{"id":"2660fc91-3a47-46e0-b6c2-e38240896eef","arxiv_id":"2505.24705","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A cross-attention RGB-thermal network reports higher PSNR and SSIM than prior low-light enhancers on LLVIP and a new 50-scene V-TIEE dataset.","lead":"A new deep network, RT-X Net, combines ordinary color images with thermal camera images to make nighttime photos brighter and sharper. The authors also introduce V-TIEE, a 50-scene visible-thermal dataset, and report higher quality than prior low-light enhancement methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LLVIP results may only show inversion of the same synthetic degradation used in training; real-world V-TIEE evidence is too thin to establish that thermal cross-attention generalizes.","rationale":"The paper is credible in intent: thermal guidance is a plausible source of nighttime structure, the cross-attention ablation is helpful, and the V-TIEE dataset is a genuine resource. However, the strongest quantitative support for 'outperforms state-of-the-art' rests on a synthetic degradation model that is also used for training. This is an empirical-verification gap rather than an internal inconsistency: the architecture could still work on real data. The reader's weakest assumption focused on V-TIEE alignment and the synthetic nature of LLVIP; I agree with both but identify the train/test degradation match on LLVIP as the more load-bearing issue, because without a real-data check the headline numbers may be a distribution-match artifact. A held-out real-data comparison of RT-X Net against its RGB-only ablation would settle this. Until then, CONDITIONAL remains the appropriate verdict.","tokens_in":6322,"tokens_out":5017,"duration_ms":58275,"concrete_test":"Use V-TIEE's multiple exposures per scene: hold out 10 of the 50 scenes, train RT-X Net and its RGB-only self-attention ablation on the low→high exposure pairs of the remaining 40 scenes using the homography-aligned thermal images, and evaluate PSNR/SSIM/LPIPS on the held-out scenes with bootstrap 95% confidence intervals over scenes. If the cross-attention minus RGB-only gain is not significantly positive on real data, the synthetic LLVIP results in Table 1 are not sufficient to support the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that cross-attention with thermal images improves low-light enhancement (abstract; Table 1). The decisive quantitative evidence is the LLVIP column, but Section 4.2.1 states: 'Low-exposure conditions are simulated by reducing exposure (×5–20) and adding noise [22].' Section 4.2.2 trains on exactly these simulated low-light pairs with an MAE loss. Thus the reported PSNR/SSIM measure how well each method inverts the same synthetic corruption family seen during training. A 1.16 dB gain over Retinexformer can therefore reflect better fit to that synthetic model rather than a genuine benefit of thermal structure in real night scenes. The V-TIEE rows are meant to supply real-world evidence, but they use only 50 scenes, no error bars, and the visible/thermal alignment is a homography between sensors with different resolutions and FOVs (1440×1080, 34° vs 640×512, 24° HFOV, Section 4.1). Residual parallax or homography error would mix non-corresponding content, making the fusion gains uninterpretable. The closest thermal-guided baseline (Cao et al. [20]) is excluded for lack of code, so the comparison does not isolate the modality benefit. The ablation in Table 2 supports cross-attention over concatenation, but only on the same synthetic LLVIP protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RT-X Net, a transformer-based network that fuses RGB and thermal images for low-light image enhancement. The architecture extracts illumination-guided self-attention features from both modalities and fuses them with a multi-head cross-attention module, followed by a reconstruction network. The authors introduce a real-world Visible-Thermal Image Enhancement Evaluation (V-TIEE) dataset with 50 co-located visible and thermal scenes, and they evaluate RT-X Net on a subset of the LLVIP dataset and on V-TIEE. Quantitative results in Table 1 report PSNR 27.75 dB / SSIM 0.85 on LLVIP, outperforming Retinexformer by 1.16 dB, and LPIPS 0.12 / SSIM 0.71 on V-TIEE. The ablation in Table 2 shows that cross-attention fusion outperforms both RGB-only self-attention and thermal channel concatenation. The central claim is that cross-modal attention to thermal structure yields a meaningful, though modest, gain for nighttime enhancement.","tokens_in":6585,"tokens_out":2747,"duration_ms":33206,"significance":"If the reported results are robust, the paper makes a useful contribution to low-light image enhancement by demonstrating that thermal imagery can be effectively integrated through cross-attention. The proposed architecture is lightweight (0.67M parameters) and the ablation supports the design choice of cross-attention over concatenation. The release of code and the V-TIEE dataset is also a concrete contribution to the community. However, the quantitative evidence for the central claim is weakened by the synthetic nature of the main LLVIP evaluation, the small size of the real-world V-TIEE set, the lack of error bars, and the absence of the most relevant thermal-guided baseline. These issues do not invalidate the approach, but they need to be addressed before the claim that RT-X Net outperforms state-of-the-art methods is fully convincing.","major_comments":[{"comment":"The LLVIP evaluation appears to measure how well each method inverts the same synthetic low-light degradation family used in training. Section 4.2.1 states that low-exposure conditions are simulated by reducing exposure (×5–20) and adding noise, and Section 4.2.2 states that the model is trained on the same simulated low-light pairs with an MAE loss. The reported PSNR/SSIM gains, including the 1.16 dB improvement over Retinexformer, could therefore reflect better fitting to this specific synthetic corruption model rather than a genuine benefit of thermal structure in real night scenes. Please provide evidence that the improvement persists under a different degradation protocol, on real low-light RGB-thermal pairs with ground truth, or at least with a held-out degradation setting not seen in training.","section":"§4.2.1, §4.2.2, Table 1"},{"comment":"The real-world evidence on V-TIEE is too thin to carry the generalization claim. The dataset contains only 50 scenes, no error bars or per-scene statistics are reported, and the visible and thermal sensors have different resolutions (1440×1080 vs. 640×512) and fields of view (34° vs. 24° HFOV), with alignment refined by a homography. Residual parallax or homography error would mix content from non-corresponding spatial locations, which is especially problematic for a cross-attention mechanism that assumes corresponding positions across modalities. Please report per-scene or bootstrap error bars, describe the alignment accuracy quantitatively, and show that the observed gains are robust to alignment perturbation.","section":"§4.1, §4.4, V-TIEE results"},{"comment":"The most relevant thermal-guided baseline, Cao et al. [20], is not compared because its code is unavailable. This baseline directly targets thermal-guided low-light visible image enhancement, so its absence means the comparison does not isolate the benefit of the proposed cross-attention fusion mechanism. Please include a re-implementation of [20] on the same protocol, or provide a clear quantitative and qualitative comparison that justifies why the omission does not affect the claimed superiority.","section":"§2.2, §4.3, Table 1"},{"comment":"The ablation study supports the cross-attention module, but it is conducted solely on the synthetic LLVIP protocol. The 0.6 dB gain over thermal channel concatenation may not transfer to real low-light conditions, where alignment errors and domain shift are present. Please report the same ablation on a real-world evaluation set, even if only with perceptual metrics, to demonstrate that the design choice is beneficial in the intended application setting.","section":"§4.5, Table 2"}],"minor_comments":[{"comment":"The title line in the paper header contains a typo: 'CROSS A TTENTION' should read 'CROSS ATTENTION'.","section":"Title and header"},{"comment":"Equation (2) defines the illumination estimator as a function of a single input I, but the text says it predicts from both RGB and thermal images. Please clarify whether the illumination estimator is shared across modalities or whether two separate estimators are used.","section":"§3.1, Eq. (2)"},{"comment":"The notation SelfAttn(xl, Fillum) suggests that illumination features are an input to the self-attention operator, but the text only says they are used as a learnable re-weighting term on the values. Please specify the exact integration mechanism in the equation or the surrounding text.","section":"§3.2, Eq. (3)"},{"comment":"The cross-attention module is written as MCA(xa_RGB ∥ xa_Therm), but the text says 'queries are derived from one modality, while keys and values come from the other.' With a concatenated input it is unclear which projection produces queries and keys from which branch. Please define the query, key, and value projections precisely.","section":"§3.3, Eq. (5)"},{"comment":"The PCA step after cross-attention is mentioned but not described; please state how the PCA is computed, whether it is learned or fixed, and what dimensionality is used.","section":"§3.3"},{"comment":"Table 1 reports a single value per metric without error bars. Given that the LLVIP test set has only 45 images and V-TIEE has 50 scenes, standard deviations or confidence intervals are needed to assess whether the reported differences are statistically meaningful.","section":"§4.2.1, Table 1"},{"comment":"The noise addition is attributed to Hasinoff et al. [22], but the exact noise model and parameters are not specified. Please describe how the simulated low-light images are generated so the experiments are reproducible.","section":"§4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is plausible and the ablation is internally consistent, but the main quantitative evidence is currently based on a synthetic degradation protocol that may favor the proposed method, and the real-world V-TIEE evaluation is too small and too coarsely reported to support the generalization claim. A revised version with real paired data or a more thorough real-world validation, error bars, and a comparison with the thermal-guided baseline Cao et al. would materially strengthen the paper. I would also ask the editor to verify that the GitHub repository and supplementary links are accessible and match the paper's description."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nPunchline: RT-X Net is a sensible, well-scoped empirical contribution. It adapts cross-attention to fuse RGB and thermal images for low-light enhancement, and it introduces V-TIEE, a small real-world dataset. The central claim—that thermal cross-attention helps—is plausible and supported by the paper's own ablation, but the headline numbers are weaker than they look because the main benchmark (LLVIP) is synthetic and small.\n\nWhat's new: To my knowledge, no prior work uses cross-attention for RGB-thermal fusion in low-light enhancement. The V-TIEE dataset, 50 co-located visible/thermal pairs with varied gain and exposure, fills a real gap. The paper also reports generalization from LLVIP to V-TIEE without fine-tuning, which is a good real-world check.\n\nWhat works: The ablation in Table 2 is the most useful evidence. Cross-attention beats channel concatenation by 0.6 dB, a meaningful margin for a design choice. The method is simple, light (0.67M parameters), and clearly described. The authors are honest about why [20] (Cao et al., the closest thermal-guided baseline) is absent: no public code. That is an understandable limitation, though it means the comparison does not isolate the benefit of thermal over visible alone.\n\nSoft spots: The LLVIP evaluation is the main quantitative claim, but the test set is 45 simulated images. Section 4.2.1 says low-exposure inputs are generated by reducing exposure ×5–20 and adding noise, and Section 4.2.2 trains on exactly the same corruption family. So the reported PSNR/SSIM largely measure how well each method inverts that synthetic degradation. That doesn't invalidate the comparison—all baselines see the same inputs—but a 1.16 dB gain over Retinexformer could be an artifact of fitting the simulation rather than a genuine advantage in real night scenes. The real-world V-TIEE evidence is too thin on its own: 50 scenes, no error bars, and alignment is homography between cameras with different resolutions and FOVs (34° vs 24°), leaving residual parallax risk. The paper would be much stronger with error bars or statistical tests, a few more real scenes, and ideally a thermal-plus-bright-reference protocol.\n\nRecommendation: This deserves a serious referee. It's a new application and a useful dataset, and the ablation is honest. But a referee should push for tighter empirical verification before acceptance. I'd review it, and I'd tell the authors to add uncertainty quantification and to compare against Cao et al. with a reimplementation if code truly is unavailable.","headline":"RT-X Net is a plausible, modest empirical contribution—new thermal cross-attention for low-light enhancement plus a fresh dataset—but its main LLVIP numbers rest on a small synthetic benchmark, so the real-world benefit needs firmer evidence.","tokens_in":7117,"tokens_out":2075,"would_cite":false,"duration_ms":22401,"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":"RT-X Net fuses RGB and thermal images with cross-attention, reporting state-of-the-art low-light enhancement on LLVIP and the new V-TIEE benchmark.","keywords":["low-light image enhancement","RGB-thermal fusion","cross-attention transformer","thermal imaging","image restoration","LLVIP dataset","V-TIEE dataset"],"falsifier":"Take RT-X Net trained on LLVIP and evaluate it on V-TIEE with the thermal image deliberately shifted by 1, 3, and 5 pixels relative to the visible image. If the PSNR or LPIPS advantage over the RGB-only ablation shrinks or reverses with small shifts, the method's real-world benefit depends on near-perfect alignment rather than on thermal structure per se.","tokens_in":6122,"feed_emoji":"🌙","tokens_out":5955,"duration_ms":65387,"temperature":0.7,"pith_summary":"Low-light image enhancement usually works on RGB alone, but at night the visible image is dominated by noise and glare while a thermal camera still sees clean structure. The paper tries to establish that a transformer that fuses RGB and thermal images through cross-attention can beat RGB-only state-of-the-art methods, and that the fusion mechanism matters more than simply having the thermal data. If this claim is right, it gives surveillance, automotive, and robotics systems with thermal sensors a concrete way to produce sharper, better-exposed visible images at night, and it provides a new benchmark dataset, V-TIEE, for testing such methods on real nighttime scenes.","feed_headline":"Thermal cross-attention lifts low-light image quality","feed_subtitle":"RT-X Net fuses RGB with thermal structure, beating Retinexformer by 1.16 dB PSNR on LLVIP.","key_machinery":"The load-bearing component is the multi-head cross-attention module (MCA), which fuses the illumination-guided self-attention features of the RGB and thermal branches. In MCA, queries are taken from one modality while keys and values come from the other, so the thermal stream can reweight and refine the RGB feature map; the module is preceded by a Retinex-based illumination estimator that supplies an illumination map and illumination features used to reweight attention values. A PCA step reduces the fused channel dimensionality, and the enhanced image is produced by an MLP reconstruction network trained with mean absolute error.","core_discovery":"RT-X Net is a transformer for low-light image enhancement that fuses a visible RGB image with a co-registered thermal image. The paper's central claim is that thermal imagery, which records scene structure independently of illumination, supplies the texture and edge cues that RGB-only enhancers miss, and that a cross-attention module is the mechanism that lets thermal features guide the RGB representation. On the LLVIP dataset the paper reports a peak signal-to-noise ratio of 27.75 dB and SSIM of 0.85, against 26.59 dB and 0.79 for Retinexformer, and on the newly introduced V-TIEE real-world dataset it reports LPIPS of 0.12 and SSIM of 0.71, also the best among the compared methods. Ablation experiments show that removing the cross-attention and using RGB only drops PSNR by more than 1.3 dB, while simply concatenating the thermal channel yields a smaller gain.","pith_inferences":["The LLVIP evaluation darkens images by simulated exposure reduction and added noise; real nighttime glare and sensor noise may make the thermal advantage larger than the 1.16 dB PSNR gain reported here.","Because the gain over Retinexformer is modest, the practical value of RT-X Net depends on the platform already having a co-located thermal sensor; for RGB-only systems the method offers no benefit.","A direct experiment shifting or blurring the thermal input would show whether the improvement comes from high-frequency thermal edges or from coarse brightness guidance, and whether the cross-attention is robust to misalignment.","The V-TIEE benchmark could double as an alignment testbed: reporting PSNR and LPIPS under deliberately misregistered thermal inputs would quantify how much of the real-world gain relies on the homography refinement."],"forward_implications":["Systems that already carry thermal cameras, such as autonomous vehicles, drones, and surveillance rigs, can use cross-attention fusion to produce visible-light images that are sharper and better exposed than RGB-only enhancers.","The V-TIEE dataset provides a real-world benchmark with genuine noise and exposure variation, so future methods can be tested without relying on simulated darkening of daytime images.","The fusion design transfers to other illumination-invariant modalities such as depth or event streams, where one sensor is robust to darkness and the other supplies color or dense context.","At 0.67 million parameters, the model is compact enough to make the architecture a candidate for embedded or mobile deployment."],"supporting_citations":[{"why":"Supplies the paired RGB-thermal images used for training and the simulated low-light evaluation on LLVIP.","marker":"[21]"},{"why":"Is the strongest RGB-only baseline that RT-X Net claims to outperform, and its illumination-estimator design is adopted in the method.","marker":"[6]"},{"why":"Provides the Retinex decomposition that underlies the illumination map and illumination features used to guide attention.","marker":"[5]"},{"why":"Is the prior thermal-guided CNN for low-light enhancement that the paper positions its cross-attention approach against.","marker":"[20]"},{"why":"Is a leading transformer baseline that incorporates SNR maps; the comparison isolates the benefit of cross-modal fusion.","marker":"[7]"},{"why":"Is a GAN-based low-light enhancement baseline included in the quantitative comparisons.","marker":"[16]"}],"fun_headline_variants":["RT-X Net fuses RGB and thermal for sharper night images","Thermal cross-attention network brightens low-light images","RGB+thermal cross-attention lifts night image quality","RT-X Net: thermal cues improve low-light enhancement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The real-world V-TIEE evaluation assumes that the visible and thermal cameras, which have different resolutions and fields of view, are aligned accurately enough at every pixel by homography for the cross-attention to fuse corresponding locations.","fun_headline_variants_meta":{"raw":{"variants":["RT-X Net fuses RGB and thermal for sharper night images","Thermal cross-attention network brightens low-light images","RGB+thermal cross-attention lifts night image quality","RT-X Net: thermal cues improve low-light enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000448,"raw_usage":{"total_tokens":2240,"prompt_tokens":901,"completion_tokens":1339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":1273}},"tokens_in":517,"tokens_out":1339,"duration_ms":11925,"temperature":1.0,"reasoning_tokens":1273,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:14:43.416526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take RT-X Net trained on LLVIP and evaluate it on V-TIEE with the thermal image deliberately shifted by 1, 3, and 5 pixels relative to the visible image. If the PSNR or LPIPS advantage over the RGB-only ablation shrinks or reverses with small shifts, the method's real-world benefit depends on near-perfect alignment rather than on thermal structure per se.","supporting_citations":[{"cited_title":"Low-light image en- hancement with normalizing flow,","cited_arxiv_id":null,"evidence_quote":"Supplies the paired RGB-thermal images used for training and the simulated low-light evaluation on LLVIP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the strongest RGB-only baseline that RT-X Net claims to outperform, and its illumination-estimator design is adopted in the method."},{"cited_title":"Quantitative comparison of RT-X Net with baseline methods on LLVIP and V-TIEE dataset","cited_arxiv_id":null,"evidence_quote":"Provides the Retinex decomposition that underlies the illumination map and illumination features used to guide attention."},{"cited_title":"Learning to see in the dark,","cited_arxiv_id":null,"evidence_quote":"Is the prior thermal-guided CNN for low-light enhancement that the paper positions its cross-attention approach against."},{"cited_title":"The experimental results demonstrate that the proposed approach outperforms recent state-of-the-art methods by effectively incorporating ther- mal information","cited_arxiv_id":null,"evidence_quote":"Is a leading transformer baseline that incorporates SNR maps; the comparison isolates the benefit of cross-modal fusion."},{"cited_title":"On the role of thermal imaging in automotive applications: A critical review,","cited_arxiv_id":null,"evidence_quote":"Is a GAN-based low-light enhancement baseline included in the quantitative comparisons."}],"review_version":1}