Score Augmentation for Diffusion Models

Boyuan Jiang; Di Zhang; Kun Gai; Liang Hou; Pengfei Wan; Qi Yan; Renjie Liao; Xin Tao; Yuan Gao

REVIEW 4 major objections 5 minor 1 references

Score Augmentation for Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Applying augmentations to noisy data, not clean data, improves diffusion training and cuts overfitting.

desk verdict 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. read the letter →

arxiv 2508.07926 v1 pith:TRHBEKPR submitted 2025-08-11 cs.LG

Liang Hou , Yuan Gao , Boyuan Jiang , Xin Tao , Qi Yan , Renjie Liao , Pengfei Wan , Di Zhang

show 1 more author

Kun Gai

This is my paper · ORCID

classification cs.LG MSC 68T07

keywords diffusionmodelsscorematchingdataaugmentationequivariantlearningoverfittingdenoisinggenerativemodelingFID

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

What'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.

Soft 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.

Who 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.

Recommendation: 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.

Referee Report

4 major / 5 minor

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.

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 (4)

[Main text (theoretical claims)] 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.
[Experiments] 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.
[Abstract / transformation family] 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.
[Abstract / data leakage claim] 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.

minor comments (5)

[Abstract] Define 'score augmentation' formally with an equation; the current abstract is imprecise about how the equivariant objective differs from a standard denoising objective.
[Experiments] 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.
[Experiments] Provide hyperparameters and sensitivity analysis for transformation type, magnitude, and application probability; these are likely important free parameters.
[Related work] The paper should compare with standard augmentation for diffusion models (e.g., ADA, conditional augmentation) and explain the difference in both objective and effect.
[Experiments] The phrase 'stable convergence properties' should be accompanied by training curves or a quantitative convergence metric.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; ScoreAug is defined as a training objective and is validated against external FID benchmarks.

full rationale

The central claim of ScoreAug is a construction: the method transforms noisy inputs and trains the denoiser to predict the transformed original target, i.e. fit D(g(x_t)) to g(x). This is a defined learning objective, not a parameter fitted to the reported evaluation metric. The reported FID improvements are measured on external benchmarks (CIFAR-10, FFHQ, AFHQv2, ImageNet), so the main empirical claims are externally falsifiable and independent of any self-citation. The theoretical statement about score relationships in transformed spaces is presented as a derivation; the readable abstract does not define 'score augmentation' as the conclusion of the theorem, so the result is not equivalent to its premise by construction. The corrupted full text prevents full verification of the theorem's assumptions, but a potential mismatch between the theory (which likely requires invertible/measure-preserving transformations) and the augmentations used in practice (e.g., crop/mask/color) is a correctness or overclaim concern, not circularity: even if the score relationship fails for non-invertible maps, the FID gains are not logically forced by the paper's own definitions. No load-bearing self-citation chain or imported uniqueness theorem is discernible in the provided text. Therefore no circular step can be exhibited with quotation, and the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

Abstract-only audit. The full text is corrupted, so the ledger reflects only what the abstract asserts. No fitted constants and no new entities are visible; the main axioms are the standard diffusion training objective and the equivariance/score-relationship assumption for the augmentation family.

free parameters (1)

augmentation family and its hyperparameters (transformation types, magnitudes, application probability)
Not specified in the abstract; such choices are typically hand-tuned and directly affect FID. The complete implementation details are unreadable in the supplied text.

assumptions (2)

standard math Standard score matching / denoising score matching objective is a valid training target for the denoiser (background from score-based generative modeling).
ScoreAug is built on the standard diffusion training objective; the paper does not re-derive this and relies on it as background.
domain assumption For the transformations used, the score in the augmented noisy space is related to the score in the original space in a way the equivariant objective can exploit (the paper's theoretical analysis is assumed to cover the augmentation family used in experiments).
The abstract claims a theory of score relationships under general transformations; the validity of that relationship for non-invertible augmentations (crops, cutout) is the fragile premise that the corrupted text does not let us check.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Score Augmentation for Diffusion Models." pith.science (2026). https://pith.science/paper/TRHBEKPR

@misc{pith2026250807926,
  author       = {Pith},
  title        = {Pith review of: Score Augmentation for Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TRHBEKPR}},
  note         = {Machine review of arXiv:2508.07926}
}

read the original abstract

Diffusion models have achieved remarkable success in generative modeling. However, this study confirms the existence of overfitting in diffusion model training, particularly in data-limited regimes. To address this challenge, we propose Score Augmentation (ScoreAug), a novel data augmentation framework specifically designed for diffusion models. Unlike conventional augmentation approaches that operate on clean data, ScoreAug applies transformations to noisy data, aligning with the inherent denoising mechanism of diffusion. Crucially, ScoreAug further requires the denoiser to predict the augmentation of the original target. This design establishes an equivariant learning objective, enabling the denoiser to learn scores across varied denoising spaces, thereby realizing what we term score augmentation. We also theoretically analyze the relationship between scores in different spaces under general transformations. In experiments, we extensively validate ScoreAug on multiple benchmarks including CIFAR-10, FFHQ, AFHQv2, and ImageNet, with results demonstrating significant performance improvements over baselines. Notably, ScoreAug effectively mitigates overfitting across diverse scenarios, such as varying data scales and model capacities, while exhibiting stable convergence properties. Another advantage of ScoreAug over standard data augmentation lies in its ability to circumvent data leakage issues under certain conditions. Furthermore, we show that ScoreAug can be synergistically combined with traditional data augmentation techniques to achieve additional performance gains.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

[1]

�� ...

work page Pith review arXiv 2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

�� ...

work page Pith review arXiv 2025