{"id":"e59ce449-6a38-493c-9b13-a898dd43d5e5","arxiv_id":"2608.09150","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"OGG-FR is a plug-and-play training update that separates redundant and innovative parts of the FFT loss gradient and gates the innovative part by a confidence score, improving UAV infrared super-resolution in most tested settings.","lead":"This paper proposes OGG-FR, a training rule that splits the frequency-loss gradient into parts that duplicate the pixel-loss gradient and parts that add new information, then adjusts how strongly the new part is used during super-resolution training. It reports improved infrared image quality on a UAV thermal benchmark, but gives no code and no comparison against other gradient-balancing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains are single-run and mostly small; without error bars or existing gradient-surgery baselines, the 'broad improvements' may be run-to-run noise.","rationale":"The reader's weakest_assumption focuses on the reliability of g_perp_freq and the s_t calibration. That is a valid mechanistic concern, but it is not the first place the central claim fails. Even if s_t were perfectly calibrated, the empirical evidence for 'broad gains' is a single-run table with many small or negative deltas. The decisive question is whether the measured improvements are reproducible. A 0.02–0.3 dB PSNR gap without error bars is the kind of result that frequently disappears when seeds are varied. The reader already lists 'no error bars' among the weaknesses but does not make it the principal assumption; I treat it as the most load-bearing because it directly governs the truth of the central claim. The s_t concern is secondary: if the headline effect vanishes under repeated seeds, the confidence-gate question is moot; if the effect survives, the gate's calibration can be refined later. Thus I recommend keeping the CONDITIONAL verdict until the multiple-seed check is run. Agreement is 'partial' because the reader and I both see unresolved risk, but we rank the concerns differently.","tokens_in":10904,"tokens_out":7675,"duration_ms":75388,"concrete_test":"Re-run Table 1 for at least three representative settings—a large-gain case (ShuffleMixer on UAV-BD ×4), a small-gain case (HiT on UAV-BI ×8), and a negative case (RGT on UAV-BD ×4)—with five independent seeds for both weighted-sum and OGG-FR, matching the paper's reported hyperparameters. Report mean±std of PSNR. If the OGG-FR advantage over the baseline is smaller than the pooled standard deviation in two or more of the three settings, the 'broad gains' claim is not statistically established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports a single training run per model/setting, with no error bars, multiple seeds, or significance tests. Many of the claimed gains are small (e.g., SMFANet+ on UAV-BI ×8 changes PSNR by −0.023 dB; HiT on UAV-BI ×8 improves by +0.018 dB), and several entries regress. In SR training, PSNR variation across seeds is commonly ±0.1–0.3 dB at these magnitudes, so a large fraction of the improvements in Table 1 could be within run-to-run noise. If that is the case, the abstract's claim of 'broad gains' across backbones/degradations/scales is not supported. This concern is independent of whether the s_t confidence gate is well-calibrated; even a perfect gate would not rescue a headline result that is statistically indistinguishable from the baseline. The absence of comparisons with existing gradient-surgery methods (PCGrad, CAGrad, MGDA alone) also means the observed differences cannot be attributed specifically to OGG-FR's orthogonal-gradient mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OGG-FR, a training-time optimization framework for UAV infrared image super-resolution. It decomposes the frequency-loss gradient into a component parallel to the pixel-loss gradient and an orthogonal component, then uses MGDA with variance rectification when the two gradients conflict (negative cosine similarity) and a residual-confidence-scaled orthogonal component when they are compatible. The method is evaluated on the Low-light UAV thermal benchmark under BI and BD degradations at x4 and x8, across eight SR backbones, reporting PSNR/SSIM/NIQE gains over standard weighted-sum training, together with gradient-conflict analyses and ablations of the confidence slope and variance-rectification coefficient.","tokens_in":11171,"tokens_out":5565,"duration_ms":53619,"significance":"OGG-FR is clearly specified: Algorithm 1 is complete, the MGDA closed-form coefficient in Eq. (6) is correct, and the method adds no inference-time parameters or FLOPs. The paper is also honest about cases without improvement and explicitly acknowledges the fixed S=1 inflection point as a limitation. If the reported gains are reproducible, this would be a practically useful plug-and-play optimizer for lightweight IR super-resolution. However, the central empirical claim of 'broad gains' is not yet established: Table 1 reports single runs without error bars or significance tests, no existing gradient-manipulation method is used as a baseline, and there are numerical inconsistencies between Table 1 and Table 3. The method itself is not circular: the update is computed from current gradients and residual statistics, not from test-set results.","major_comments":[{"comment":"All entries in Table 1 are based on a single training run with no error bars, multiple seeds, or significance tests. Many reported gains are small (e.g., HiT on UAV-BI x8 improves by +0.018 dB PSNR, and SMFANet+ on UAV-BI x8 regresses by -0.023 dB), and SR training commonly exhibits seed-to-seed PSNR variation on the order of 0.1-0.3 dB at these magnitudes. The abstract's 'broad gains' claim is therefore not supported without seed-level statistics; a large fraction of the table could be consistent with run-to-run noise.","section":"Table 1"},{"comment":"OGG-FR is never compared with existing gradient-manipulation or loss-balancing methods such as PCGrad, CAGrad, or MGDA alone, even though the related-work section positions OGG-FR against them. Since OGG-FR combines MGDA with an orthogonal-gradient injection, the experiments cannot attribute the observed differences specifically to the orthogonal-gradient mechanism; direct comparisons with these methods are needed to support the claimed novelty and to test the assumption that the orthogonal FFT-gradient component is reliable frequency-domain innovation.","section":"Experiments / Main Quantitative Results"},{"comment":"The text states that backbones were included 'for which paired baseline and OGG-FR training records were available,' and the set of reported models differs across settings (e.g., SwinIR, HAT, and RGT appear only in selected rows of Table 1). This is not a principled evaluation protocol and introduces potential selection bias. The authors should specify how all models were chosen, report the same complete model set across all BI/BD and x4/x8 settings, and ideally release training records or code to make the selection auditable.","section":"Experimental Setup"},{"comment":"The default-kappa and default-beta results in Table 3 do not match the OGG-FR results for the same setting in Table 1. For example, HAT on UAV-BI x4 is reported as 35.0089 dB in Table 3 but 34.3125 dB in Table 1, and HiT on UAV-BI x4 is 35.7043 dB in Table 3 but 33.7178 dB in Table 1. Since both tables are described as UAV-BI x4, this numerical inconsistency undermines reproducibility and must be resolved or explicitly explained (e.g., different training epochs or data splits).","section":"Table 3 vs Table 1"}],"minor_comments":[{"comment":"The small constant epsilon in the denominator of S(r_H) is not defined; please state its value.","section":"Eq. (11)"},{"comment":"The phrase 'approximately 500 training epochs' is too vague; report the exact epoch count or the stopping criterion used for each backbone.","section":"Experimental Setup"},{"comment":"The term 'gradient gaming cosine' is used without a formal definition; define rho_t clearly as the cosine similarity and consider using a more standard name such as 'gradient alignment cosine.'","section":"Method, Eq. (3)"},{"comment":"The distinction between 'ShuffleMixer' and 'ShuffleMixer (base)' is not explained; clarify what configuration each row corresponds to.","section":"Table 1"},{"comment":"Several NIQE values worsen under OGG-FR even when PSNR and SSIM improve; the paper should discuss this perceptual-metric trade-off rather than only reporting NIQE in the tables.","section":"Tables 1 and 2"},{"comment":"The paper references supplementary sensitivity diagnostics but does not include the supplementary material in the reviewed text; ensure it is available and clearly linked.","section":"Ablation Studies"},{"comment":"The caption does not specify which backbones and training settings correspond to the infrared and visible-domain curves; add this information.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The main concerns are the single-run evaluation, the absence of gradient-surgery baselines, the unconventional backbone-selection wording, and the Table 1/Table 3 numerical discrepancies. I would ask the editor to require the authors to provide multi-seed results with significance statements, to add comparisons against PCGrad/CAGrad/MGDA, and to resolve the Table 1/Table 3 inconsistencies. Clarifying how 'paired training records were available' and releasing code or training logs would also help establish reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth a look if you work on multi-loss SR training. The genuinely new piece is the two-regime update rule: OGG-FR decomposes the FFT gradient into parallel and orthogonal components relative to the pixel gradient, uses MGDA as a safe base in the conflict regime with a variance-rectified orthogonal addition, and in the compatible regime discards the redundant parallel part and injects the orthogonal innovation scaled by a residual-confidence gate. That specific combination is not in the cited literature. The math is correct, the algorithm is fully specified, there is no inference-time cost, and the paper honestly reports settings where OGG-FR does not help. The gradient-conflict measurements in Fig. 1 and Fig. 5 are a useful piece of motivation, and the plug-and-play property across several backbones is appealing.\n\nNow the soft spots, in proportion. The empirical case is the weak section. Table 1 reports a single training run per model and setting, with no error bars, no seeds, no significance tests. Many of the gains are small (HiT on UAV-BI x8 improves by 0.018 dB; SMFANet+ on UAV-BI x8 regresses by 0.023 dB), and in SR training seed-to-seed PSNR variation at these magnitudes is commonly 0.1-0.3 dB. That means a sizeable fraction of the \"broad gains\" could be run-to-run noise. The paper also omits comparisons with existing gradient manipulation methods — PCGrad, CAGrad, GradNorm, MGDA alone — so even where OGG-FR does better than weighted sum, we cannot attribute the difference to the orthogonal-innovation mechanism specifically. The phrase \"for which paired baseline and OGG-FR training records were available\" is concerning; it reads as post-hoc selection and needs clarification. The s_t confidence gate with a fixed S=1 inflection is acknowledged as a limitation, which is honest, but it is still a hand-set heuristic and the ablation covers only three backbones.\n\nThe central optimization idea is coherent and the derivation holds; the issue is evidential, not logical. If the authors add multi-seed results with error bounds and compare against at least PCGrad and MGDA, this would be a respectable contribution. Without those, the abstract's claim of broad gains is not supported.\n\nWho is this for? Researchers working on optimization for multi-loss SR or gradient surgery generally. It deserves a serious referee, but the referee should push for the missing baselines and statistics. I would not desk-reject it.","headline":"Plausible optimizer for multi-loss IR super-resolution training, but the headline claim of broad gains rests on single runs with no error bars and no existing gradient-surgery baselines.","tokens_in":11647,"tokens_out":1613,"would_cite":true,"duration_ms":18369,"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":"Separating the frequency gradient into a redundant parallel part and an orthogonal innovation part stabilizes multi-loss training for UAV infrared super-resolution.","keywords":["infrared image super-resolution","UAV thermal imaging","multi-loss training","frequency-domain loss","gradient conflict","orthogonal gradient decomposition","MGDA","lightweight super-resolution"],"falsifier":"Train on clean infrared images with added synthetic sensor noise (for example Gaussian noise) and compare OGG-FR against standard weighted-sum training: if the PSNR/SSIM gains shrink or reverse as the orthogonal gradient becomes dominated by noise, then the method is not separating innovation from noise. A sharper ablation would replace $g_{\\perp}^{\\rm freq}$ with a random orthogonal vector of the same norm while keeping the confidence gate, to test whether the specific content of the orthogonal component matters.","tokens_in":10714,"feed_emoji":"📡","tokens_out":6104,"duration_ms":50980,"temperature":0.7,"pith_summary":"The paper addresses a training instability in UAV infrared image super-resolution: when the pixel-domain L1 loss and the frequency-domain FFT loss are combined as a fixed weighted sum, their gradients often point in conflicting directions, and the frequency gradient blends real high-frequency detail with sensor noise. It proposes OGG-FR, a plug-and-play update rule that decomposes the frequency gradient into a component parallel to the pixel gradient (treated as redundant or harmful) and an orthogonal component (treated as the only useful frequency innovation). Conflicting iterations use a Pareto-safe MGDA direction plus a variance-suppressed orthogonal term; compatible iterations discard the parallel part and inject the orthogonal term scaled by a confidence score derived from the high-frequency residual. On the Low-light UAV thermal benchmark, most tested backbones improve in PSNR/SSIM at x4 and x8 under both bicubic and blur-downsampling degradations, with larger gains than on a visible-domain reference.","feed_headline":"Gradient split yields broad gains in UAV infrared super-resolution","feed_subtitle":"A plug-and-play update that separates useful frequency signal from redundant and noisy parts lifts thermal-image PSNR/SSIM.","key_machinery":"The central object is the orthogonal-gradient decomposition of the frequency gradient with respect to the pixel gradient, gated by the gradient gaming cosine $\\rho_t$ that selects one of two regimes. In the conflict regime, the Multiple Gradient Descent Algorithm (MGDA) provides a Pareto-safe base direction and a variance rectifier $R(g_{\\perp}^{\\rm freq}) = g_{\\perp}^{\\rm freq}/(1+\\beta \\operatorname{Var}(g_{\\perp}^{\\rm freq}))$ stabilizes the orthogonal injection. In the compatible regime, a confidence score $s_t = \\sigma(\\kappa(S(r_H)-1))$ computed from the TV-to-$L^1$ ratio of the high-frequency residual $r_H$ decides how strongly the orthogonal innovation is added.","core_discovery":"The central claim is that the gradient of the FFT loss should be interpreted through its geometry relative to the pixel gradient before being merged into the update. Writing $g_{\\rm freq} = g_{\\parallel}^{\\rm freq} + g_{\\perp}^{\\rm freq}$, where $g_{\\parallel}^{\\rm freq}$ is the projection onto $g_{\\rm pix}$ and $g_{\\perp}^{\\rm freq}$ is the remainder, the paper argues that $g_{\\parallel}^{\\rm freq}$ is redundant in cooperative settings and harmful in conflict, while $g_{\\perp}^{\\rm freq}$ carries the frequency-domain innovation that the pixel loss cannot provide. The update is $h_t = g_{\\rm pix} + \\gamma_{\\rm boost} s_t g_{\\perp}^{\\rm freq}$ when the gradient cosine $\\rho_t \\ge 0$, and $h_t = h_{\\rm mgda} + \\gamma_{\\rm safe} R(g_{\\perp}^{\\rm freq})$ when $\\rho_t < 0$, where $h_{\\rm mgda}$ is the minimum-norm point in the convex hull of the two gradients and $R$ suppresses high-variance components. The paper reports that this rule improves reconstruction across lightweight and transformer backbones, and that the gains are consistent with the measured higher gradient conflict in infrared than in visible images.","pith_inferences":["A testable extension is to replace the fixed threshold $S=1$ with a per-dataset or per-batch adaptive statistic; if the residual ratio's scale varies with sensor noise, an adaptive inflection should improve gains on high-noise thermal sensors.","The variance rectifier in the conflict regime acts as a trust region on the orthogonal component; this suggests a prediction that the optimal $\\gamma_{\\rm safe}$ should shrink as sensor noise increases, which could be checked on synthetic-noise corruptions of the benchmark.","The zero boundary on $\\rho_t$ could be generalized to a soft or asymmetric transition; the few negative results on stable baselines are consistent with the confidence gate occasionally overweighting weak orthogonal components, so a softened gate may trade a little gain for fewer regressions."],"forward_implications":["The same gradient decomposition can be applied to any multi-loss training setup where one loss's gradient is noisy or partially redundant, not just infrared super-resolution.","Because OGG-FR changes only the training update and adds no inference cost, lightweight edge-deployed super-resolution models can adopt it without architectural changes.","The reported gains are larger on infrared than on visible-domain data, which ties the method's benefit to the degree of gradient conflict present in the training signal.","The fixed inflection point $S(r_H)=1$ and hyperparameters $\\kappa,\\beta$ are kept constant across all settings; the paper notes that dataset-adaptive calibration is a remaining direction.","The framework is not guaranteed to improve every architecture: the paper documents a few model-setting pairs, such as RGT on UAV-BD at x4 and SwinIR on UAV-BI at x4, where OGG-FR slightly lowers PSNR."],"supporting_citations":[{"why":"Supplies the Low-light UAV thermal benchmark and the BI/BD, x4/x8 evaluation splits used for all main results.","marker":"Zhao et al. 2025"},{"why":"Provides the MGDA formulation for the Pareto-safe base gradient in the conflict regime.","marker":"Sener and Koltun 2018"},{"why":"The pairwise gradient-conflict removal method (PCGrad) that motivates the conflict-aware manipulation and serves as a contrast in related work.","marker":"Yu et al. 2020"},{"why":"Documents the low-contrast, high-frequency-noise properties of infrared images that motivate treating the frequency gradient as a mixed signal.","marker":"Huang et al. 2025"},{"why":"Supplies the DIV2K visible-domain reference dataset used for the cross-domain gradient-behavior comparison.","marker":"Agustsson and Timofte 2017"}],"fun_headline_variants":["Orthogonal gradient split lifts UAV infrared super-resolution","When gradients clash, split orthogonal for thermal super-resolution","Separate redundant and useful gradients to boost UAV thermal SR","Frequency rectification sharpens UAV infrared super-resolution","A smarter gradient rule for UAV infrared super-resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the orthogonal part of the frequency gradient is genuine detail worth injecting rather than sensor noise, and that the confidence score computed from the high-frequency residual with a fixed inflection point correctly measures how much of that detail to trust.","fun_headline_variants_meta":{"raw":{"variants":["Orthogonal gradient split lifts UAV infrared super-resolution","When gradients clash, split orthogonal for thermal super-resolution","Separate redundant and useful gradients to boost UAV thermal SR","Frequency rectification sharpens UAV infrared super-resolution","A smarter gradient rule for UAV infrared super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000741,"raw_usage":{"total_tokens":3348,"prompt_tokens":1026,"completion_tokens":2322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":2248}},"tokens_in":642,"tokens_out":2322,"duration_ms":15319,"temperature":1.0,"reasoning_tokens":2248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:32:30.648872+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train on clean infrared images with added synthetic sensor noise (for example Gaussian noise) and compare OGG-FR against standard weighted-sum training: if the PSNR/SSIM gains shrink or reverse as the orthogonal gradient becomes dominated by noise, then the method is not separating innovation from noise. A sharper ablation would replace $g_{\\perp}^{\\rm freq}$ with a random orthogonal vector of the same norm while keeping the confidence gate, to test whether the specific content of the orthogonal component matters.","supporting_citations":[],"review_version":1}