acceptodds
Under review as a conference paper at ICLR 2027

CAGS: Class-Adaptive Guidance Strength for Training-Free Diffusion-Based Dataset Distillation

Abstract

Training-free diffusion-based dataset distillation methods such as MGD typically use a single classifier-free guidance (CFG) strength for all classes. This class-agnostic design overlooks substantial variation in class complexity: structurally diverse classes may be under-guided, whereas simpler classes can suffer from excessive sharpening. Our key insight is that class-wise generation statistics can reveal how strongly each class should be guided. Based on this observation, we propose CAGS (Class-Adaptive Guidance Strength), a training-free method that assigns a class-specific CFG strength using cluster entropy and intra-class variance. A systematic ablation and weight search over an initial four-factor design shows that these two factors consistently drive the gains, while mode count and inter-class separability can be removed. We further provide a theoretical explanation under a Gaussian-mixture model, where the optimal guidance satisfies and increases with both intra-class variance and entropy; consequently, the benefit of adaptation grows with cross-class heterogeneity. On ImageNet-100 at IPC=10, CAGS achieves 30.81% on ResNet-18, 29.74% on ResNetAP-10, and 25.55% on ConvNet-6, improving over MGD by 7.21, 3.94, and 2.15 percentage points, respectively, and outperforming other published baselines including IDC-1 and MinMaxDiff. In matched scaling experiments, the gains increase to 3.58–6.97 points on ImageNet-1000, while remaining near-neutral on a low-heterogeneity 10-class subset and reaching +5.28 points on ImageNette. CAGS also generalizes across generators, improving Stable Diffusion 1.5 results on Food-101 from 2.27% to 4.17%, with zero trainable parameters and negligible inference overhead.

Primary area:
generative models
Keywords:
Dataset Distillation, Diffusion Models, Classifier-Free Guidance, Adaptive Guidance, Data Condensation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_76XwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAGS, a training-free method for dataset distillation that assigns different classifier-free guidance strengths to different classes. The method estimates class complexity using cluster entropy and intra-class variance, and maps these statistics to class-specific guidance values through a sigmoid function. The authors also provide a theoretical analysis based on Gaussian mixture models to motivate the relationship between class complexity and guidance strength. Experiments show improvements over unguided generation and several published dataset distillation baselines.

Critical strengths
  • The method is simple and does not require training additional models. It can be incorporated into an existing diffusion-based generation pipeline with relatively little modification.
Critical weaknesses
  • Adaptive guidance has already been explored in diffusion models, and the proposed method does not introduce a new generation or optimization framework. The novelty mainly lies in applying this idea to dataset distillation.

  • One of the main claims is that class-adaptive guidance is better than using a single guidance strength. However, the main experiments only include fixed λ=0.10, while CAGS uses guidance values within [0, 0.06]. Since the fixed λ=0.10 baseline performs worse than unguided generation, it is not clear whether the improvement comes from class adaptation or simply from choosing a smaller guidance strength.

  • The authors derive an optimal guidance strength using a manually defined utility function that balances discriminative benefit and diversity cost. However, this utility is not directly connected to the downstream classification objective used in dataset distillation.

Questions for authors
  • In Appendix D.6, the proposed complexity score has only a weak correlation with per-class accuracy gain, whereas separability shows a stronger correlation. How do the authors reconcile this observation with the claim that entropy and intra-class variance are the key factors for guidance adaptation?

  • The related work discusses several existing adaptive CFG methods, but the experiments do not directly compare with them. Could the authors include at least one representative adaptive guidance baseline under the same dataset distillation protocol?

  • The method relies on VAE features and K-means clustering to estimate class complexity. How sensitive are the results to the feature extractor, clustering method, and number of clusters? Would a simpler variance-based heuristic achieve comparable performance?

4 comments

Thank you for recognizing the method's simplicity and ease of integration, and for your questions about novelty, controls and theory. We carefully checked the implementation and experiments and revised the manuscript, including the main overview figure, to clarify the method and address your concerns. We would greatly appreciate your consideration of the revised manuscript and look forward to further discussion.

1. Contribution relative to adaptive CFG (W1).

CAGS maps real-data cluster entropy and feature dispersion to class-specific mode-guidance coefficients, computed once without a learned predictor. With the generator frozen and CFG fixed at g=4, CAGS varies a separate update toward cached mode targets; adaptive CFG changes conditional/unconditional denoising. Our contribution concerns this distinct control variable for dataset distillation (Sections 2.3 and 3).

2. Is the gain due to class allocation (W2)?

CAGS improves over CFG-only across all three classifiers. Evaluation: ImageNet-100/IPC=10, final epoch 2000, generation seeds 100–104 × three classifier seeds.

MethodConvNet-6ResNet-18ResNetAP-10
CFG-only21.91 ± 0.6326.14 ± 0.5226.08 ± 0.72
CAGS24.15 ± 0.5430.03 ± 0.2629.12 ± 0.44
Fixed (CAGS mean)24.36 ± 0.5230.03 ± 0.3629.16 ± 0.40

Top-1 mean ± sample SD (%) across five generation replicates after averaging classifier seeds; architectures reuse datasets. Bold values mark the highest unrounded mean, not significance. pp denotes percentage points.

Paired gains over CFG-only are +2.245/+3.881/+3.041 pp, with Holm p=0.0299/0.00105/0.00666 (six comparisons per architecture). Fixed-mean differences have CIs spanning zero. Only ResNet-18 exceeds the three-permutation average (+0.260 pp, Holm p=0.00561), versus −0.231/−0.110 pp on ConvNet-6/ResNetAP-10 (Sections 5.2–5.4; Tables 1, 2, 6, 7).

Independent selection uses a 10% training holdout excluded from complexity/target estimation: 14 fixed and 13 CAGS candidates (81 fits). It selects fixed 0.03 and entropy/dispersion weights 0.25/0.75; 0.5/0.5 remains the default. The 300 reporting fits use seeds 200–204 × three classifier seeds, ImageNet-100/IPC=10/ConvNet-6, final epoch 2000. Default/selected CAGS gain +2.396/+2.368 pp over CFG-only (Holm p=0.0231/0.0372; 40-contrast family). Differences from selected fixed are −0.211/−0.239 pp (both Holm p=1); own-mean fixed and permutation-average contrasts are also inconclusive.

These controls show similar observed means to fixed guidance, without establishing an additional allocation benefit or statistical equivalence. Earlier validation-set use remains disclosed (Section 5.5; Appendix F.1; Tables 3, 8–10).

3. Would a simpler rule suffice (Q3)?

The two-factor default has the highest observed mean among these factor variants (seeds 200–204; same mean/SD convention). Mean matching controls for average guidance strength.

RuleRaw (%)Mean-matched (%)Holm p
Default CAGS24.68 ± 0.5424.68 ± 0.54—
Entropy24.22 ± 0.2124.62 ± 0.321.000
Dispersion24.56 ± 0.2524.56 ± 0.401.000
Four factors24.57 ± 0.3024.65 ± 0.211.000

Bold marks the highest mean per accuracy column. Holm p compares default with each matched rule (40-contrast family). None of the six raw/matched factor contrasts is significant. We retain the compact default without claiming both factors are necessary (Table 4; Tables 9–10). Dispersion is mean Euclidean distance to the class centroid; covariance trace is untested (Section 3.2).

4. Theory and complexity interpretation (W3, Q1).

The bounded quadratic analysis characterizes the oracle allocation gap under clipped optima, with a weighted-square identity for interior optima. It provides intuition, not a classification-risk derivation or CAGS guarantee (Section 4). Appendix B includes the binary entropy counterexample and removes unsupported general entropy-monotonicity and separability-invariance claims.

The separability association remains exploratory. Complexity–gain correlations against fixed-mean guidance change sign across seeds on every architecture (Figure 3); CFG-only comparisons test a different intervention. These results do not establish unique roles for entropy/dispersion.

5. Direct adaptive-CFG and estimator tests (Q2–Q3).

The 18-condition campaign (seeds 300–304 × three classifier seeds) is incomplete (Appendix D.3). It tests beta-CFG adapted to DiT/DDPM, alone and with CAGS, plus feature/PCA/clustering/K changes with raw and mean-matched coefficients, keeping mode targets fixed. Beta-CFG uses one frozen published preset, not an equal-tuning-budget search. We will report completed, audited five-generation results.

Anonymous code: repository.

We provide additional results on guidance and estimator choices to further address the concerns raised in your original review.

CAGS computes and caches guidance coefficients from cluster entropy and feature dispersion once, keeping CFG fixed and the generator frozen, with no added trainable parameters. The completed additional campaign (90 datasets; 270 classifier fits) adds paired replication and evaluations of the β-CFG adaptation and estimator choices.

1. Replicated gains and the β-CFG comparison (W2, Q2).

Default CAGS improves over CFG-only by +2.012 percentage points (pp), positive in all five generation replicates, with Holm p=0.044586.

Protocol: ImageNet-100, IPC=10, ConvNet-6, frozen DiT, 50 sampling steps, final classifier epoch 2000; all generation seeds 300-304 × classifier seeds 0-2, paired on A800. All rows were rerun in this cohort, separate from earlier cohorts and best-epoch results. Top-1 is mean ± sample SD across five generation means, each averaging three classifier seeds. Differences are default CAGS minus the row. Two-sided paired t-tests use Holm correction over all 34 planned contrasts.

ConditionTop-1 (%)CAGS minus row (pp)Holm p
CFG-only22.23 ± 0.30+2.0120.044586
Default CAGS24.25 ± 0.60——
Fixed: CAGS mean24.55 ± 0.53-0.3080.274678
Selected fixed 0.0324.34 ± 0.30-0.0891.000000
Selected CAGS24.15 ± 0.71+0.0931.000000
β-CFG adaptation12.56 ± 0.30+11.6850.000134
CAGS + β-CFG20.30 ± 0.37+3.9470.001665

The CFG-only gain has an unadjusted 95% paired t CI of [1.292,2.732] pp (n=5, df=4). Mean-matched fixed guidance is higher in all five generations: CAGS minus fixed is -0.308 pp, with an unadjusted CI of [-0.497,-0.119] and Holm p=0.274678. This cohort replicates the complete configuration's benefit over CFG-only; an additional allocation advantage over matched fixed guidance remains unestablished.

CAGS exceeds the tested β-CFG adaptation and CAGS + β-CFG by +11.685/+3.947 pp, both significant after correction. β-CFG uses the frozen public preset (scale=7.5, a=2, b=3, gamma=1), adapted from SD/DDIM to DiT/DDPM with three guided channels, guidance-norm normalization and no beta-density peak normalization. Ordinary CAGS and fixed controls retain CFG=4. This comparison concerns that preset and adaptation; matched selection budgets and comparable CFG-scale selection remain to be evaluated.

Selected rows reuse the training-holdout choices without retuning: fixed 0.03 and CAGS weights 0.25/0.75, versus default 0.5/0.5. Default/selected CAGS gain +2.012/+1.919 pp over CFG-only (Holm p=0.044586/0.064445); their direct paired difference is +0.093 pp (Holm p=1). The original default retains its observed CFG-only benefit without adopting the selected weights.

2. Feature, clustering and K sensitivity (Q3).

The same cohort and statistical convention apply. We compare raw coefficients and coefficients matched to the default's mean.

Complexity estimatorRaw Top-1 (%)Mean-matched Top-1 (%)
Default VAE / auto K24.25 ± 0.6024.25 ± 0.60
ResNet-18 / auto K24.00 ± 0.2924.28 ± 0.35
VAE / K=524.11 ± 0.6724.10 ± 0.70
VAE / K=1024.13 ± 0.4524.16 ± 0.53
VAE / PCA=16, auto K24.22 ± 0.6724.12 ± 0.41
VAE / Ward, K=1024.13 ± 0.9124.34 ± 0.84

Observed mean accuracies span 24.00-24.25% with raw coefficients and 24.10-24.34% after mean matching, ranges of 0.25/0.24 pp. Mode targets remain fixed; only complexity estimation and coefficient assignment change. Ward and K-means K=10 share K and PCA dimension; the default row is one shared reference.

All ten default-versus-estimator contrasts have Holm p=1 in the full 34-test family. The observed mean variation is limited across these tested choices. The dispersion-only control was reported earlier; this estimator table does not test a distinct covariance-trace heuristic.

3. Class-level evidence (Q1).

Across five generations, Pearson correlations with CAGS-minus-CFG-only gains are 0.133-0.312 for separability and -0.056 to 0.086 for complexity. Complexity correlations with CAGS-minus-fixed-mean gains range from -0.079 to 0.131. Each correlation uses 100 class accuracies after averaging classifier seeds within a generation. Separability has the stronger observed association with CFG-only gains; complexity correlations with the allocation contrast are small.

The additional experiments show a CFG-only gain across fresh generation seeds and limited observed mean variation across the tested estimators. These results support the practical value of CAGS as a simple, cached sampling rule in this setting. We hope its demonstrated utility and the explicitly stated comparison limits will both inform your assessment.

Thank you for your responses. I believe my original rating is fair and would prefer to keep it unchanged.

Thank you for considering our responses and clarifying your assessment. We respect your judgment and sincerely appreciate the time and effort you have devoted to reviewing our work. Your constructive comments have been valuable in guiding our revisions and improving the manuscript.

Reviewer_PrL8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper observes that training-free diffusion-based dataset distillation (MGD³) applies a single global CFG strength to all classes, while classes differ in complexity. It proposes CAGS: a one-time, off-line analysis of real data (VAE features + K-means) computes four per-class complexity factors, of which cluster entropy and intra-class variance are retained; these are mapped through a bounded sigmoid to a per-class guidance strength ω_c ∈ [0, 0.06]. The frozen DiT is then sampled with per-class CFG. A Gaussian-mixture analysis predicts ω*_c ∝ v_c/σ_c² and that the adaptivity gain grows with cross-class heterogeneity. On ImageNet-100 (IPC=10) CAGS reaches 30.81% / 29.74% / 25.55% (ResNet-18 / ResNetAP-10 / ConvNet-6), with larger gains at 1000 classes, and transfers to SD 1.5 on Food-101.

Critical strengths
  1. Well-scoped, genuinely cheap intervention. The method changes one scalar schedule, adds zero trainable parameters, and costs ~22 min of one-time cached analysis. This is a meaningful practical advantage over methods requiring real-image initialization or per-step OT matching.

  2. Unusually honest reporting of negative results. TAGS, IAST, the CAGS+D⁴M hybrid, and CAGS+herding are all reported as ineffective or sub-additive, and the paper offers a consistent explanation (adaptation along axes where the optimum is constant). This raises confidence in the positive results.

Critical weaknesses
  1. The headline comparison is inflated by protocol. The abstract advertises +7.21 over MGD³ on ResNet-18, but the paper's own matched unguided baseline (26.52%) already exceeds MGD³'s published 23.6% by ~2.9 points. The improvement attributable to CAGS is therefore ~+4.3, not +7.2. The same correction applies to ResNetAP-10 (+3.94 → ~+3.2). Comparisons against published numbers obtained under a different protocol should not headline the abstract.

  2. Internal numerical inconsistencies directly undermine the central design claim. Table 2 reports Entropy-only 25.31±0.62 and Full CAGS 24.24±0.26 (entropy alone better); Table 19 reports 24.35±0.47 and 25.29±0.27 (ordering reversed); Appendix F's text again cites a different pair. Since the two-factor "optimality" rests on a margin of +0.24 (well inside seed noise), these discrepancies must be resolved before the claim can be assessed.

  3. The theory is illustrative rather than predictive, and two empirical findings contradict it. The linear benefit term in U_c(ω) is assumed rather than derived, so ω* ∝ v/σ² follows from the modeling choice; Theorem 1's non-negativity is a variance decomposition. Meanwhile (a) inter-class separability is the strongest per-class predictor of gain (r=0.205, p=0.041) although Lemma 2 declares it irrelevant, and (b) Appendix H shows the four-factor design wins on the 44-class animal subset, contradicting "two-factor is uniformly optimal.

Questions for authors
  1. Please re-state the main results against the matched unguided baseline, and clarify whether MGD³'s published numbers were produced with its own fixed guidance value and augmentation protocol. What is the relationship between your ω ∈ [0, 0.06] parameterization and the CFG = 4.0 used in your D⁴M implementation?

  2. Table 2, Table 19, and Appendix F disagree on Entropy-only vs. Full CAGS (25.31/24.24 vs. 24.35/25.29). Which numbers come from the same generation run? If entropy alone matches the two-factor design within noise, what justifies the two-factor default?

4 comments

We appreciate your recognition of our low-cost approach and transparent reporting of negative results. We added controls, clarified the protocol, ablations and theory, and improved the overview figure for clarity. We hope you will read the revised manuscript and welcome further discussion.

1. What improves under matched evaluation (W1, Q1)? The abstract removes the +7.21 cross-protocol headline. Published MGD³ uses its own evaluation/augmentation; best-epoch results remain separate (Table 11; Appendix E). The table answers with ImageNet-100/IPC=10 results: all generation seeds 100-104, three classifier seeds each, final epoch 2000.

MethodConvNet-6ResNet-18ResNetAP-10
CFG-only21.91 ± 0.6326.14 ± 0.5226.08 ± 0.72
CAGS24.15 ± 0.5430.03 ± 0.2629.12 ± 0.44
Fixed (CAGS mean)24.36 ± 0.5230.03 ± 0.3629.16 ± 0.40
MGD³ official21.90 ± 0.73——

Top-1 mean ± sample SD (%): average classifier seeds first, then summarize five generation replicates; architectures reuse datasets. Bold marks highest accuracy by unrounded mean, not significance. pp denotes percentage points.

CAGS improves over CFG-only in every replicate: +2.245/+3.881/+3.041 pp, Holm p=0.0299/0.00105/0.00666 (six comparisons per architecture). The fixed-mean differences have CIs spanning zero (Sections 5.2–5.4; Tables 1, 2, 6, 7).

The common-protocol MGD³ rerun shares the split, targets, paired noise and evaluator, retaining coefficient 0.1 and 24 guided updates versus CAGS's 25. CAGS gains +2.251 pp (unadjusted 95% CI [1.533,2.969], Holm p=0.00288; three comparisons). This tests whole configurations, not allocation alone or published-table reproduction (Section 5.2; Table 2). CAGS and our D⁴M-style code keep CFG at g=4; the coefficient in [0,0.06] instead scales latent mode attraction. CFG-only disables that attraction while retaining conditioning (Section 3; Appendix C).

Permutation controls limit attribution: only ResNet-18 exceeds the three-permutation average (+0.260 pp, Holm p=0.00561); ConvNet-6/ResNetAP-10 give -0.231/-0.110. Non-significance is not equivalence (Section 5.4).

2. Which factor results are reliable (W2, Q2)? Paired experiments do not establish two-factor optimality. This complete factor comparison uses ImageNet-100/IPC=10/ConvNet-6, final epoch 2000, all seeds 200-204 × three classifier seeds and the same mean/SD convention.

RuleRaw (%)Mean-matched (%)Holm p
Default CAGS24.68 ± 0.5424.68 ± 0.54—
Entropy24.22 ± 0.2124.62 ± 0.321.000
Dispersion24.56 ± 0.2524.56 ± 0.401.000
Four factors24.57 ± 0.3024.65 ± 0.211.000

Bold marks the highest mean per column among displayed factor rules; default is the shared reference. Holm p compares default with each matched rule (all 40 contrasts). None of the six raw/matched contrasts is significant; matching mean guidance reduces the CAGS–entropy gap from 0.457 to 0.057 pp (Section 5.5; Table 4; Tables 9-10). Nominal ranking does not prove factor necessity.

The 25.31/24.24 best-epoch entropy/four-factor pair is traced to separate logs, without verified paired generation (Table 12). Unresolved 24.35/25.29 references are withdrawn (Tables 25-26); neither assumed generation noise nor the old +0.24 margin establishes optimality.

3. Default versus holdout-selected configuration (Q2). The 0.5/0.5 default was fixed before reporting. A 10% training holdout, excluded from complexity/target estimation, selects among 14 fixed and 13 CAGS candidates (81 fits): entropy/dispersion weights 0.25/0.75 and fixed 0.03. All 300 reporting fits on seeds 200-204 are complete: selected CAGS 24.65%, selected fixed 24.89%, CFG-only 22.29%.

Default/selected CAGS gain +2.396/+2.368 pp over CFG-only (Holm p=0.0231/0.0372, full 40-contrast family); neither significantly exceeds its own mean-matched or selected fixed control. Prior use of the reporting validation set during development remains disclosed (Section 5.5; Appendix F.1; Tables 3, 8-10).

4. Theory and contrary observations (W3). Section 4 assumes a bounded quadratic surrogate with clipped optima and an interior weighted-square identity, without identifying its coefficients with entropy/dispersion or deriving classification risk. Appendix B provides the entropy counterexample and covariance/marginal-density corrections; general entropy monotonicity and separability irrelevance are withdrawn. Figure 3 reports sign-changing complexity-gain correlations against fixed guidance. Appendix H retains the exploratory animal-only best-epoch four-factor advantage; neither it nor correlations identify causal factors or uniformly optimal weights.

Adaptive-CFG and estimator tests remain incomplete (Appendix D.3). We will supplement audited results as they complete.

Manuscript's anonymous code link: repository.

Building on our earlier response to your original review, we report a newly completed paired evaluation and further clarify the scope of our conclusions.

CAGS uses class statistics computed once to set cached mode-guidance coefficients, keeping CFG fixed and the pretrained generator frozen without adding trainable parameters. The completed Appendix D.3 cohort (90 generated datasets; 270 classifier fits) strengthens the evidence for this sampling intervention: the unchanged default improves over CFG-only in all five new generation replicates (+2.012 pp; Holm p=0.044586).

1. Paired replication and attribution (W1, Q1).

Protocol: ImageNet-100, IPC=10, ConvNet-6, frozen DiT, 50 sampling steps, final classifier epoch 2000; all generation seeds 300-304 × classifier seeds 0-2, paired on A800. All rows belong to this cohort, separately from earlier cohorts and best-epoch results. Data preparation, mode targets, sampling noise and evaluation are shared. CFG is g=4; MGD³ retains its official coefficient 0.1 and 24 guided updates, versus CAGS's 25.

Top-1 is mean ± sample SD across five generation means, each averaging three classifier seeds. Differences are default CAGS minus the row in percentage points (pp). Paired 95% t CIs are unadjusted (n=5, df=4); two-sided paired t-tests use Holm correction over all 34 planned contrasts.

ConditionTop-1 (%)CAGS minus row (pp)95% CI (pp)Holm p
CFG-only22.23 ± 0.30+2.012[1.292,2.732]0.044586
Default CAGS24.25 ± 0.60———
Fixed: CAGS mean24.55 ± 0.53-0.308[-0.497,-0.119]0.274678
Selected fixed 0.0324.34 ± 0.30-0.089[-0.547,0.368]1.000000
Selected CAGS24.15 ± 0.71+0.093[-0.318,0.504]1.000000
MGD³ official window22.32 ± 0.56+1.929[0.974,2.884]0.137561

Mean-matched fixed guidance is higher in all five generations. The CAGS-minus-fixed difference is -0.308 pp: its unadjusted CI excludes zero, while the full-family Holm test does not reject (p=0.274678). The cohort replicates the full configuration's benefit over CFG-only; an additional allocation advantage over the fixed controls remains unestablished.

The MGD³ difference is positive in all five replicates (+1.929 pp), but is not significant after this cohort's correction (Holm p=0.137561). The earlier significant comparison used a separate cohort and correction family. Both are configuration-level comparisons retaining coefficient and update-window differences, rather than isolated tests of class allocation.

2. Frozen choices and class-level interpretation (Q2, W3).

Selected rows reuse the training-holdout choices unchanged: fixed guidance 0.03 and entropy/dispersion weights 0.25/0.75, versus default 0.5/0.5. No configuration was retuned on these reporting outcomes.

Default and selected CAGS have mean gains over CFG-only of +2.012/+1.919 pp (Holm p=0.044586/0.064445). Their direct difference is +0.093 pp (Holm p=1). The unchanged default retains its observed CFG-only benefit on fresh generation seeds. We retain the original two-factor weights as the predefined reference across cohorts. The historical provenance gaps discussed in our earlier reply remain unresolved.

Across the five generations, Pearson correlations with CAGS-minus-CFG-only gains are 0.133-0.312 for separability and -0.056 to 0.086 for complexity; complexity correlations with CAGS-minus-fixed-mean gains are -0.079 to 0.131. Each correlation uses 100 classes after averaging classifier seeds within a generation. The complexity correlations with the allocation contrast are small in this cohort. The theoretical model remains illustrative; the performance conclusions here rest on the paired experiments.

CAGS contributes an explicit, cached rule for incorporating class statistics into mode guidance, accompanied by paired evidence for its configuration-level benefit. The unchanged setting retains its CFG-only gain on fresh generation seeds. The controls refine the scope of this contribution, and we hope the additional evidence and clarified claims will be considered alongside the remaining limitations.

Thank you for your responses, which solve my concerns. I will maintain my positive rating.

Thank you for your thoughtful review and positive assessment. We are pleased that our responses have addressed your concerns, and we appreciate your constructive feedback, which helped us improve the clarity and rigor of the paper.

Reviewer_7hx5OpenReview ↗
Rating
3(+1)
Confidence
2(−2)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses a limitation in training-free, diffusion-based dataset distillation: the use of a globally fixed classifier-free guidance (CFG) scale for all classes. The authors hypothesize that structurally diverse classes require different guidance strengths than simpler classes. To solve this, they propose CAGS, a training-free method that computes a complexity score for each class based on intra-class variance and cluster entropy, and dynamically assigns a class-specific CFG scale. The empirical design is strongly supported by a theoretical analysis using a Gaussian-mixture model, which proves that the optimal guidance scale is proportional to intra-class variance and positively correlated with cluster entropy. Furthermore, the theory predicts an adaptivity gap, showing that performance gains depend on cross-class heterogeneity rather than mere class count. Extensive experiments validate these claims, showing significant improvements over SOTA methods across multiple architectures and datasets, while adding zero trainable parameters and negligible computational overhead.

Critical strengths
  1. Simplicity and negligible cost. CAGS changes only one scalar per class, adds no trainable parameters, and needs about 22 minutes of one-time analysis. It plugs directly into existing training-free diffusion distillation pipelines. If the gains are truly due to per-class adaptation, this is a cheap and easily reproducible improvement.
  2. Broad evaluation with honest reporting of negative results. The experiments span 10 to 1000 classes and three architectures. Gains over the authors' own unguided baseline are consistently positive on IN-100/200/1K. The paper also reports the failures of TAGS and IAST, the neutral 10-class result, and the IPC=1 degradation, which helps delimit where the method applies.
Critical weaknesses
  1. The evidence does not isolate per-class adaptation as the source of the gains, and the effect size is comparable to generation noise.
    • Missing key controls. Because of the sigmoid mapping, ρc\rho_c lies at most in [σ(−0.6),σ(2.4)]≈[0.35,0.92][\sigma(−0.6), \sigma(2.4)] ≈ [0.35, 0.92], so every class actually receives λc\lambda_c of roughly 0.021–0.055. Quartile means in Table 11 range only from 0.028 to 0.041, with a std of 0.0049. Within such a narrow range, the gains may come mainly from the overall guidance level rather than from how guidance is allocated across classes. The only fixed-λ control on 100 classes uses λ=0.10, which lies outside the CAGS range. Two controls are missing: a fixed λ equal to the CAGS mean, and a random permutation of λcλ_c across classes. Notably, the near-constant separability-only configuration already gains about 1 point over unguided (Table 2).
    • Generation randomness is not included in the reported variance. The reported stds reflect only classifier-training seeds. The same four-factor configuration obtains 25.29 (Tables 7/19), 24.24 (Table 2), and 23.27 (Table 12, λmax=0.06) on three independently generated datasets. Entropy-only obtains 25.31 in Table 2 and 24.35 in Table 19. This 1–2 point between-generation variation is of the same order as the main gain (+2.65), and it reverses the factor-ablation ranking across tables. The t-tests and effect sizes in Table 3 therefore substantially overstate the statistical evidence.
    • Comparisons with published baselines conflate protocol differences. The authors' unguided baseline already reaches 26.52 on ResNet-18, above MGD3's published 23.6. Most of the headline "+7.21 over MGD3" therefore comes from the training protocol rather than from CAGS. Under DMGD's setting (Table 17), CAGS (62.00 on ImageNette, 38.86 on ImageWoof) falls below MGD3's published 66.4 and 40.4. The authors attribute this to heavier augmentation, but they do not reproduce MGD3 under their own protocol.
  2. The theory does not support the design, contains errors, and is contradicted by some of the paper's own empirical results.
    • The monotonicity claim does not hold. Propositions 1 and 2 claim that σ2c=Varπ[logπk]\sigma^{2}c = Var{π}[log π_{k}] decreases monotonically in H_c, so that λ*_c increases with H_c. This does not hold in general. Take K=2, which the paper says covers 94% of classes: σ² = p(1−p)·(log(p/(1−p)))², which vanishes both at p=0.5 (maximum entropy) and as p→0 (minimum entropy), so it is non-monotone in H. The Schur-convexity argument in Appendix B.2 is not rigorous. In addition, λ*_c→∞ as σ²_c→0, which conflicts with the small-λ regime the analysis assumes.
    • The utility function is postulated, and the implementation is linked to it only through monotonicity. The benefit term (1+λ)v_c is assumed rather than derived; the trace of the sharpened covariance is in fact v_c/(1+λ). The diversity cost accounts only for mode reweighting and ignores covariance shrinkage. The implementation, a sigmoid of a sum of normalized features, is not v_c/σ²_c. The only remaining link between theory and implementation is monotonicity, which is itself problematic per (a).
    • Lemma 2 follows from its own assumption, and the per-class analysis contradicts it. Lemma 2 is obtained by absorbing the p(x)^(−λ) term into the normalization, yet this term is exactly where inter-class information enters. Empirically, the per-class analysis in Appendix D.6 finds that the only factor significantly correlated with per-class gain is separability. Entropy shows almost no correlation, and the combined complexity score is not significant either.
  3. Additional inconsistencies
    • The legend of Figure 6 (22.79% / 25.7%) disagrees with its caption (22.90% / 25.55%).
    • The opening of Appendix F states that entropy-only beats the four-factor design by 1.07 points, but Tables 18 and 19 in the same appendix show the opposite.
    • Appendix I states that multiple factors help at 100 classes, which contradicts the main text.
Questions for authors
  1. How do you explain the values of 24.24 / 25.29 / 23.27 for the same four-factor configuration in Tables 2, 7, and 12?
  2. Were the weights, s, c0, and the λ range selected on the same validation set used for reporting? Can you instead select them on a held-out split of the training set and then report on the validation set?
  3. What exactly is "Unguided"? Does it retain MGD3's mode discovery and mode guidance? What does MGD3 achieve when reproduced under your protocol? D4M uses λ_CFG=4.0 (Appendix E) while CAGS uses λ≤0.06; are the two CFG conventions the same?
  4. The K=2 case is a counterexample to the claimed monotonicity of σ² in H in Proposition 1. How will you revise the theory?
4 comments

Thank you for your thorough and insightful review. We added controls, corrected the theory and clarified the overview figure. We hope you will read the revised manuscript and welcome further discussion.

1. Controlled gains and class allocation (W1, Q3).

CAGS improves over CFG-only on all three classifiers and over the official MGD³ configuration on ConvNet-6. Evaluation: ImageNet-100, IPC=10, final epoch 2000, all generation seeds 100–104 × three classifier seeds.

MethodConvNet-6ResNet-18ResNetAP-10
CFG-only21.91 ± 0.6326.14 ± 0.5226.08 ± 0.72
CAGS24.15 ± 0.5430.03 ± 0.2629.12 ± 0.44
Fixed (CAGS mean)24.36 ± 0.5230.03 ± 0.3629.16 ± 0.40
MGD³ official21.90 ± 0.73——

Top-1 mean ± sample SD (%) over five generation seeds after averaging classifier seeds; architectures reuse datasets. Bold numbers mark the highest unrounded mean, not significance; pp denotes percentage points.

CAGS gains +2.245/+3.881/+3.041 pp over CFG-only (Holm p=0.0299/0.00105/0.00666; six contrasts per architecture), positive for all five generation seeds. On ResNet-18, it also exceeds the average of three prespecified permutations by +0.260 pp (Holm p=0.00561), supporting this allocation relative to the tested permutations. ConvNet-6/ResNetAP-10 differences are −0.231/−0.110 pp; fixed-mean differences have CIs spanning zero (Sections 5.2–5.4).

Against MGD³, CAGS gains +2.251 pp (95% CI [1.533,2.969], Holm p=0.00288; three contrasts). The split, targets, noise and evaluator are shared; MGD³ retains coefficient 0.1 and 24 guided updates versus CAGS's 25. This measures a configuration-level gain. The cross-protocol +7.21 headline and unsupported augmentation attribution are removed (abstract; Appendix E).

2. Holdout-based selection and fresh-seed evaluation (Q2).

The original default retains its gain over CFG-only on fresh seeds without adopting holdout-selected weights. A 10% training holdout, excluded from complexity/target estimation, selects among 14 fixed and 13 CAGS candidates (81 fits). It selects fixed 0.03 and entropy/dispersion weights 0.25/0.75; the default remains 0.5/0.5, with slope/center/bound 3/0.6/0.06 unchanged. Key controls from 300 reporting fits: ImageNet-100, IPC=10, ConvNet-6, final epoch 2000, seeds 200–204 × three classifier seeds.

MethodTop-1 (%)Gain vs CFG-only (pp)Holm p
CFG-only22.29 ± 0.19——
Default CAGS24.68 ± 0.54+2.3960.0231
Selected CAGS24.65 ± 0.51+2.3680.0372
Fixed: default mean24.74 ± 0.25——
Fixed: selected mean24.93 ± 0.66——
Selected fixed 0.0324.89 ± 0.47——

Same reporting rules; Holm covers all 40 contrasts; dashes denote tests not reported here. CAGS means are similar to mean-matched and selected fixed guidance; these controls leave the additional allocation benefit unresolved and do not establish statistical equivalence (Section 5.5; Appendix F.1). CIs are unadjusted. Earlier development used the reporting validation set; this limitation remains disclosed.

3. Method and theory clarified (W2, Q3–Q4).

CAGS adjusts mode attraction in [0,0.06] while keeping CFG at g=4. CFG-only retains conditioning and disables mode attraction. The sigmoid is corrected to σ(s(Rc−c0))\sigma(s(R_c-c_0)), matching the code (Section 3; Appendix C).

Section 4's bounded quadratic surrogate explains the potential cost of sharing one coefficient across classes with different utility optima. Under its assumptions, it gives clipped optima and an interior weighted-square gap. This provides conditional intuition, without deriving the entropy/dispersion mapping, classification risk or sigmoid optimality.

Your binary counterexample is correct. Section 4 and Appendix B withdraw entropy monotonicity, Schur-convexity and separability invariance. The marginal-density factor cannot generally be absorbed into normalization; Gaussian powering contracts covariance, while the retained discrete KL concerns mode weights only. Figure 3's correlations remain exploratory, without causal claims.

4. Updated evidence and record clarification (W3, Q1).

The gains reported above come from the new controlled evaluations. The 25.31/24.24 best-epoch values trace to separate logs; paired generation remains unverified. The 25.29/24.35 references, 23.27 range table and former Figure 6 were withdrawn; their provenance remains unresolved (Tables 12, 25, 26).

Fresh raw/mean-matched ablations compare the two-factor default with entropy-only, dispersion-only and four-factor rules; none differs significantly from the default (Table 4; Tables 9–10). We retain the compact two-factor default and report the animal-subset reversal (Appendix H).

Adaptive-CFG and estimator tests remain incomplete (Appendix D.3). We will supplement completed, audited results.

Anonymous code: repository.

To further address the questions in your original review, we supplement our earlier response with the newly completed control experiments.

CAGS computes class-specific mode-guidance coefficients once from cluster entropy and feature dispersion, keeping CFG fixed and the pretrained generator frozen, with no additional trainable parameters. The completed Appendix D.3 campaign (90 datasets; 270 classifier fits) evaluates this rule on fresh generation seeds and adds the controls discussed below.

1. Paired replication and the allocation control (W1.1-W1.3; Q3).

Default CAGS improves over CFG-only in all five generation replicates: +2.012 pp, significant after Holm correction. The fixed-guidance controls distinguish this full-configuration benefit from the additional effect of class-specific allocation.

Protocol: ImageNet-100, IPC=10, ConvNet-6, frozen DiT, 50 sampling steps, final classifier epoch 2000; generation seeds 300-304 × classifier seeds 0-2, paired on A800. All rows were rerun with shared data preparation, mode targets, sampling noise and evaluation. This cohort is separate from earlier cohorts and best-epoch results. CFG is g=4; MGD³ retains its official coefficient 0.1 and 24 guided updates, versus CAGS's 25.

Top-1 is mean ± sample SD across five generation means, each averaging three classifier seeds. The statistical unit is the generation replicate. Differences are default CAGS minus the row in percentage points (pp). Paired 95% t CIs are unadjusted (n=5, df=4); two-sided paired t-tests use Holm correction over all 34 planned contrasts.

ConditionTop-1 (%)CAGS minus row (pp)95% CI (pp)Holm p
CFG-only22.23 ± 0.30+2.012[1.292,2.732]0.044586
Default CAGS24.25 ± 0.60———
Fixed: CAGS mean24.55 ± 0.53-0.308[-0.497,-0.119]0.274678
Selected fixed 0.0324.34 ± 0.30-0.089[-0.547,0.368]1.000000
Selected CAGS24.15 ± 0.71+0.093[-0.318,0.504]1.000000
MGD³ official window22.32 ± 0.56+1.929[0.974,2.884]0.137561

Mean-matched fixed guidance is higher in all five replicates. CAGS minus fixed is -0.308 pp; its unadjusted CI excludes zero, whereas the full-family Holm test does not reject (p=0.274678). The new cohort replicates the complete configuration's gain over CFG-only; an additional advantage over matched fixed guidance remains unestablished.

The MGD³ difference is positive in all five replicates (+1.929 pp), but is not significant after this cohort's correction (p=0.137561). The earlier significant comparison used a separate cohort and correction family. Both compare sampling configurations with their coefficient and update-window differences retained.

2. Transfer of frozen selection choices (Q2).

Selected rows reuse the training-holdout choices unchanged: fixed 0.03 and entropy/dispersion weights 0.25/0.75, versus the default 0.5/0.5. No selection used these new reporting outcomes.

Default and selected CAGS have positive mean gains over CFG-only (+2.012/+1.919 pp; Holm p=0.044586/0.064445). Their direct difference is +0.093 pp (Holm p=1), and selected CAGS minus selected fixed is -0.183 pp (Holm p=1). The default retains its observed CFG-only benefit without adopting the selected weights. These fresh seeds extend generation-level replication; the reporting validation split also informed earlier method development.

3. Class-level associations and their scope (W2.3).

Across the five generations, Pearson correlations between complexity and CAGS-minus-fixed-mean gains range from -0.079 to 0.131. For CAGS-minus-CFG-only gains, correlations are 0.133-0.312 for separability and -0.056 to 0.086 for complexity. Each correlation uses 100 classes after averaging classifier seeds within a generation.

The complexity-score correlations with matched-fixed gains are small in this cohort. We report these descriptive class-level associations alongside the paired comparisons that assess the allocation rule's performance.

The new cohort extends the evidence for CAGS beyond the earlier generation seeds: its unchanged configuration improves over CFG-only in all five new replicates. Computing class coefficients once provides an explicit way to incorporate class statistics into sampling. The paired controls support this configuration-level benefit and define the scope of the allocation claim. We hope these demonstrated benefits, considered alongside the stated limitations, help clarify the contribution of CAGS.

Thanks for the additional results. Most of my concerns have been solved. I have updated my rating.

Thank you very much for your thoughtful and thorough review, and for taking the time to examine our additional results and update your assessment. We sincerely appreciate the care and effort you have devoted to reading our paper. Your constructive comments have helped us identify areas for improvement and strengthen our work.

We will carefully incorporate the additional results and clarifications into the revised manuscript and continue to refine the work. We will make our code publicly available to support reproducibility, and we remain firmly committed to the integrity, accuracy, and transparent reporting of our data and experimental results. We take these commitments seriously and will follow through on them.

Thank you again for your time, careful consideration, and valuable feedback.

Comments on the paper

2 comments

Dear reviewers,

The authors have submitted their response. Please carefully check whether the new content has addressed your concerns, particularly the fairness of the numerical comparison, theoretical contradiction, guidance variation range, and novelty against prior work. If not, please follow up with the authors ASAP so that the authors can further respond before the discussion period ends (Nov 18).

Meanwhile, I also encourage the authors to improve the clarity of their response, which would help communication with reviewers.

Best,

AC

Dear Area Chair,

Thank you for overseeing the review process. We carefully revisited the manuscript in response to the reviewers’ comments. As the discussion concludes, we would like to summarize CAGS’s contributions and the concrete improvements made during the rebuttal.

1. A simple mechanism for incorporating class structure into diffusion-based dataset distillation.
CAGS uses cluster entropy and intra-class feature dispersion to compute class-specific mode-guidance coefficients before sampling. These coefficients guide attraction toward cached mode targets while keeping classifier-free guidance fixed and the pretrained generator frozen. This provides a concrete way to incorporate class-level statistics into the sampling process without introducing additional trainable parameters.

2. Consistent empirical gains in the revised core evaluation.
On ImageNet-100 at ten images per class, paired final-epoch evaluations across five generation seeds and three classifier seeds show gains over CFG-only of 2.25, 3.88, and 3.04 percentage points on ConvNet-6, ResNet-18, and ResNetAP-10, respectively. The gains are positive in every generation replicate for each classifier. These results support the practical utility of the CAGS sampling configuration across the three evaluated architectures.

3. Stronger controls addressing the source of the gains.
Following the reviewers’ questions, we added mean-matched fixed guidance, coefficient-permutation controls, and a comparison with the official MGD³ sampling configuration under a common evaluation protocol. The latter shows a 2.25-percentage-point gain on ConvNet-6. We explicitly distinguish this configuration-level improvement from the effect of class-specific allocation; an additional advantage over mean-matched fixed guidance is not established. These controls make the empirical contribution more precisely defined.

4. Improved parameter selection and experimental reporting.
We added a training-holdout selection procedure, excluded the holdout from class-statistic and mode-target estimation, and evaluated the selected configurations on fresh generation seeds. We also clarified sampling windows, hyperparameters, and statistical procedures, reported uncertainty across generation replicates, and separated best-epoch results from fixed final-epoch evaluations. Unsupported cross-protocol comparisons were removed.

5. Corrected theoretical analysis and clearer presentation.
We corrected the theoretical analysis and explicitly stated its assumptions and scope. We clarified the distinction between mode guidance and classifier-free guidance, aligned the mathematical specification with the implementation, and revised the main overview figure to make the method easier to follow.

The red highlighting records manuscript revisions, including clearer explanations, additional evidence, wording refinements, and necessary corrections. It should therefore not be interpreted as indicating that every highlighted passage was erroneous in the original submission.

We respectfully ask that your final assessment consider CAGS’s concrete methodological contribution, its demonstrated empirical benefits, and the strengthened evidence and clarified scope in the revised manuscript. We appreciate the reviewers’ constructive feedback and your careful consideration of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.