{"id":"25a237a3-0ee3-409d-a96d-833ca7beedab","arxiv_id":"2411.16327","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CapHDR2IR combines HDR visible-light input with dense caption features to generate infrared images, reporting state-of-the-art scores on the HDRT dataset.","lead":"The paper presents CapHDR2IR, a method that generates infrared images from high dynamic range visible-light photos, adding a dense captioning branch that supplies scene semantics. It reports higher PSNR and SSIM than existing visible-to-infrared translation methods on the HDRT dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is under-supported because three RGB-to-IR methods cited in Related Work (PAS-GAN, InfraGAN, IR-GAN) are absent from Table 1; the claim should be conditioned on adding these comparisons.","rationale":"The reader's conditional verdict is appropriate, and the dataset concern about HDRT being the only benchmark is real but secondary. The more immediate threat to the strongest claim is that the comparison set is incomplete: the paper itself identifies PAS-GAN, InfraGAN, and IR-GAN as related deep RGB-to-IR methods in Related Work, yet none appears in Table 1. A state-of-the-art claim cannot be verified without comparing against all directly relevant state-of-the-art methods that the authors acknowledge. An additional internal inconsistency supports the need for caution: in the Loss Weight section, the row for the chosen weights (alpha=10, beta=0.1) reports PSNR 1.879, SSIM 6.163, MSE 2.574, LPIPS 3.141, which matches the SDR2IRV2 row in Table 2 rather than the full HDR-input CapHDR2IR row (1.976, 6.359, 2.242, 3.035). This suggests the hyperparameter sweep may not have been run on the final model configuration, further weakening confidence in the reported numbers. Still, the core architectural idea is coherent and the ablations are directionally consistent, so the paper merits conditional acceptance pending stronger comparative validation and artifact release.","tokens_in":12201,"tokens_out":8953,"duration_ms":89091,"concrete_test":"Run PAS-GAN (Wang et al. 2024), InfraGAN (Özkanoğlu and Ozer 2022), and IR-GAN (Ma et al. 2024) on the HDRT 8k/2k split using their official implementations and the paper's metrics (PSNR, SSIM, MSE, LPIPS), supplying both SDR and HDR inputs as in Table 1. If any omitted method achieves higher PSNR/SSIM or lower MSE/LPIPS than the corresponding CapHDR2IR entry, the SOTA claim fails; otherwise the concern is resolved. This also requires releasing the exact split and preprocessing used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is 'our model surpasses traditional methods and other deep learning approaches' (Experiments, Comparison with Other Methods). Table 1 compares only CycleGAN, Pix2Pix, MUNIT, sRGB-TIR, and RGB2IR among deep methods. Yet Related Work explicitly names PAS-GAN, InfraGAN, and IR-GAN as visible-to-infrared translation methods. None is evaluated or explained as unavailable. Because the claim is stated as SOTA over 'other deep learning approaches,' omitting three directly relevant methods makes the headline result unverifiable. The absence is load-bearing: if any omitted method outperforms CapHDR2IR on HDRT (with the same train/test split and metrics), the SOTA statement is false. No code, weights, or per-method training details are provided for the baselines, so the table cannot be independently reproduced.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CapHDR2IR, a two-branch generator for visible-to-infrared translation. One branch is an image generation encoder-decoder; the other is a dense captioning branch based on ETDC (Shao et al. 2023), whose features are fused into the encoder at multiple scales through caption fusion modules. The input is an HDR visible image, tone-mapped for the caption branch, and the training loss combines perceptual and GAN terms. Experiments on the HDRT dataset compare the method with traditional and deep baselines in both SDR-input and HDR-input settings, with ablations for HDR input, caption branch, caption fusion, and loss weights. The authors claim state-of-the-art performance on HDRT and argue that HDR input and dense captioning jointly address detail loss and pseudo-thermal-crossover artifacts.","tokens_in":12382,"tokens_out":3339,"duration_ms":34018,"significance":"If the reported results hold, the paper makes a useful contribution by identifying two concrete failure modes in visible-to-infrared translation and addressing them with HDR input and dense captions. The motivation is clearly argued, and the ablation study in Table 2 consistently shows gains from HDR input and from the caption branch, which supports the internal consistency of the design. The paper is also clearly written and the qualitative figures illustrate the intended effects. However, the central SOTA claim is currently supported only by a single self-collected benchmark, omits several directly relevant baselines, and relies on test-set-based hyperparameter selection; these issues prevent the claim from being taken at face value.","major_comments":[{"comment":"The comparison in Table 1 omits methods that the Related Work section explicitly identifies as visible-to-infrared translation approaches: PAS-GAN, InfraGAN, and IR-GAN. Because the text states that 'our model surpasses traditional methods and other deep learning approaches,' the absence of these three directly relevant baselines means the SOTA claim is not verifiable from the reported evidence. Please add these comparisons, or state concretely why they cannot be run (e.g., no public implementation), and condition the claim accordingly.","section":"Experiments, Comparison with Other Methods, Table 1"},{"comment":"The loss weights alpha and beta in Eq. (1) are selected using the test set: Table 3 reports reconstruction metrics on the test set for different weight combinations, and the chosen values (alpha=10, beta=0.1) are then used for the main results. This makes the reported numbers in Table 1 the result of test-set selection. Please use a validation split for hyperparameter selection and report the corresponding test-set results, and include error bars or significance tests over multiple runs.","section":"Loss Weight, Table 3"},{"comment":"All experiments are conducted on HDRT, a dataset introduced by the same research group (Peng et al. 2024a), and no evaluation is performed on established RGB-IR benchmarks such as KAIST or FLIR. Since the central contribution depends on the claim that HDR inputs and caption fusion generalize across visible-to-infrared translation, the single-benchmark evaluation leaves the generalization claim unsupported. Please add at least one external dataset or explicitly restrict the conclusion to HDRT.","section":"Experiments, Dataset"},{"comment":"The improvements from the caption fusion module are small in Table 2 (e.g., PSNR 1.966 vs 1.976 for HDR2IRV3 vs CapHDR2IR), and no variance or significance information is reported anywhere in the ablation. Without repeated-run statistics, the marginal contribution of the final fusion module is difficult to distinguish from noise. Please provide standard deviations or significance tests for the ablation rows.","section":"Ablation Study, Table 2"}],"minor_comments":[{"comment":"Eq. (3) reuses alpha as a tone-mapping scaling factor, while Eq. (1) uses alpha as the perceptual-loss weight; this overloaded notation is confusing and should be resolved by renaming one of the two symbols.","section":"Methodology, Eq. (3)"},{"comment":"In Eq. (7), F^i_fused_cap is undefined; it should presumably be F^i_aligned_cap from Eq. (6).","section":"Methodology, Eq. (7)"},{"comment":"Table 3 is difficult to read because the alpha and beta values are not formatted as separate columns with clear headers (e.g., rows like 'e-1 e-1'); please reformat the table so that each hyperparameter value is explicit.","section":"Table 3"},{"comment":"No code, trained weights, or detailed baseline training protocols are provided, which makes exact reproduction of Table 1 difficult; please consider releasing code and weights or providing more complete experimental details.","section":"Experiments, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The paper's evaluation relies on the authors' own HDRT dataset, and the dense captioning backbone ETDC has overlapping authorship. This alone is not disqualifying, but the editor should weigh the single-benchmark evidence and the test-set-based hyperparameter selection when evaluating the strength of the SOTA claim. The ablation study is the strongest part of the paper and suggests the core idea is sound; the requested baseline additions and statistical rigor would make the contribution much more convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CapHDR2IR does something genuinely new: it combines HDR visible images with dense caption features for visible-to-IR translation. The ablation study supports the central design, and the physical motivation is sound. The SOTA claim, however, is currently overbroad and under-supported.\n\nWhat is actually new and works: using HDR as input is a sensible fix for the dynamic-range problem, and the dense caption branch is fused at multiple scales with spatial attention. Table 2 shows every added component (HDR input, pre-processing, caption branch, caption fusion) improves PSNR/SSIM/LPIPS. That is real evidence. The qualitative examples show reduced pseudo-thermal-crossover, and the architecture is described clearly enough to reproduce in principle.\n\nSoft spots:\n\n1. Table 1 omits PAS-GAN, InfraGAN, and IR-GAN, all cited in Related Work as RGB-to-IR methods. The abstract and comparison section claim SOTA over “deep learning approaches,” but those directly relevant baselines are absent without explanation. That is load-bearing for the headline.\n\n2. The only benchmark is HDRT, from the same group, and the dense caption backbone ETDC shares authors. The assumption that HDRT is well-registered, representative, and unbiased is not independently established. One external dataset (e.g., KAIST or FLIR) would substantially raise confidence.\n\n3. The loss weights alpha and beta are chosen from a sweep (Table 3) that appears to be on the evaluation split. There is no separate validation set and no error bars or significance tests, so the margins in Table 1 could be noise.\n\n4. The “first to integrate VLMs into image generation” claim is likely false in that general form. The defensible claim is narrower: first dense-caption fusion for RGB-to-IR with HDR input.\n\nThe citation pattern is not egregious, but the self-referential elements are real. Overall this is a coherent method with a solid ablation story. Readers working on RGB-to-IR, HDR imaging, or vision-language-guided generation will get value from it.\n\nI would send it to peer review, but acceptance should be conditional on adding the three missing baselines, separating validation from test for hyperparameter selection, and reporting variance or significance. If those are addressed, the paper is competitive.","headline":"A well-ablated HDR+caption RGB-to-IR method whose SOTA claim is conditional on adding three missing baselines and more rigorous evaluation.","tokens_in":12911,"tokens_out":3186,"would_cite":true,"duration_ms":31904,"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":"The paper claims that combining high-dynamic-range visible input with dense caption features produces state-of-the-art infrared image generation on the HDRT dataset.","keywords":["visible-to-infrared translation","high dynamic range imaging","dense captioning","vision-language models","image-to-image translation","pseudo-thermal-crossover","HDRT dataset"],"falsifier":"Take an independently collected set of paired HDR-visible and infrared images, and compare the full CapHDR2IR against the same generator without the caption branch and against an SDR-input version, measuring whether distinct objects with similar temperatures remain separable in the output. If the caption branch offers no region-level improvement, or if SDR inputs match HDR inputs once the dataset is chosen to emphasize low-dynamic-range scenes, then the claim that HDR and captions drive the reported gains is falsified.","tokens_in":12029,"feed_emoji":"🌡️","tokens_out":10007,"duration_ms":81624,"temperature":0.7,"pith_summary":"CapHDR2IR is a proposed framework for translating visible-light images into infrared images without specialized IR hardware. The paper's central claim is that two shortcomings explain why prior RGB-to-IR translation loses fidelity: standard dynamic range (SDR) visible input cannot represent extreme luminance, and the generator has no scene-level understanding, so objects with similar temperatures collapse into what the authors call pseudo-thermal-crossover artifacts. CapHDR2IR addresses both by taking high-dynamic-range (HDR) visible input and adding a dense caption branch, a pretrained vision-language model (ETDC) that emits per-region textual descriptions which are fused into the generator's encoder. On the HDRT dataset, the authors report the best PSNR, SSIM, MSE, and LPIPS in both the SDR-input and HDR-input comparison groups, and confirm the role of each component in an ablation study. If these results hold, visible-to-infrared synthesis becomes more practical for low-light and high-contrast scenes, because the input does the exposure work and the captions do the semantic work.","feed_headline":"HDR input plus dense captions wins infrared image synthesis","feed_subtitle":"A dense caption branch plus HDR input yields sharper, more meaningful infrared images.","key_machinery":"The machinery is a two-branch generator. The caption branch tone-maps HDR input to SDR using the log-average luminance formula, then runs ETDC, a textual-context-aware dense captioner with a ResNet-101 backbone and region proposal network, to obtain region features and caption features. The image generation branch is an encoder-decoder: caption fusion modules use a spatial attention map computed from concatenated visual and caption features to align caption features with visual features at multiple scales, and up-sampling fusion modules concatenate center-cropped encoder features with transposed-convolution upsampled features. The training objective is perceptual loss from a pre-trained VGG network weighted at 10 plus a GAN loss weighted at 0.1.","core_discovery":"The discovery the paper is trying to establish is that HDR input plus dense captions, not just a better generator architecture, is what moves visible-to-infrared translation forward. In the HDR group of Table 1, CapHDR2IR reports the best numbers on PSNR, SSIM, MSE, and LPIPS, beating RGB2IR, sRGB-TIR, CycleGAN, Pix2Pix, MUNIT, and classical methods on every metric. The authors argue that the HDR input preserves information in clipped shadows and highlights, while the caption branch retains objects that the naive generator would blur into dark, uniform regions that mimic thermal crossover. The ablation table supports the ordering: HDR input alone, the caption branch alone, and the caption fusion module each add measurable gains, and the full model with HDR pre-processing for the caption branch is the best configuration.","pith_inferences":["A test the paper does not run: evaluating the same method on independently collected RGB-IR benchmarks would settle whether the HDR advantage generalizes beyond HDRT.","If the caption branch is doing the work, an ablation replacing dense captions with a segmentation or object-detection feature map would tell whether language is necessary or merely a source of high-level features.","A direct artifact metric—counting how often known distinct objects collapse to the same gray level in generated infrared output—would quantify pseudo-thermal crossover; the paper relies on qualitative examples.","Because the caption branch is pretrained on SDR imagery, the log-average tonemapping of HDR input is safety-critical, and any mismatch could silently degrade the semantic benefit."],"forward_implications":["Infrared generation pipelines should ingest HDR or multi-exposure visible images instead of single SDR frames when targeting dark and high-contrast scenes.","Dense captioning can act as a semantic prior that suppresses artifacts in domain transfer, a mechanism not limited to infrared imaging.","Each component of CapHDR2IR—HDR input, the caption branch, and the caption fusion module—contributes measurable improvement, and the full configuration is the only one that tops every metric on the HDRT benchmark.","The chosen loss weights (perceptual weight 10, GAN weight 0.1) place far more emphasis on feature-level fidelity than on adversarial realism, and the weight ablation identifies this as the best operating point."],"supporting_citations":[{"why":"Supplies the HDRT dataset, the single benchmark whose HDR-visible/IR pairs ground every quantitative and qualitative result.","marker":"Peng et al. 2024a"},{"why":"Provides the ETDC dense captioning model that forms the caption branch and supplies semantic region features.","marker":"Shao et al. 2023"},{"why":"RGB2IR is the strongest RGB-to-IR baseline the method must outperform in Table 1.","marker":"Huang, Huang, and Wu 2024"},{"why":"sRGB-TIR is the edge-guided RGB-to-thermal baseline that represents the tailored prior work.","marker":"Lee et al. 2023"},{"why":"Pix2Pix is the paired image-to-image translation baseline against which improvement is measured.","marker":"Isola et al. 2017"},{"why":"CycleGAN is the unpaired translation baseline in the comparison.","marker":"Zhu et al. 2017"},{"why":"MUNIT is the multimodal translation baseline in the comparison.","marker":"Huang et al. 2018"},{"why":"Color transfer is the classical statistical baseline in both SDR and HDR groups.","marker":"Reinhard et al. 2001"}],"fun_headline_variants":["HDR and captions sharpen IR synthesis","Caption-aware HDR boosts IR image fidelity","HDR plus dense captions bests IR translation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the HDRT dataset is well-registered, representative, and unbiased in favor of HDR input or caption-friendly scenes; if that fails, the reported state-of-the-art may not hold on other visible-to-infrared benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["HDR and captions sharpen IR synthesis","Caption-aware HDR boosts IR image fidelity","HDR plus dense captions bests IR translation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1296,"prompt_tokens":938,"completion_tokens":358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":554,"tokens_out":358,"duration_ms":4338,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:13:32.813681+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an independently collected set of paired HDR-visible and infrared images, and compare the full CapHDR2IR against the same generator without the caption branch and against an SDR-input version, measuring whether distinct objects with similar temperatures remain separable in the output. If the caption branch offers no region-level improvement, or if SDR inputs match HDR inputs once the dataset is chosen to emphasize low-dynamic-range scenes, then the claim that HDR and captions drive the reported gains is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ETDC dense captioning model that forms the caption branch and supplies semantic region features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RGB2IR is the strongest RGB-to-IR baseline the method must outperform in Table 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"sRGB-TIR is the edge-guided RGB-to-thermal baseline that represents the tailored prior work."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Pix2Pix is the paired image-to-image translation baseline against which improvement is measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CycleGAN is the unpaired translation baseline in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MUNIT is the multimodal translation baseline in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Color transfer is the classical statistical baseline in both SDR and HDR groups."}],"review_version":1}