{"id":"f3c70fe1-75a6-44c4-b490-92607f5e45ea","arxiv_id":"2508.07926","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"ScoreAug trains diffusion denoisers on augmented noisy inputs with augmentation-prediction targets, reporting reduced overfitting and better FID in data-limited settings.","lead":"This paper proposes ScoreAug, a data augmentation method for diffusion models that transforms noisy inputs instead of clean images and trains the denoiser to predict the augmented target, reporting reduced overfitting and better generation quality on CIFAR-10, FFHQ, AFHQv2, and ImageNet. The supplied full text is encoding-corrupted, so the technical claims could not be independently checked.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-invertible augmentations (crop/mask/color) break the claimed score relationship unless the theory restricts to measure-preserving bijections; corrupted text leaves this unresolved.","rationale":"The paper's central contribution is an equivariant objective that is supposed to learn scores in transformed spaces. The strongest claim depends on the score-relationship theorem covering the actual augmentation set. My concern is not about whether flips are useful (they are), but about internal consistency: the L2 objective used to define ScoreAug has a unique minimizer for any augmentation, while a score exists only under structural conditions. For non-invertible maps, these differ. Because the supplied full text is corrupt, I cannot verify whether Sec. 3 restricts to the safe class; the reader correctly marked the paper UNVERDICTED, and my read does not change that verdict. If the recovered theory already restricts to measure-preserving bijections and the experiments match, no issue; otherwise the reported gains need another explanation. The proposed test is designed to settle this directly.","tokens_in":14619,"tokens_out":6832,"duration_ms":83631,"concrete_test":"For a fixed σ and a Gaussian-mixture p_data, compute the exact minimizer D*(y)=E[g(x)|g(x+σε)=y] and the score σ²∇_y log p_Y(y) for g = random coordinate mask (m⊙v, m_i~Bernoulli(0.5)). The mask makes p_Y singular (no Lebesgue density), while D* is well-defined; hence the two cannot be equal. Then repeat for random crop/resize on a simple image data set; if the same mismatch appears, ScoreAug's objective is not score matching for augmentations in the experimental protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that, for every augmentation g used in the experiments, minimizing E||D(g(x_t))−g(x)||² yields a denoiser whose output is the score (or an unbiased denoising target) of the corresponding augmented noisy distribution. This holds only under a strong equivariance condition: g(x+σε)=g(x)+σε′ with ε′∼N(0,σ²I) and g a measure-preserving bijection (e.g., orthogonal flips/rotations). For augmentations usually used in the reported benchmarks—random crop, cutout, color jitter, resize—the condition fails: crop/mask is non-injective and produces a degenerate or lower-dimensional y; color jitter changes the noise covariance; nonlinear mappings like resize do not commute with additive Gaussian noise. Then the L2 minimizer is D*(y)=E[g(x)|g(x_t)=y], which is not a score in the original data space and, for a mask, is identically zero on masked coordinates. The abstract's phrase 'general transformations' therefore overclaims unless Sec. 3 contains an additional restriction; the supplied text is corrupted, so this cannot be checked. If the experiments are restricted to measure-preserving bijections, the concern is moot; if they include crops/cutouts, the reported FID gains are empirical and not grounded in the stated score relationship.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ScoreAug, a data augmentation scheme for diffusion models. The method applies a transformation g to the noisy input x_t and trains the denoiser to output g(x), the transformed clean target, rather than x. The authors argue that this gives an equivariant objective that lets the denoiser learn scores in multiple denoising spaces, effectively augmenting the score itself. The abstract claims a theoretical analysis of the score relationship under general transformations and reports FID improvements on CIFAR-10, FFHQ, AFHQv2, and ImageNet, with reduced overfitting at small data scales and varying model capacities, stable convergence, and compatibility with standard augmentation. In the copy made available to me, only the abstract is legible; the main text, equations, and tables are garbled, so the derivations and quantitative results cannot be checked.","tokens_in":14905,"tokens_out":6889,"duration_ms":86803,"significance":"The idea is simple and potentially valuable: standard augmentation for diffusion is usually applied to clean training images, while ScoreAug exploits the denoiser's role and uses transformed noisy inputs with augmented targets, which may increase effective sample size and improve robustness. If the score relationship is correct for a nontrivial, practically useful class of transformations, the method would offer a principled alternative to heuristic augmentation and a possible anti-overfitting tool for data-limited regimes. The benchmark selection is appropriate and the claims are falsifiable. However, the paper's contribution cannot currently be assessed: no readable proof or experiment table is available, and the exact assumptions on g are not stated. I therefore cannot vouch for significance beyond plausibility.","major_comments":[{"comment":"In the supplied copy the body text after the abstract is unreadable; equations and theorem statements are missing or garbled. The abstract's central theoretical assertion—that one can relate scores in different spaces under 'general transformations'—is therefore not derivable from the submitted material. This is load-bearing: ScoreAug's correctness depends on the L2 objective E||D(g(x_t))−g(x)||² having a Tweedie/score interpretation for the specific transformations used. I cannot verify the assumptions (invertibility, measure preservation, equivariance of noise, Jacobian corrections) or the proof. The authors must provide a readable theory section and state the exact class of allowed g.","section":"Main text (theoretical claims)"},{"comment":"The experimental section is not verifiable: the abstract reports 'significant performance improvements' and overfitting mitigation on CIFAR-10, FFHQ, AFHQv2, and ImageNet, but the table fragments visible in the supplied text contain no interpretable FID values, baselines, or error bars. There are no readable training configurations, augmentation hyperparameters (types, magnitudes, probabilities), or evaluation protocols. The claims of significance and stable convergence are therefore not supported by any quantitative evidence in the copy under review.","section":"Experiments"},{"comment":"The paper uses the phrase 'general transformations' without definition. The score identity is known to hold exactly for measure-preserving bijections that commute with additive Gaussian noise (e.g., orthogonal flips/rotations). For non-injective linear maps such as cropping, the objective still has a valid posterior-mean interpretation in the image space, but for masking/cutout the noise covariance degenerates, and for nonlinear color/resize augmentations the noise is not preserved. If the experiments use any of these, the theoretical grounding needs to be extended or qualified. Please state the transformation family and prove or cite the identity for each experimental augmentation.","section":"Abstract / transformation family"},{"comment":"The abstract's statement that ScoreAug 'circumvent[s] data leakage issues under certain conditions' is unsupported. The conditions are not specified, no definition of leakage is given, and no experiment isolating leakage is reported in the readable portion. This should be clarified or removed.","section":"Abstract / data leakage claim"}],"minor_comments":[{"comment":"Define 'score augmentation' formally with an equation; the current abstract is imprecise about how the equivariant objective differs from a standard denoising objective.","section":"Abstract"},{"comment":"If crop and resize are used, clarify how the network output is projected back to the original image space at sampling; the output dimensionality of the denoiser appears ambiguous.","section":"Experiments"},{"comment":"Provide hyperparameters and sensitivity analysis for transformation type, magnitude, and application probability; these are likely important free parameters.","section":"Experiments"},{"comment":"The paper should compare with standard augmentation for diffusion models (e.g., ADA, conditional augmentation) and explain the difference in both objective and effect.","section":"Related work"},{"comment":"The phrase 'stable convergence properties' should be accompanied by training curves or a quantitative convergence metric.","section":"Experiments"}],"recommendation":"uncertain","confidential_remarks":"The supplied manuscript is unreadable, so I cannot form a soundness verdict. If the editor has a clean PDF, it should be sent for review. Based on the abstract, the method is plausible but the theory and experiments need close checking, especially the transformation assumptions behind the score identity and whether the reported FID gains are reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take: the only thing we can evaluate from this copy is the idea in the abstract, because the full text is corrupted (encoding garbage). That's an infrastructure problem, not necessarily the authors' fault, but it means no one should rely on any of the experimental numbers until we see the real PDF.\n\nWhat's actually interesting: the method is a genuine variant of augmentation for diffusion. Applying the transform to the noisy input and making the denoiser predict the augmented target is an equivariant objective, not just standard augmentation on clean images. That could be a useful trick for data-limited regimes, and the claim that it composes with ordinary augmentation has practical value.\n\nSoft spots, in proportion: (1) 'general transformations' almost certainly overclaims. The score transformation identity can't hold for arbitrary g. It holds for measure-preserving bijections — flips, maybe orthogonal rotations — and the stress-test is right that crops, masks, color jitter, and resize break the exact score relationship because the noise doesn't transform covariantly and non-invertible maps give degenerate conditionals. If their experiments rely on those augmentations, the gains are empirical regularization effects, not consequences of the stated score theory. That's not fatal — empirical tricks are fine — but the paper needs to say which family the theory covers and not conflate it with what the experiments use. (2) We can't check the related-work section, so the novelty claim is weaker than stated. (3) We can't verify that the benchmarks exist or the FID numbers are honest.\n\nWho it's for: people working in data-limited diffusion or augmentation strategies. If the real paper is readable and the theory section restricts to invertible, measure-preserving maps, this could be a solid contribution. As it stands, I'd want to see a clean copy and have a referee check Sec. 3 before believing the theoretical framing. I would not cite this from the abstract alone.\n\nRecommendation: send a clean copy to a serious referee. The idea is worth the referee time; the current manuscript can't be evaluated, but the underlying proposal is not obviously wrong and could be useful. So: would accept for review, with heavy revision expected.","headline":"Idea is plausible and worth a look, but this copy is unreadable and the 'general transformations' claim needs a hard check before anyone cites it.","tokens_in":15412,"tokens_out":1961,"would_cite":false,"duration_ms":22464,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Applying augmentations to noisy data, not clean data, improves diffusion training and cuts overfitting.","keywords":["diffusion models","score matching","data augmentation","equivariant learning","overfitting","denoising","generative modeling","FID"],"falsifier":"Train ScoreAug on a fixed dataset with two augmentation sets: flips only, and random crops/cutouts. At several noise levels, estimate the true score $\\nabla \\log p_t$ of the transformed noisy distribution by Monte Carlo from the training set, and compare it with what the denoiser predicts. If the crop-trained denoiser deviates far from the true score while the flip-trained denoiser matches it, the 'general transformations' theorem does not cover the transformations actually used.","tokens_in":14498,"feed_emoji":"🔄","tokens_out":7004,"duration_ms":80331,"temperature":0.7,"pith_summary":"ScoreAug is a data-augmentation scheme aimed directly at the diffusion denoiser. It applies a transformation to the noisy input and changes the training target to the transformed clean image, so the denoiser learns an equivariant map rather than a fixed denoiser. This lets one network estimate scores across many denoising spaces, which the paper argues reduces overfitting and improves generation when data is limited. The paper derives how scores relate across such transformed spaces under general transformations, and reports lower FID on CIFAR-10, FFHQ, AFHQv2, and ImageNet, plus synergy with conventional augmentation.","feed_headline":"Augment noisy diffusion inputs to cut overfitting","feed_subtitle":"ScoreAug makes the denoiser predict the augmented target, improving FID on small datasets.","key_machinery":"Equivariant denoising objective. The transformation $g$ is applied simultaneously to the noisy input and to the clean target: training pairs $(x_t, x_0)$ become $(g(x_t), g(x_0))$, so the denoiser must be equivariant under $g$. This turns a single network into an estimator of the score — the gradient of the log-density $\\nabla \\log p_t$ — across the family of transformed denoising spaces. The paper's score-relationship identity is the machinery that connects those estimated scores back to the original generative task.","core_discovery":"The paper's central claim is that augmentation for diffusion models should be applied after noise is added and should be reflected in the denoising target, not applied to clean images as a preprocessing step. Concretely, for a transformation $g$, ScoreAug feeds the network the noisy-input pair $(g(x_t), t)$ and asks it to predict $g(x_0)$, the transformed clean image. Because the target and the input transform together, the denoiser becomes equivariant and its score estimate — the gradient of the log-density that drives generation — is defined with respect to the transformed denoising space. The paper proves a relationship between scores in the transformed and original spaces under general t","pith_inferences":["The paper does not spell out the boundary of its 'general transformations' theorem. A direct test is to compare ScoreAug with only flips against ScoreAug with crops: if crops still help but the predicted scores no longer match the true score of the augmented noisy distribution, the method is acting as a regularizer rather than as exact score augmentation.","The noisy-space design suggests a mechanism for the claimed leakage avoidance: augmented images never appear as clean training examples, so the model cannot memorize them as real data; this could be verified by checking whether generated samples reproduce augmented artifacts.","The same equivariant-target idea should transfer to other denoising tasks — audio, video, or graph generation — whenever a cheap transformation can be applied to the noisy input and the clean target; this is an extension the paper does not run."],"forward_implications":["A single ScoreAug-trained denoiser learns scores for many transformed denoising spaces, so it needs fewer clean examples per effective target.","ScoreAug reports lower FID than both plain training and standard clean-data augmentation on CIFAR-10, FFHQ, AFHQv2, and ImageNet.","The method closes much of the overfitting gap across data scales and model capacities, and training converges stably.","ScoreAug can be stacked with ordinary clean-data augmentation for additional gains.","Under the conditions stated in the paper, ScoreAug avoids the data-leakage problem associated with standard augmentation."],"supporting_citations":[],"fun_headline_variants":["Augment noisy diffusion data to fight overfitting","ScoreAug: predict transformed targets for diffusion","Equivariant denoising: augment inputs and targets","Boost diffusion FID by augmenting noisy pairs","Cut overfitting in diffusion with score augmentation"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the score-relationship theory covers the non-invertible augmentations actually used in the experiments, such as crops and cutouts; if it only holds for simple flips, the learned scores are biased and the reported gains are empirical rather than grounded in the stated theory.","fun_headline_variants_meta":{"raw":{"variants":["Augment noisy diffusion data to fight overfitting","ScoreAug: predict transformed targets for diffusion","Equivariant denoising: augment inputs and targets","Boost diffusion FID by augmenting noisy pairs","Cut overfitting in diffusion with score augmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":1059,"prompt_tokens":751,"completion_tokens":308,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":237}},"tokens_in":495,"tokens_out":308,"duration_ms":3607,"temperature":1.0,"reasoning_tokens":237,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:46:03.928037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ScoreAug on a fixed dataset with two augmentation sets: flips only, and random crops/cutouts. At several noise levels, estimate the true score $\\nabla \\log p_t$ of the transformed noisy distribution by Monte Carlo from the training set, and compare it with what the denoiser predicts. If the crop-trained denoiser deviates far from the true score while the flip-trained denoiser matches it, the 'general transformations' theorem does not cover the transformations actually used.","supporting_citations":[],"review_version":1}