{"id":"df872abc-facb-4695-b15b-6cefadd9f86e","arxiv_id":"2506.20969","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A conditional diffusion model with self-attention at half input resolution translates RGB images to thermal images with better object-level thermal detail than lower-resolution attention, evaluated on four datasets.","lead":"This paper tests a conditional diffusion model that turns ordinary RGB camera images into synthetic thermal images, aiming to give self-driving and field robots thermal data where none exists. The authors report that adding attention at higher image resolutions helps preserve hot objects like vehicles and pedestrians, and that training separate day and night models works better.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ablation evidence cited for the central claim measures only global SSIM/PSNR/FID, not the salient-object thermal signatures the paper says are improved; object-level or downstream validation is missing.","rationale":"I read the paper in good faith. The method is not internally inconsistent: attention at higher resolution plausibly helps retain spatial detail, and Table II is directionally consistent with the architecture claim. My concern is that the central claim as worded is about salient-object thermal signatures, yet Table II's metrics are global and the prior-method comparisons are qualitative. This is an evidence gap rather than a demonstrated error. The reader's weakest_assumption focuses on RGB underdetermining thermal physics; that is related but not identical, because even a model that perfectly learned all available RGB-thermal correlations in the training domain could still fail to preserve object-level signatures, and global metrics would not reveal it. Hence partial agreement. The verdict remains unchanged: with an object-level or downstream test the claim would either be substantiated or need to be scaled back to 'global image statistics improve.' I do not see grounds to reject, since the architecture choice is plausible and the reported internal numbers directionally support it.","tokens_in":8335,"tokens_out":4629,"duration_ms":57981,"concrete_test":"On the Freiburg daytime test split or KAIST (which provides pedestrian bounding boxes), annotate or obtain masks for pedestrians and vehicle tires. For Self-attention Models I and II, retrained with the same budget and at least 3 seeds, compute: (1) object thermal contrast = mean intensity inside GT object mask minus mean intensity of the surrounding local background; (2) average precision of a thermal pedestrian detector trained on real thermal data when applied to synthetic thermal images. Also compute the same metrics for PID and LDM outputs on the same test images. If Model II is not significantly better than Model I on both object contrast and detection AP, the claimed 'identification and correlation of high temperature objects' is not supported; if no baseline exceeds chance, the proposed synthetic data cannot be used for downstream navigation as claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central added value (Section I contribution 2; Section III-C) is that inserting self-attention at height/2 x width/2 lets the model learn and preserve the thermal signatures of salient objects such as pedestrians and vehicle tires. The only quantitative support is Table II, comparing Self-attention Model I and Model II on Freiburg daytime: PSNR 10.78->11.14, SSIM 0.59->0.65, FID 265.21->156.53. These are whole-image statistics. They can improve because of better background texture, global contrast, or histogram matching while object-level thermal errors—pedestrians rendered cooler than the road, tires missing their friction heat—remain or even worsen. The qualitative figures (Figures 1-3) cannot settle this, and the comparison against PID and LDM in Figures 2-3 is also qualitative only. Thus the evidence cited for the strongest claim does not measure the property the claim is about. Section III-D itself concedes that RGB appearance underdetermines thermal appearance (water intensity reverses day/night; vegetation and soil have near-identical night intensities), so systematic object-level failures are a genuine risk rather than a remote possibility. If an object-level evaluation showed no improvement for Model II, the paper's central contribution would be unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ThermalDiffusion, a conditional denoising diffusion probabilistic model (DDPM) for RGB-to-thermal image translation, and claims that adding self-attention at height/2 x width/2 improves the preservation of thermal signatures of salient objects such as pedestrians and vehicle tires. It also studies the effect of training on daytime versus nighttime data and evaluates fine-tuning across the Freiburg, Caltech Aerial, KAIST, and FLIR datasets. Quantitative results are reported as PSNR/SSIM/FID for the self-attention ablation and the day/night split, while comparisons with PID and LDM are shown only qualitatively.","tokens_in":8594,"tokens_out":4663,"duration_ms":52775,"significance":"If the central claim holds, the contribution is practically useful: a simple architectural change in a conditional diffusion model that improves object-level thermal fidelity could help synthesize thermal training data for autonomous navigation. The paper's strengths include evaluation on multiple public datasets, a clear day/night ablation, and comparison with two recent diffusion-based baselines. However, the load-bearing evidence is incomplete: the claimed object-level improvement is quantified only with whole-image metrics, and the comparisons to PID/LDM are qualitative. The paper does not include a downstream task validation, which would be the natural way to substantiate the data-augmentation motivation.","major_comments":[{"comment":"The central claim that adding self-attention at height/2 x width/2 preserves thermal signatures of salient objects (pedestrians, vehicle tires) is supported only by global PSNR/SSIM/FID on the whole image. These metrics can improve through background texture, global contrast, or histogram changes while object-level thermal errors remain or worsen. Please add object-level quantitative evidence, for example pedestrian/vehicle bounding-box or segmentation-based thermal fidelity (mean absolute error or SSIM within ROIs), or a downstream detection/segmentation experiment using the generated thermal images. Without this, the claim in Contribution 2 is not directly tested.","section":"Section III-C, Table II, Figure 5"},{"comment":"The claimed superiority over PID and LDM is supported only by qualitative side-by-side images. No quantitative metrics are reported for the baseline comparisons. Please report PSNR/SSIM/FID (and, if feasible, object-level metrics) for all methods and datasets, ideally with multiple random seeds and error bars or significance tests. This is necessary because the paper's contribution includes a claim of improved fidelity over prior diffusion-based translators.","section":"Section IV-D, Figures 2-3"},{"comment":"The day/night conclusion is based on one dataset with single-run metrics and no error bars. In addition, the text says \"One exception arises where the combined model performs worse when evaluated on nighttime data,\" but Table I shows the combined model (FID 164.89) is close to the night-only model (FID 162.44) and better than the day-only model on night data (FID 241.16). Please clarify the exception and provide variance estimates across training runs; otherwise the claim that time-specific models are systematically better is not established.","section":"Section III-D, Table I"},{"comment":"The motivating application is augmenting existing RGB datasets with synthetic thermal data for downstream robotics tasks, but the paper never evaluates a downstream task (e.g., semantic segmentation, object detection, or depth estimation) trained on synthetic thermal data and tested on real thermal data. Given Section III-D's own admission that RGB underdetermines thermal appearance (water intensity reversal between day and night, vegetation and soil nearly identical at night), a downstream evaluation is the most direct way to show that the synthetic data are useful despite these physical ambiguities. Please add such an experiment or temper the data-augmentation claim.","section":"Abstract, Section I, Section III-D"}],"minor_comments":[{"comment":"The reverse-process distribution is written with x_t although the derivation uses y_t; please make the notation consistent throughout Section III-A.","section":"Section III-A, Eq. (6)"},{"comment":"There are typos such as \"noisyy\" and \"y_t instance given the image at time t-1\"; please correct these textual errors.","section":"Section III-A, text around Eq. (8)"},{"comment":"The sentence \"We extends upon the idea\" is grammatically incorrect; it should be \"We extend upon the idea.\"","section":"Section III-C"},{"comment":"Please clarify that the rows are test periods and the columns are training sets, since the current caption does not state this explicitly.","section":"Table I caption"},{"comment":"Reference [2] does not point to the original Pix2Pix paper; please cite Isola et al. and check whether [28] and [32] are duplicates of the same latent diffusion paper.","section":"References"},{"comment":"The statement that \"the combined model tends to have a larger variance for contrast\" is presented without supporting evidence; either add a quantitative histogram or variance analysis or remove the speculative explanation.","section":"Section III-D"},{"comment":"Please state the input resolution, training compute, and memory usage for the two self-attention variants so readers can assess the cost of the added attention level.","section":"Section IV-B, Table II"}],"recommendation":"major_revision","confidential_remarks":"The novelty is modest and the main new result is an architectural tweak whose quantitative support is currently weak. The absence of quantitative baseline comparisons and downstream validation, rather than any methodological error, is the main barrier to acceptance. A major revision with object-level and downstream experiments would be the appropriate path."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a workmanlike incremental paper. It takes the conditional DDPM of Saharia et al., adds an extra self-attention block at height/2 resolution, and tests it on four RGB-thermal datasets. The genuinely useful parts are the empirical studies of day/night model splitting and cross-dataset fine-tuning; the architectural change is small and its benefit is not well evidenced.\n\nWhat the paper does well: Section III-D gives a physically grounded explanation of why day and night thermal images differ, and Table I shows that separate day/night models beat a combined model, which is useful for anyone building thermal data pipelines. The transfer results from Freiburg to Caltech and KAIST are plausible and show the model generalizes to new terrains.\n\nThe soft spots are load-bearing. The paper's main claim is that higher-resolution self-attention preserves the thermal signatures of pedestrians and tires, but the only quantitative support is a single ablation on Freiburg daytime with PSNR, SSIM, and FID scores. Those are global statistics that can improve without any improvement at the object level. The comparison against PID and LDM is qualitative only, so no quantitative superiority is established. There are no error bars over seeds, no code release, and no downstream task evaluation. The paper itself concedes in Section III-D that RGB appearance underdetermines thermal appearance, which is honest but underscores the need for a detector trained on synthetic thermal and evaluated on real thermal data. None of that is present.\n\nThis paper is for robotics researchers who want to augment RGB datasets with synthetic thermal images. They will get a reasonable baseline and a useful day/night split finding. It deserves a serious referee because the problem is practical and the method is sane, but it needs major revision: quantitative comparison with at least PID and LDM, object-level or downstream validation, error bars, and code or full training details.\n\nMy recommendation: treat it as a borderline paper that requires hard revision, not as a desk reject.","headline":"An incremental but honest application of conditional diffusion to RGB-to-thermal translation; the self-attention extension is plausible but the evidence for its benefit is thin.","tokens_in":9109,"tokens_out":2995,"would_cite":false,"duration_ms":34483,"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":"A conditional diffusion model with self-attention at half input resolution generates thermal images from RGB that preserve the heat signatures of pedestrians and tires better than attention at lower resolutions alone.","keywords":["RGB-to-thermal translation","conditional diffusion models","self-attention resolution","thermal imaging","autonomous driving","day-night domain gap","image-to-image translation","dataset augmentation"],"falsifier":"Segment pedestrians and vehicle tires in the RGB image, then compare the average intensity of the generated thermal image inside those segments against the ground-truth thermal image; if the half-resolution attention model fails to beat the lower-resolution baseline on this object-level thermal fidelity metric, the paper's central claim is unsupported.","tokens_in":8168,"feed_emoji":"🌡️","tokens_out":9544,"duration_ms":86589,"temperature":0.7,"pith_summary":"The paper tries to show that a conditional denoising diffusion model, with self-attention run at half the input height and width, can convert ordinary RGB images into thermal images that retain the heat signatures of salient objects such as pedestrians and vehicle tires. This matters because robotics and driving datasets rarely include thermal imagery, and synthetic thermal data could let existing RGB-only datasets be augmented for downstream perception tasks without collecting paired thermal measurements. The paper also argues that thermal imagery differs so strongly between day and night that separate models trained on each period outperform a single model trained on combined data. The evidence is an ablation on the Freiburg dataset, where the half-resolution attention model raises SSIM from 0.59 to 0.65 and lowers FID from 265.21 to 156.53 relative to the lower-resolution attention baseline.","feed_headline":"Finer attention sharpens synthetic thermal images","feed_subtitle":"RGB-only driving data could gain usable thermal images for night and degraded-weather perception.","key_machinery":"The key mechanism is the conditional denoising diffusion probabilistic model (conditional DDPM): a U-Net that denoises a Gaussian-noise image into a thermal image while conditioning on the RGB source image at every reverse step. The paper's modification is to add a self-attention layer at the highest resolution of half the input height and width, on top of the attention levels inherited from the guided-diffusion backbone. Self-attention at this finer scale is what lets the model associate particular visual objects with their characteristic thermal intensities, such as hotter tires and warmer bodies, which are lost when attention operates only at coarser resolutions.","core_discovery":"The central claim is that pushing self-attention in a conditional DDPM up to the resolution of $\\frac{H}{2} \\times \\frac{W}{2}$, rather than only at $\\frac{H}{4}$, $\\frac{H}{8}$, and $\\frac{H}{16}$, lets the model correlate visual appearance with object-level thermal properties at a scale where those properties are actually visible. As a result, the synthesized thermal images better capture warm human bodies and friction-heated tires, and they carry more high-frequency detail. A second finding is that models trained and evaluated on the same time of day (day or night) outperform a combined day-plus-night model, because water's apparent thermal intensity reverses between day and night, while vegetation and soil contrast collapses at night.","pith_inferences":["If the mapping is reliable, synthetic thermal images could serve as pseudo-labels for training downstream detectors and segmenters, but the paper itself does not demonstrate this downstream benefit; testing on such tasks would be the natural next step.","The day/night split implies that any RGB-to-thermal translation system should be conditioned on time-of-day metadata or trained separately per period; the paper leaves this as a training-data choice rather than a learned conditioning variable.","The improvement from higher-resolution attention may partly reflect increased model capacity rather than attention per se; an ablation that matches parameter count would clarify the mechanism.","Because thermal appearance depends on factors invisible in RGB, such as recent solar exposure and water's day/night reversal, purely appearance-based translation is likely to have a ceiling; hybrid models that inject physical priors could extend it."],"forward_implications":["Existing RGB-only driving datasets can be augmented with synthetic thermal images, enabling supervised training of thermal-based perception without collecting paired thermal data.","Separate day and night models should be used in practice, because combined training degrades nighttime thermal fidelity due to the dominance of high-contrast daytime samples.","Fine-tuning a model pretrained on a richer dataset (Freiburg) on a smaller target dataset (KAIST) improves thermal translation metrics over training from scratch, suggesting transfer learning is beneficial.","Adding self-attention at half resolution increases compute and memory cost, so the choice of attention resolution is a practical trade-off between fidelity and resource use."],"supporting_citations":[{"why":"Supplies the self-attention principle for diffusion-based image-to-image translation that the paper extends to RGB-to-thermal.","marker":"[5]"},{"why":"Provides the base guided-diffusion architecture whose attention resolution levels the paper modifies by adding a half-resolution level.","marker":"[6]"},{"why":"Gives the conditional DDPM formulation the paper uses to condition thermal generation on the RGB source image.","marker":"[7]"},{"why":"The paired RGB-thermal Freiburg dataset used for the central daytime/nighttime training and attention-resolution experiments.","marker":"[8]"},{"why":"A physics-informed diffusion baseline that the paper compares against when arguing its model better represents pedestrians and vehicles.","marker":"[26]"},{"why":"Latent diffusion baseline that the paper's model is compared against on KAIST and FLIR; the comparison supports the claim of superior thermal detail.","marker":"[28]"},{"why":"Documents day/night thermal intensity reversals that motivate the paper's separate day and night models.","marker":"[31]"}],"fun_headline_variants":["High-res attention sharpens synthetic thermal images","ThermalDiffusion uses fine self-attention to improve synthetic thermal quality","Higher-res attention yields crisper synthetic thermal images","Day-specific thermal diffusion beats mixed-day model","ThermalDiffusion fine attention sharpens RGB-to-thermal translation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that an object's thermal appearance is determined well enough by its visible RGB appearance that a model trained on paired RGB-thermal images can synthesize realistic thermal images for new scenes that have no thermal sensor.","fun_headline_variants_meta":{"raw":{"variants":["High-res attention sharpens synthetic thermal images","ThermalDiffusion uses fine self-attention to improve synthetic thermal quality","Higher-res attention yields crisper synthetic thermal images","Day-specific thermal diffusion beats mixed-day model","ThermalDiffusion fine attention sharpens RGB-to-thermal translation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3094,"prompt_tokens":867,"completion_tokens":2227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2149}},"tokens_in":483,"tokens_out":2227,"duration_ms":17624,"temperature":1.0,"reasoning_tokens":2149,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:36:57.604090+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Segment pedestrians and vehicle tires in the RGB image, then compare the average intensity of the generated thermal image inside those segments against the ground-truth thermal image; if the half-resolution attention model fails to beat the lower-resolution baseline on this object-level thermal fidelity metric, the paper's central claim is unsupported.","supporting_citations":[{"cited_title":"Palette: Image-to-image diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the self-attention principle for diffusion-based image-to-image translation that the paper extends to RGB-to-thermal."},{"cited_title":"Diffusion models beat gans on image syn- thesis,","cited_arxiv_id":null,"evidence_quote":"Provides the base guided-diffusion architecture whose attention resolution levels the paper modifies by adding a half-resolution level."},{"cited_title":"Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images,","cited_arxiv_id":null,"evidence_quote":"The paired RGB-thermal Freiburg dataset used for the central daytime/nighttime training and attention-resolution experiments."},{"cited_title":"Pid: Physics-informed diffusion model for infrared image generation,","cited_arxiv_id":null,"evidence_quote":"A physics-informed diffusion baseline that the paper compares against when arguing its model better represents pedestrians and vehicles."},{"cited_title":"High-resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Latent diffusion baseline that the paper's model is compared against on KAIST and FLIR; the comparison supports the claim of superior thermal detail."},{"cited_title":"Unmanned ground vehicle perception using thermal infrared cameras,","cited_arxiv_id":null,"evidence_quote":"Documents day/night thermal intensity reversals that motivate the paper's separate day and night models."}],"review_version":1}