{"id":"67fa6f1b-ce65-43de-820d-3c5531ec8f09","arxiv_id":"2506.12738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A weighted, layer-wise annealed dropout applied at intermediate layers of blind super-resolution networks improves generalization on unseen degradations over prior regularization methods.","lead":"This paper introduces Adaptive Dropout, a regularization method that applies a weighted dropout to intermediate layers of blind super-resolution networks and anneals it layer by layer during training. The authors report consistent PSNR gains over prior dropout-based and alignment-based regularization methods on synthetic and real-world benchmarks, plus gains on denoising, deraining, dehazing, and generative restoration.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the headline claim: RRDB+Explicit Adaptive Dropout scores 21.00 dB on Urban100, 0.05 dB below Simple-Align's 21.05 dB, so the statement that the method outperforms Simple-Align on all synthetic datasets is false as written.","rationale":"The central claim of the paper is that Adaptive Dropout outperforms all past regularization methods on synthetic and real-world benchmarks. The supporting evidence for the synthetic half of that claim is Table 2, which compares directly with Simple-Align. A careful reading shows a row in that very table where the proposed method loses: RRDB+Explicit Adaptive Dropout (21.00 dB) and RRDB+Implicit Adaptive Dropout (20.96 dB) are both below RRDB+Simple-Align (21.05 dB) on Urban100. This is not a question of statistical significance or mechanism; it is a factual inconsistency between the stated conclusion and the reported data. Even if the entry is a transcription error, the paper as submitted contains a false claim, so the verdict should remain conditional pending correction. This concern is more load-bearing than the reader's variance-shift mechanism concern: the method could be right and the mechanism story wrong, but a direct counterexample in the headline comparison cannot be ignored. The reader's concern about missing error bars is related and supports the same conclusion that the empirical claim is established only conditionally.","tokens_in":15852,"tokens_out":8997,"duration_ms":84011,"concrete_test":"Extract the raw per-degradation PSNR logs for Table 2 rows RRDB+Simple-Align and RRDB+Explicit/Implicit Adaptive Dropout on Urban100, and confirm whether the reported 21.05/21.00/20.96 values are correct. If correct, revise the abstract and Section 5 claims to exclude or qualify the Urban100 result; if they are typographical, correct the table and rerun the entire Simple-Align comparison with at least 3 random seeds, reporting mean±std, because most synthetic differences in Table 2 are below 0.05 dB.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 states: \"Our proposed method outperforms Simple-Align [56] on all synthetic datasets, with notable improvements on the more complex datasets Manga109 and Urban100.\" Table 2, however, lists RRDB+Simple-Align at 21.05 dB on Urban100 and RRDB+Explicit Adaptive Dropout at 21.00 dB, with RRDB+Implicit Adaptive Dropout at 20.96 dB. This is a direct counterexample to the abstract's central claim that the method \"outperforms all past regularization methods\" on synthetic benchmarks. The discrepancy is not a subtle statistical issue: the compared numbers are 0.05–0.09 dB in the opposite direction. Many other synthetic entries in the same table also show margins of only 0.01–0.05 dB, which without error bars or seed averaging are within typical run-to-run variation. The paper's own Table 1 reports per-degradation PSNRs for the main method, while Table 2's synthetic columns appear to be averages over the eight degradation settings; the caption does not state this, making the comparison opaque. Whatever the intended protocol, the claim as written is not supported by the table.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive Dropout, a plug-in regularization method for blind super-resolution that replaces direct dropout at intermediate layers with a convex combination f(x)=w x+(1-w)dropout(x,p), trained with a layer-wise annealing schedule (implicit learnable weights or explicit block-by-block annealing). The authors derive a variance-shift quantity s_k, argue that direct dropout degrades intermediate features through training-testing inconsistency and unequal layer-wise generalization requirements, and report experiments on SRResNet, SwinIR, and RRDB across multiple synthetic and real-world datasets, as well as denoising, deraining, dehazing, and generative restoration tasks.","tokens_in":16014,"tokens_out":6331,"duration_ms":77231,"significance":"If the empirical claims are sustained, Adaptive Dropout would be a simple and broadly applicable regularizer for blind SR and adjacent restoration tasks, with code release, tests across several architectures, and consistent directional gains in many settings. The use of SRGA as an auxiliary generalization metric is also a positive feature. However, the central quantitative claim is contradicted by the paper's own Table 2, the variance-shift derivation is incomplete, and no repeated-seed statistics or error bars are reported, so the case as presented is not fully established.","major_comments":[{"comment":"The text and abstract claim that Adaptive Dropout outperforms all past regularization methods on synthetic benchmarks, but Table 2 provides a direct counterexample: on Urban100, RRDB+Simple-Align achieves 21.05 dB, while RRDB+Explicit Adaptive Dropout achieves 21.00 dB and RRDB+Implicit Adaptive Dropout 20.96 dB. This is not a subtle rounding issue; it is a loss on one of the datasets highlighted in the text. The authors must either correct the claim or clarify the protocol, including how the synthetic columns are aggregated across degradations and where the per-degradation comparisons with Simple-Align are reported, since the referenced supplementary material is not included.","section":"Section 5, Table 2"},{"comment":"The variance-shift calculation for adaptive dropout is not correct as written. If the dropout mask is Bernoulli as defined in the text, Var[x̂_k] is (1−p)σ²+p(1−p)μ², not the expression given in Section 3.2. If the authors intend inverted dropout with a 1/(1−p) scaling, then for f(x)=w x+(1−w)D(x) the variance includes a w²σ² term and a 2w(1−w)σ² cross-covariance term because the same x appears in both branches; the paper's s_k^ad formula drops these terms. Thus the claim that the adaptive format reduces the variance shift by a factor of (1−w)² is not established and the derivation should be redone with an explicit masking convention.","section":"Section 3.2, Eq. (1)"},{"comment":"The paper attributes the performance benefit to variance-shift reduction, but no experiment isolates variance shift from other effects of dropout, such as reduced effective capacity, altered gradient flow, or implicit ensembling. The learned/annealed schedule could be improving generalization through any of these channels. I recommend adding a control experiment: for example, compare adaptive dropout against standard dropout with matched expected mask density, or measure the variance shift directly during training and show that it predicts the observed relative performance across layers and schedules.","section":"Section 3 and 4.2"},{"comment":"The ablation table is inconsistent with Table 1: Table 5 lists no dropout as 25.66 dB and standard dropout as 24.89 dB, whereas Table 1 reports SRResNet with no regularization as 24.89 dB on Set5 clean and Dropout as 25.67 dB. The caption does not state the dataset, degradation, or evaluation protocol, so the reader cannot verify the component contributions. Please specify the protocol and align the numbers with Table 1, or explain the discrepancy explicitly.","section":"Section 5, Table 5"},{"comment":"No repeated-seed statistics or error bars are reported anywhere, yet many of the claimed advantages are 0.01–0.05 dB, which is within typical run-to-run variation for SR training. For the headline comparisons against Simple-Align and Dropout, please report mean ± std over at least three seeds, or clearly state if the reported numbers are single runs and adjust the strength of the claims accordingly.","section":"Section 5, Tables 1 and 2"}],"minor_comments":[{"comment":"The phrase 'also highly effective in other image restoration tasks' is grammatically incomplete; it should read 'and is also highly effective in other image restoration tasks.'","section":"Abstract"},{"comment":"In the applications paragraph, 'image draining' should be 'image deraining.'","section":"Section 5"},{"comment":"Algorithm 1 contains placeholder symbols (? and ???) instead of actual variable names; the algorithm should be typeset with the real variables so that the block-by-block annealing rule is understandable.","section":"Figure 3"},{"comment":"The caption does not define the real-world columns 'mild,' 'difficult,' and 'wild,' nor does it state how the synthetic columns are aggregated over the eight degradation settings; please add the missing definitions.","section":"Table 2"},{"comment":"The text says 'the inconsistency is manifested as the shift at intermediate layers after applying dropout,' but the corresponding figure (Figure 2) is not described in enough detail to see the accumulation along layers; a clearer annotation of the measured shift would help.","section":"Section 3.2"},{"comment":"The GitHub URL in the abstract contains a typo ('Adpative-Dropout'); if the repository name is intentionally misspelled this should be noted, otherwise the link should be corrected.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the method is simple enough to be broadly useful if the results hold. The Urban100 discrepancy in Table 2 and the inconsistent ablation baselines must be resolved before the paper can be considered for publication; the authors should also provide the supplementary material promised in the text, since the detailed Simple-Align comparisons are referenced there. I do not see evidence of misconduct, but the abstract's 'outperforms all past regularization methods' claim is too strong given the authors' own numbers and the absence of uncertainty quantification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and honestly presented: replace standard dropout at intermediate layers with f(x) = w x + (1-w) dropout(x,p), and anneal w block-by-block from shallow to deep during training. The authors show this improves generalization across multiple architectures (SRResNet, SwinIR, RRDB), many synthetic degradations, and several real-world datasets, and they also test it on denoising, deraining, dehazing, and generative SR. That is a solid amount of work, and the improvements, while modest (roughly 0.1 to 0.5 dB), are consistent. The variance-shift derivation is self-contained and gives a plausible mechanism, even if it is not causally isolated from other dropout effects like reduced capacity.\n\nThe soft spots are real but mostly fixable. The biggest issue is internal: Section 5 states the method outperforms Simple-Align on all synthetic datasets, but Table 2 shows RRDB+Explicit Adaptive Dropout at 21.00 dB on Urban100 versus Simple-Align's 21.05 dB, a direct counterexample. The abstract's blanket claim is also too strong. Many of the winning margins in Table 2 are 0.01-0.05 dB, and there are no repeated-seed statistics or error bars anywhere in the paper, so those differences could be noise. Second, the ablation table lists the no-regularization baseline as 25.66 dB, while Table 1 lists the same baseline as 24.89 dB; the mismatch is unexplained and makes the ablation hard to interpret. Third, Section 4.2 says the explicit schedule sets w to 0, but Algorithm 1 says it sets w to 1; these mean opposite things in Equation 1. Fourth, the layer-wise annealing schedule was reverse-engineered from the learnable-w variant on the same benchmarks used for the final claims, so there is selection risk. The code link has a typo and the supplementary materials are referenced but not provided.\n\nNone of this is fatal. The central empirical trend seems defensible, and the method is a reasonable extension of dropout to intermediate layers for SR. I would send this to peer review, but the authors must fix the Table 2 claim, either by restating it or by reporting the Urban100 number honestly, and they should add seed averages or confidence intervals. The paper will be useful for anyone working on regularization for blind SR or low-level vision training strategies. It does not open a new direction, but it is a competent step forward.","headline":"A useful, clearly written empirical paper on adaptive dropout for blind SR, but its own Table 2 contradicts the claim that it beats Simple-Align on all synthetic datasets, and the missing error bars make the small margins hard to trust.","tokens_in":807,"tokens_out":821,"would_cite":true,"duration_ms":27146,"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":"Dropout, applied as a weighted blend of original and dropped features with a layer-wise annealing schedule, becomes a generalizable regularizer for blind super-resolution instead of a destabilizing one.","keywords":["blind super-resolution","dropout regularization","training-testing inconsistency","variance shift","layer-wise annealing","feature generalization","image restoration"],"falsifier":"Measure the PSNR of a blind super-resolution model trained with standard intermediate dropout whose output is rescaled so that training and inference variances match exactly; if that rescaling reproduces the Adaptive Dropout gains, the variance-shift mechanism is confirmed, and if it does not, the gains come from the interpolation or the annealing schedule rather than from variance alignment.","tokens_in":15544,"feed_emoji":"🖼️","tokens_out":8601,"duration_ms":84074,"temperature":0.7,"pith_summary":"Blind super-resolution recovers a high-resolution image from a low-resolution one when the degradation process is unknown, and such models tend to overfit the degradations seen during training. Earlier dropout-based regularizers for this task only placed dropout (randomly masking features during training) before the final convolutional layer, because applying it to intermediate layers hurt performance. This paper argues that the harm comes from two inconsistencies: a variance shift between training and inference, and mismatched generalization needs of shallow versus deep layers. It proposes Adaptive Dropout, which replaces plain dropout with $f(x)=w x+(1-w)\\operatorname{dropout}(x,p)$ and anneals the weight $w$ layer by layer during training. In the paper's experiments this consistently beats prior regularization methods on synthetic and real-world benchmarks and transfers to denoising, deraining, dehazing, and generative super-resolution.","feed_headline":"Dropout across all layers now helps, not hurts, blind super-resolution","feed_subtitle":"A weighted per-layer blend of original and dropped features stops blind SR models overfitting training degradations.","key_machinery":"The key machinery is the adaptive dropout operation $f(x)=w x+(1-w)\\operatorname{dropout}(x,p)$, treated as a convex combination of the faithful feature $x$ and its randomly masked version, with $w$ either learned or annealed per block. It does two jobs at once: by retaining a fraction of the original feature it reduces the training-testing variance shift, whose size the paper measures for standard dropout as $s_k=\\frac{p}{1-p}\\left(\\frac{\\mu^2}{\\sigma^2}+1\\right)$; and by letting the value of $w$ vary across layers it balances fitting ability against generalization, annealing perturbation out of shallow layers earlier and keeping deeper layers in a longer generalization stage.","core_discovery":"The central claim is that dropout can be applied to intermediate layers of a blind super-resolution network if it is made adaptive rather than binary. The adaptive form $f(x)=w x+(1-w)\\operatorname{dropout}(x,p)$ keeps the original feature path intact while the dropped path regularizes representations, shrinking the training-testing variance shift by a factor of $(1-w)^2$ relative to standard dropout. Because shallow layers supply general features that need little perturbation while deep layers encode degradation-specific information that needs strong regularization, the authors couple this format with a layer-wise annealing schedule that removes perturbation from shallow blocks first and keeps deep blocks perturbed longer. The resulting regularizer, attached at intermediate layers and still integrated with final-layer dropout, is what produces the reported gains across architectures and restoration tasks.","pith_inferences":["One testable extension is to vary $p$ and $w$ independently on the same degradation suite and check whether the measured variance-shift quantity $s_k$ predicts the PSNR gain; if it does not, the adaptive format's success is not explained by variance reduction alone.","The convex-combination form is a learned interpolation between identity and noise injection, so the method may transfer to other regression tasks with distribution shift, such as depth estimation or deblurring, where intermediate features also mix general and task-specific content.","If the annealing schedule is the active ingredient, a variant that removes perturbation according to layer depth rather than a hand-set iteration count would be a natural next test of the mechanism.","The channel-ablation evidence suggests that explicit intermediate regularization, not final-layer regularization, rebalances channels; replicating that ablation with an independent protocol would confirm whether channel balance is the causal channel for the gains."],"forward_implications":["A single plug-in training change gives consistent PSNR gains over no regularization and over final-layer-only dropout on SRResNet, SwinIR, and RRDB across clean, blur, noise, JPEG, and combined degradations.","The layer-wise annealing schedule implies that deep layers, not shallow ones, should carry most of the regularization load in blind super-resolution, and that this asymmetry is what lets the method avoid the collapse that plain intermediate dropout causes.","Because the method also improves GAN-based and diffusion-based restoration models, it can serve as a general training strategy for image restoration rather than a fix limited to one architecture.","Adaptive Dropout integrates with the existing final-layer Dropout, yielding a unified regularization pipeline that explicitly constrains features at every depth."],"supporting_citations":[{"why":"Supplies the final-layer dropout baseline, the three empirical findings about intermediate dropout, and the experimental protocol this paper extends.","marker":"[29]"},{"why":"Simple-Align is the prior regularization method that this paper compares against on both synthetic and real-world datasets.","marker":"[56]"},{"why":"Provides the variance-shift analysis of dropout that the paper uses to quantify training-testing inconsistency.","marker":"[33]"},{"why":"Real-ESRGAN's degradation simulation is the training paradigm for the blind super-resolution experiments.","marker":"[57]"},{"why":"The practical degradation model defines the multi-degradation pool used to generate low-resolution training images.","marker":"[68]"},{"why":"SwinIR is one of the main backbone architectures whose performance with and without Adaptive Dropout is reported.","marker":"[34]"},{"why":"SRResNet is the other backbone architecture used for the main comparisons and ablations.","marker":"[30]"}],"fun_headline_variants":["Adaptive dropout unlocks intermediate layers for blind SR","Layer-wise dropout that actually boosts blind super-resolution","Blind super-resolution gains from adaptive layer dropout","Making dropout work on every layer of blind super-resolution","Adaptive dropout tames intermediate layers for generalizable SR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the performance drop from applying dropout at intermediate layers is caused mainly by a training-versus-testing variance shift, quantified by $s_k$, and that the adaptive format's reduction of that shift is what produces the gains; the paper does not isolate this variance-shift effect from other changes such as reduced capacity or altered gradient flow.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive dropout unlocks intermediate layers for blind SR","Layer-wise dropout that actually boosts blind super-resolution","Blind super-resolution gains from adaptive layer dropout","Making dropout work on every layer of blind super-resolution","Adaptive dropout tames intermediate layers for generalizable SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1292,"prompt_tokens":966,"completion_tokens":326,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":252}},"tokens_in":582,"tokens_out":326,"duration_ms":4132,"temperature":1.0,"reasoning_tokens":252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:44:07.398789+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the PSNR of a blind super-resolution model trained with standard intermediate dropout whose output is rescaled so that training and inference variances match exactly; if that rescaling reproduces the Adaptive Dropout gains, the variance-shift mechanism is confirmed, and if it does not, the gains come from the interpolation or the annealing schedule rather than from variance alignment.","supporting_citations":[{"cited_title":"Reflash dropout in image super-resolution","cited_arxiv_id":null,"evidence_quote":"Supplies the final-layer dropout baseline, the three empirical findings about intermediate dropout, and the experimental protocol this paper extends."},{"cited_title":"Navigating beyond dropout: An intriguing solution towards generalizable image super resolution","cited_arxiv_id":null,"evidence_quote":"Simple-Align is the prior regularization method that this paper compares against on both synthetic and real-world datasets."},{"cited_title":"Under- standing the disharmony between dropout and batch normal- ization by variance shift","cited_arxiv_id":null,"evidence_quote":"Provides the variance-shift analysis of dropout that the paper uses to quantify training-testing inconsistency."},{"cited_title":"Designing a practical degradation model for deep blind image super-resolution","cited_arxiv_id":null,"evidence_quote":"The practical degradation model defines the multi-degradation pool used to generate low-resolution training images."},{"cited_title":"Swinir: Image restoration us- ing swin transformer","cited_arxiv_id":null,"evidence_quote":"SwinIR is one of the main backbone architectures whose performance with and without Adaptive Dropout is reported."},{"cited_title":"Photo- realistic single image super-resolution using a generative ad- versarial network","cited_arxiv_id":null,"evidence_quote":"SRResNet is the other backbone architecture used for the main comparisons and ablations."}],"review_version":1}