem, not a nitpick.\n\nIt is not the whole story. The detection results on LLVIP (Table VI) show real improvements in precision, recall, and mAP, which is an extrinsic evaluation less affected by the circularity. The human study is tiny (8 volunteers) and not more than suggestive. The novelty overlap with LENFusion is real, but the authors cite it and differentiate by architecture, which is acceptable.\n\nMy bottom line: this paper deserves a serious referee, not a desk reject. The architecture and code are worth engaging with, and the joint training idea is a legitimate line of work. But the revision must state the PSNR reference, recompute numbers against specified references, and reconcile the 0-1 normalization in training with the apparent 0-255 metric scale. Without that, the central quantitative claim is unverifiable. If the authors fix that, I'd be comfortable citing this as a representative one-stage low-light fusion method.","headline":"A worthwhile one-stage low-light fusion architecture, but the headline PSNR is unverifiable without a stated reference — and possibly trained into the model.","tokens_in":17894,"tokens_out":2360,"would_cite":true,"duration_ms":22500,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A joint cascaded multi-task network performs low-light enhancement and visible-infrared fusion in one pass, beating two-stage baselines on night fusion metrics.","keywords":["visible-infrared image fusion","low-light enhancement","cascaded multi-task learning","illumination disentanglement","hyper cross-attention","latent feature extraction","LLVIP dataset","nighttime fusion"],"falsifier":"Recompute PSNR, MSE, and CC with the visible image, the infrared image, and both sources (or a no-reference metric) as references on the same test set; if DFVO's margin over two-stage baselines disappears when the reference is not the infrared image, the headline metric is an artifact of the loss design.","tokens_in":16802,"feed_emoji":"🌙","tokens_out":4590,"duration_ms":40793,"temperature":0.7,"pith_summary":"The paper claims that the usual two-stage recipe for darkness-free infrared-visible fusion — first brighten the visible image, then fuse it with the infrared image — loses information at each transfer and can be replaced by a single network trained on three tasks at once: infrared reconstruction, illumination disentanglement, and fusion. If correct, DFVO gives clearer, better-lit fused images in night driving scenes with less noise and over-exposure than either conventional fusion or two-stage enhancement-then-fusion baselines, and it improves downstream pedestrian and vehicle detection. The headline evidence is 63.258 dB PSNR and 0.724 correlation with source images on the LLVIP dataset, plus a preference score of 62.5 percent in a human perception study.","feed_headline":"One network fuses and brightens night images in a single pass","feed_subtitle":"DFVO replaces two-stage enhancement-then-fusion pipelines, cutting information loss and lifting LLVIP PSNR to 63.258 dB.","key_machinery":"The load-bearing mechanism is the latent-common feature extractor (LCFE), which combines a details-extraction module (DEM) and a hyper cross-attention module (HCAM) to obtain high-frequency semantic and low-frequency texture features from both modalities. Three task-specific heads share these latent features: infrared image reconstruction, illumination disentanglement (via a separate encoder-decoder that estimates the illumination image under Retinex theory), and the main image fusion task, which concatenates visible and infrared latent features. Task-specific losses include a Retinex decomposition loss, a histogram-equalization prior loss, a content loss with adaptive weights that directly minimizes L2 distance between the fused image and the corrected-enhanced/infrared images, a structural loss, and an RGB color consistency loss; a growing task weight for the fusion loss balances the cascade during training.","core_discovery":"DFVO's central claim is that cascaded multi-task learning, in which shared latent features feed three task-specific heads trained simultaneously, outperforms the two-stage cascaded training used by existing darkness-free fusion methods. The network extracts common latent features from visible and infrared inputs, reconstructs the infrared image, decomposes the visible image into reflectance and illumination under Retinex theory, and fuses the features in one holistic pass. The authors report state-of-the-art quantitative results on the LLVIP dataset (PSNR 63.258 dB, MSE 0.032, SF 0.082, AG 6.783, CC 0.724), best perceptual preference in a human ranking, and higher YOLOv5 pedestrian precision and recall on fused images, which they attribute to avoiding the information entropy loss of hierarchical data transmission.","pith_inferences":["The paper never defines what reference image the PSNR/MSE numbers are computed against; given that the content loss directly pushes the fused image toward the infrared image in L2, the headline 63.258 dB may largely measure similarity to the infrared input rather than fusion quality.","The same cascaded-multi-task principle could be tested on other modality pairs, such as RGB-depth or RGB-event fusion, where a two-stage restoration-then-fusion pipeline would suffer the same information-loss problem.","A straightforward ablation would hold the loss functions fixed and swap the joint training schedule for a sequentially frozen two-stage schedule with identical modules; if the gap persists, the 'all at once' claim is about optimization, not architecture.","The HCAM's quadratic attention complexity is acknowledged as a limitation; a linear-attention or windowed variant of the same cross-modal attention would be the natural test of whether the fusion gain survives a lightweight deployment."],"forward_implications":["Night-vision fusion systems for autonomous driving could skip a dedicated low-light enhancement preprocessing stage, cutting runtime and avoiding error accumulation between stages.","Because the same shared latent features feed all three tasks, the network is trained end-to-end with no separate enhancement dataset or paired ground-truth illumination, making it applicable where only visible-infrared pairs are available.","The fused outputs improve object detection and pedestrian recognition in the paper's experiments, suggesting the all-at-once objective transfers to high-level vision tasks better than two-stage pipelines.","A single-stage network trained jointly on reconstruction, enhancement, and fusion generalizes across datasets (MSRS, SMOD, KAIST) without retraining, indicating the approach is not overfit to one data distribution."],"supporting_citations":[{"why":"Supplies the single-scale two-stage darkness-free method that DFVO contrasts with and reports beating; its noise and over-exposure motivate the all-at-once design.","marker":"[17]"},{"why":"Provides the two-stage low-light enhancement and fusion baseline used in quantitative and qualitative comparisons.","marker":"[20]"},{"why":"Represents the multi-scale two-stage data transmission approach that DFVO argues still loses information between stages.","marker":"[21]"},{"why":"Classic CNN fusion encoder-decoder used both as a direct baseline and inside two-stage comparisons.","marker":"[26]"},{"why":"Unsupervised fusion baseline that defines the feature-extraction-fusion-reconstruction paradigm DFVO compares against.","marker":"[27]"},{"why":"Supplies the Retinex illumination decomposition used by the illumination disentanglement task and its loss.","marker":"[45]"},{"why":"Source of the illumination smoothness and mutual-consistency losses, and a two-stage enhancement baseline.","marker":"[47]"},{"why":"The visible-infrared nighttime dataset on which the model is trained and the headline metrics are measured.","marker":"[52]"}],"fun_headline_variants":["One-pass darkness-free fusion beats two-stage pipelines","Single network fuses IR and visible, lifts LLVIP PSNR to 63.258","All-at-once fusion eliminates two-stage info loss, best on LLVIP","Darkness-free fusion: one network, both tasks, top results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative superiority rests on an unstated choice of reference image for PSNR and MSE; if the reference is the infrared image, the large PSNR partly follows from a loss term that pushes the fused image toward the infrared image.","fun_headline_variants_meta":{"raw":{"variants":["One-pass darkness-free fusion beats two-stage pipelines","Single network fuses IR and visible, lifts LLVIP PSNR to 63.258","All-at-once fusion eliminates two-stage info loss, best on LLVIP","Darkness-free fusion: one network, both tasks, top results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001203,"raw_usage":{"total_tokens":4995,"prompt_tokens":1022,"completion_tokens":3973,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":3894}},"tokens_in":638,"tokens_out":3973,"duration_ms":31859,"temperature":1.0,"reasoning_tokens":3894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:26:06.068885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute PSNR, MSE, and CC with the visible image, the infrared image, and both sources (or a no-reference metric) as references on the same test set; if DFVO's margin over two-stage baselines disappears when the reference is not the infrared image, the headline metric is an artifact of the loss design.","supporting_citations":[{"cited_title":"Divfusion: Darkness-free infrared and visible image fusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the single-scale two-stage darkness-free method that DFVO contrasts with and reports beating; its noise and over-exposure motivate the all-at-once design."},{"cited_title":"Lenfusion: A joint low- light enhancement and fusion network for nighttime infrared and visible image fusion,","cited_arxiv_id":null,"evidence_quote":"Provides the two-stage low-light enhancement and fusion baseline used in quantitative and qualitative comparisons."},{"cited_title":"Enlighten fusion multiscale network for infrared and visible image fusion in dark environments,","cited_arxiv_id":null,"evidence_quote":"Represents the multi-scale two-stage data transmission approach that DFVO argues still loses information between stages."},{"cited_title":"U2fusion: A unified unsupervised image fusion network,","cited_arxiv_id":null,"evidence_quote":"Unsupervised fusion baseline that defines the feature-extraction-fusion-reconstruction paradigm DFVO compares against."}],"review_version":1}