acceptodds
Under review as a conference paper at ICLR 2027

Diffusion-Informed Single-pass Classifier

Abstract

Classifiers built from diffusion models (i.e. diffusion classifiers) are reported to be robust and well aligned with human perception. However, a single prediction requires about forward passes, which is far too costly for deployment. We ask where this alignment (shape bias, out-of-distribution accuracy, and error consistency with human mistakes) comes from, and whether an ordinary single-pass discriminative classifier can have it too. We carefully decompose the human alignment gap between diffusion classifiers and discriminative classifiers that is reported in the literature into the axes along which their designs differ: training data, resolution, aggregation of evidence across noise levels and training objective (generative vs discriminative). We isolate “aggregation” as an important contributor (more than training data or the generative objective alone) to this alignment. We measure how much each noise level contributes to a diffusion classifier's prediction through reductions in KL divergence to its final prediction. The normalized average contributions define a noise-sampling distribution, , and their cumulative sum, , provides a label-smoothing schedule. We use these two curves to set the noise levels and label confidence when training an ordinary single-pass classifier (e.g., a ResNet-50), transferring aggregation with no change to architecture or inference cost; and, unlike naive noise augmentation, at no cost in accuracy. Our recipe, DISC (Diffusion-informed single-pass classifier), matches the diffusion classifier's out-of-distribution accuracy, exceeds its error consistency, narrows its shape-bias gap, and improves robustness to corruptions and distribution shifts without losing clean accuracy. The curve also serves as an adaptive guidance schedule, improving diffusion sampling quality.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
generative classifiers, human-alignment, robustness, diffusion, augmentation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ABdNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the root cause of the high consistency between diffusion classifiers and human perception, as well as how the ability of diffusion classifiers to align with humans can be extended to ordinary classifiers. Therefore, the author proposes DISC, a classifier based on curve-driven methods, which maintains accuracy on clean data while narrowing the gap in shape deviation compared to diffusion classifiers, and improves robustness to data corruption and distribution shifts. Additionally, the wgenw_{gen} curve performs better than the constant guidance in ADM-64 and DiT-XL/2 samplers.

Critical strengths
  1. DISC transfers the useful properties of complex and resource-consuming generative models to ordinary classifiers. Moreover, DISC achieved an accuracy rate of 78.6% on clean data, and has demonstrated excellent performance in multiple scenarios with distribution shifts. The experimental results show that DISC can improve corrosion robustness while maintaining a high accuracy rate. It has good practicality and application value.

  2. The author conducted extensive ablation experiments, conducting experiments under multiple seeds and making comparisons. The experimental results are convincing, and the author clearly stated that the contribution of a single sample could be negative, and provided curve comparisons for different sample sizes.

Critical weaknesses
  1. The paper directly converts the KL contribution of the noise layer to the teacher's final prediction into the label smoothing coefficient. However, the meanings of the two are not equivalent: the KL contribution measures the degree to which the running prediction is close to the teacher's final output, while the label confidence should reflect the reliability of the category information in the current noise input. The position where the decision is formed is not necessarily equal to the position where the information is reliable, and the two are not naturally equivalent. It is suggested that the authors clarify the basis and applicable conditions of this mapping.

  2. The paper uses the attribution curve itself as the noise sampling distribution, and uses its cumulative quantity as the label smoothing coefficient. The former determines the training allocation at each noise level, while the latter determines the confidence level of the supervisory signal. The two functions are different, and there is insufficient basis for this. It is suggested that the authors explain separately the two design choices. Why can the same attribution curve provide reasonable basis for both of them at the same time?

  3. How many independent noise samples were used for the single noise level baseline? Can it provide the result of averaging scores with a fixed noise level and the same number of evaluation times for the model?

  4. Have you compared the noise distribution after adjusting parameters within the same budget, such as log-normal sampling, combined with SNR label smoothing or fixed label smoothing? For Table 3, how did the original SNR label scheme perform at a 60% noise ratio?

Questions for authors
  1. Under the same sampler, number of steps and parameter adjustment budget, how does the curve in this paper compare with the one guided by a finite interval?
  2. Please clarify how the input normalization during DISC training is mapped to the sampling state of ADM, and whether the corresponding gradient scaling is included in the implementation.
3 comments

Thank you for the careful review. We answer W1 and W2 first; revisions are mainly led by these.

W1 and W2. The basis of the method. We added it to the Introduction ("How an ordinary classifier can have it", L100–143) and to Sec. 3.1 (L246–254). In the paper's words:

  1. A diffusion classifier operates on a score field ∇xlog⁡p(xσ∣c)\nabla_x \log p(x_\sigma \mid c), which is ideally learned over the whole input space.
  2. From this score field, a diffusion classifier models a scalar field p(c∣x)p(c \mid x) over the whole input space, and each query is an integration (aggregation) over a trajectory in this field, from the noisiest to the clean level.
  3. To transfer this class-posterior field to a single-pass discriminative classifier, we keep diffusing the data points during training. At each noised point, we need to know how far the diffusion classifier's decision has formed by that noise level, and set the label confidence accordingly.
  4. Since querying the diffusion classifier at every training point is infeasible, we estimate the expected log class-posterior log⁡p(c∣xσ)\log p(c \mid x_\sigma) on a small subset of the training set, as a function of the noise level σ\sigma only and normalized between the noisiest and the clean level. We denote it ρ(σ)\rho(\sigma) and calculate it as the cumulative sum of the normalized average contributions wgenw_{\mathrm{gen}} of each noise level to the final prediction.
  5. Each training noise level is sampled from wgenw_{\mathrm{gen}}. At that level, ρ\rho interpolates between the one-hot ground-truth label and the uniform vector, as in label smoothing.
  6. The curves are defined through the diffusion classifier's own final prediction, whereas ρ\rho will be applied to the one-hot labels of the true classes. The two coincide under two assumptions. (A1) Every image has one-hot true label cc. (A2) The diffusion classifier is ideal: its final prediction is the one-hot true label, and it recovers this label monotonically as levels are added. Then ρ\rho is the expected posterior of the true class, normalized between the noisiest level and the clean image.

W1. Points 3 and 4 above give the basis, and point 6 the conditions. We checked them (new App. B): the diffusion classifier is 96.5% accurate on the images used for the curves. The curves recomputed with gains in the log-probability of the GT label, instead of reductions in KL to the full prediction, are close to ours (Fig. 6b,c), and a classifier trained with them gives results similar to the original DISC (Table 5; L864–869). On the reliability of the class information in the noisy input: the gap between the class-conditional and the unconditional denoising error, which prior work links to mutual information, overlaps wgenw_{\mathrm{gen}} (Fig. 6b; L875–882).

W2. Points 4 and 5 above: both curves come from one measurement. wgenw_{\mathrm{gen}} is the contribution of each noise level, so we draw more training samples where the decision is formed. ρ\rho is its cumulative sum, which sets the label confidence. There is no free parameter. Fig. 2b ablates the importance of combining the two visually on the checkerboard. Table 2 ablates this on ImageNet: removing wgenw_{\mathrm{gen}} lowers clean accuracy from 73.6 to 64.0, and removing ρ\rho lowers it to 71.0.

W4. Log-normal sampling; SNR labels at 60% noise.

  • Log-normal sampling (new Table 5 and Fig. 8; App. C, L910–919). With the log-normal that EDM tuned for its own training as the sampler, and the recipe and budget of DISC, clean accuracy is 72.7 with our labels, 72.2 with SNR labels, and 71.1 with fixed smoothing, against 73.6 for DISC (AutoAttack is higher for these rows: 30.5 against 18.2). We also fitted a log-normal to our curves: our curves fit well to a log-normal distribution, but with parameters that no one would include in a grid search, N(−1.01,1.332)\mathcal{N}(-1.01, 1.33^2) for wgenw_{\mathrm{gen}} and N(−1.05,1.552)\mathcal{N}(-1.05, 1.55^2) for wgenw_{\mathrm{gen}} (GT-logprob).
  • Limited ablation in Table 3. We have now extended the ablations, including the matched baselines at 60% noise setting.

W3. Noise draws at the single noise level. In the submission, the single-level rows of Table 1 used one noise draw; they now use 250, as many evaluations as 250 levels. Aggregation still adds +7.6 / +10.2 / +4.7 in shape bias / OOD accuracy / error consistency for EDM, and +5.0 / +17.2 / +9.5 for Stable Diffusion.

Q1. Guidance limited to an interval. We added it on DiT-XL/2 with the same sampler, number of steps and samples, sweeping the guidance weight for each method (new Table 6; App. D, L1001–1007). The interval alone reaches FID 1.99, better than wgenw_{\mathrm{gen}} alone (2.11). wgenw_{\mathrm{gen}} inside the interval is the best: 1.73.

Q2. Input normalization and gradient scaling. Yes. The ADM state xtx_t is normalized to the input that DISC saw in training at the corresponding noise level. The gradient is taken with respect to xtx_t, so it needs no separate rescaling (new App. D).

AuthorsOpenReview ↗

Reminder to Reviewer ABdN

Dear Reviewer ABdN,

We would like to kindly follow up on our rebuttal. We would appreciate any further feedback, and we would be glad to address any remaining concerns before the discussion closes.

Thank you again for your time and consideration.

Reviewer_ABdNOpenReview ↗

Reply to rebuttal

Thank you for the detailed response. The clarification of the methods for W1 and W2, along with the additional experiments, helps readers better understand the motivation behind the proposed approach. The explanation that wgenw_{\mathrm{gen}} and ρ\rho are derived from the same underlying measurement, with ρ\rho representing its cumulative form, clarifies their relationship. The ImageNet ablation studies also support the usefulness of both components. Furthermore, the note regarding the revised experiments using 250 noise levels addresses my concern about relying on a single sampling noise level. Overall, the authors’ responses have resolved my concerns, and I find the contributions of this paper to be highly meaningful. I therefore increase my rating from “Weak Rejection” to “Weak Acceptance.”

Reviewer_VL77OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • The paper identifies that the aggregation of noise levels in "Diffusion Classifiers" has the most significant effect on performance metrics (SB, OOD, EC) out of the four considered axes (training data, input resolution, aggregation of noise levels, training objectives). cf. Table 1b)
  • The paper uses a pre-trained diffusion model to weigh the importance of different noise levels using a KL-Div formulation ⇒\Rightarrow wgen,ρ ⁣:[0,∞)→Rw_\text{gen},\rho\colon [0, \infty) \rightarrow \mathbb{R} (cf. Fig. 1 and Section 3.1)
  • The paper uses those weights to accomplish improvements:
    • Basic discriminative classifiers: Train a classifier using a mixture of noisy inputs and label smoothing (Eq. (7)).
    • Improved guidance: Use the above for classifier guidance and the noise weightings for classifier-free guidance
Critical strengths
  • The paper establishes the insight that
    1. the amount of information change (gain/loss) depends heavily on the dataset and the noise level. The shape of this function differs substantially between datasets (Figure 1b, 2c)
    2. this information-change function wgenw_\text{gen} can be used in different contexts (discriminative models/guidance) to improve performance.
Critical weaknesses
  1. Experiments on ImageNet discrimination and guidance show the usefulness of the obtained function wgenw_\text{gen}. It remains unclear:
    • how sensitive the choice is and whether or not wgenw_\text{gen} is optimal. Figure 2c) shows an alternative choice (based on Heinonen et al., 2025). How does this choice compare to your choice on the two main experiments (Table 1, Figure 3)? The improvements of your method in Table 3 are only minor, which suggests that the expensive wgenw_\text{gen} can be replaced by a substantially cheaper version.
    • Figure 2c) shows 1. a slight offset of the mean and 2. a more complicated overall function. Does the offset in the mean already explain the improvements over Heinonen et al., or is it really the more complicated overall function?
  2. Missing comparisons with other guidance scale scheduling approaches. In Section 4.3 and Figure 3 your classifier and classifier-free guidance approaches are presented. However, it is unclear how relevant wgenw_\text{gen} and ρ\rho are for those. Helpful would be the comparison with
    • a ResNet trained just like yours but with noise, e.g., like setting uniform+hard in Table 2,
    • a ResNet trained just like yours but with the wgenw_\text{gen} schedule from Heinonen,
    • concurrent work on scheduled scales. A quick search yielded Chen et al., 2026, "Information-Theoretic Classifier-Free Guidance with Adaptive Schedule Optimization" and V. S. et al., 2024, "Diffusion Inference with Dynamic Classifier-free Guidance".
  3. Conclusions from Table 1 are no more than mild evidence (lines 44 to 88)
    • The axes in Table 1b) are concluded from ≤2\leq 2 pairs of noisy models
    • (1) data, R1 -> R6: Also changes training objective
    • (2) resolution, R1 -> R7: Also changes model architecture (stem)
    • (3) aggregation: Trivially improves all metrics. But improvements differ substantially (e.g., ΔEC\Delta \text{EC} between +5.6+5.6 and +15.8+15.8)
Questions for authors

Clarification:

  • What are the clean accuracies for the other methods in Table 1a)? Why are they left out? Especially R4, R8-R11 would be important to understand the diffusion-based classifier baselines.
  • Where does the error consistency gain in R16 come from? To my understanding, you obtained wgenw_\text{gen} from EDM (with ΔEC=15.6\Delta \text{EC} = 15.6) and somehow that improved the ResNets ΔEC\Delta \text{EC} from 1818 to 2323?

Future work:

  • See weaknesses
4 comments

Thank you for the detailed review. Most of the comparisons you asked for are now in the paper.

Q1. Clean accuracies in Table 1a. We added as many as we could run ourselves or gather from prior work. Diffusion classifiers are extremely slow, so even prior work reports these numbers on small subsets. See the footnote of Table 1.

Q2. Where does the error-consistency gain of DISC come from? DISC does not change the training data, the resolution, the architecture, or the training objective (it is still cross-entropy). So we attribute the error-consistency gain of DISC to the aggregation over noise levels, which DISC performs during training instead of at test time (Introduction: "How an ordinary classifier can have it", L100–143, see its latest revision). The numbers are consistent with this: enabling aggregation for EDM (R9 → R8) raises EC by 4.7, and passing the same property to DISC (R7 → R18) raises EC by 5.0.

W1a. How sensitive is the choice, and is wgenw_{\mathrm{gen}} optimal?

  • We now extended our tests with different combinations of the curves proposed in the literature in Tables 2, 3 and 5, and compared their shapes and means with our curves in Figures 6, 7 and 8 (Appendices B and C). The improvements may be only minor, but we should not ignore that there is an improvement in every setting that one of our curves touches, with no other change to the setting.
  • More importantly, our curves have no free parameters to tune. Although prior work did a good job of tuning its heuristics and saturated the benchmarks, our curves can still match or improve on any benchmark, even if only marginally. Our curves look close to some curves in prior art, but with parameters that one would hardly include in a grid search, e.g., see L910–919.
  • Regarding your comment that the "expensive wgenw_{\mathrm{gen}} can be replaced by a substantially cheaper version", we do not agree that wgenw_{\mathrm{gen}} is substantially more expensive: it is computed once per dataset on a small subset, and after that it adds no cost to the training or the inference of the classifier. Moreover, Heinonen et al., the paper of the “cheaper version” as you call it, has pages of ablations with a significant amount of tuning experiments, which might be as expensive as training a diffusion model and obtaining our curves.

W1b. See Figure 7 and Appendix C for the comparison of the curves of Heinonen et al. with ours at ImageNet scale. There is both an offset and a different function shape. The third block of Table 5 shows that their curve at their scale is 4.1 points below DISC in clean accuracy (69.57 against 73.64). Shifting their curve to the mean of ours recovers 2.8 of these points (72.34); the remaining 1.3 points come from the function's shape. We added the same comparison to Table 3, where the effect is smaller because the hyperparameters were already tuned for the curves of Heinonen et al.

W2a–b. We added an FID column (at s=1s = 1) to Table 2, so every variant of Table 2 is also tested as a guide for classifier guidance.

  • An interesting observation is that only the settings that combine noise augmentation with coupled label smoothing improve FID, while settings with hard labels or constant label smoothing fail, regardless of the noise sampler used (numbers in Table 2).
  • Another interesting observation is the overlap in Figure 3a between the ADM classifier, which is a noise-conditioned half U-Net, and the ResNet-50 trained with a uniform noise sampler and hard labels.

W2c. Comparison with other guidance schedules. We ran the ones that are compatible with our settings. We also cite the literature on classifier guidance, including the ones we couldn't compare against. See Table 6 and Figure 3.

W3. Table 1 gives only mild evidence. We hope that this revision removes the confounds you name and makes it more tightly controlled:

  • Data (R1 → R6). R6 is now the SWSL ResNet-50 (Yalniz et al., 2019): the same architecture and the same loss as R1, trained with additional web images. The change is +7.2 / +11.8 / +0.4.
  • Resolution. R1 → R7 changes the stem of the network, and all later layers are unchanged. In a sense, changing the first layers is part of changing the resolution. Nevertheless, we added a new row, R12, which takes the R1 network unchanged and feeds it images reduced to 64 pixels and enlarged again: +36.5 / −11.2 / −2.2, the same direction as R1 → R7.
  • Aggregation. In the submission, the diffusion classifiers with a single noise level used one noise draw, so they had 250 times less budget than the ones with 250 noise levels. They now use 250 noise draws at the single level, which is the same number of evaluations. Aggregating over levels still adds +7.6 / +10.2 / +4.7 for EDM (R9 → R8) and +5.0 / +17.2 / +9.5 for Stable Diffusion (R11 → R4). The two values of Δ\DeltaEC are now +4.7 (DISC inherits +5.0) and +9.5 (they were +5.6 and +15.8).
Reviewer_VL77OpenReview ↗

Reply to rebuttal

Thank you for your detailed reply. I’m happy with how you addressed W1+W2+Q1.

Q2. Where does the error consistency gain in R18 come from? You argue that it comes from the aggregated noise levels. Since aggregating the noise of diffusion classifiers decreased EC by 4.7 (R9->R8), it seems plausible that this transfers to your discriminator models (DISC; R7 -> R18). However, as far as I can tell, R1 and R19 are the same setting as R7 and R18, only on 224px instead of 64px, and there we see only an increase of 0.1 in EC.

W3. Table 1 gives only mild evidence. I like the added rows. But I still don’t think that the table gives a sufficiently clear picture to draw strong conclusions.

Considering the new results, I am more convinced of the value of the contributions of this paper. I increase my rating from “Weak Rejection” to “Weak Acceptance”.

Thank you very much for your reply. We are glad that the revision addressed some of your concerns and that you now find the paper worth accepting.

One small practical note: on our side the rating still shows as 2, so we would be grateful if you could check that the update was saved.

About the remaining comments:

W3. Our aim in Table 1 is to decouple the axes that are confounded in the models compared in previous work, rather than to draw correlations over many models, which [1] already does. Such a careful and detailed decomposition over different axes is one of our contributions (contribution #4).

  1. Table 1b changes exactly one axis at a time, assuming that the change of architecture and hyperparameters is tied to the change of objective: axis 4 includes both the objective and the architecture, which makes sense because the architecture depends on the task (generation requires a decoder, while a classifier does not), and a change of architecture implies a change of hyperparameters.

  2. With the training data fixed (axis 1), we follow the line:

    • ResNet-50 at 224 px (R1) → ResNet-50 at 64 px (R7) changes the resolution (axis 2);
    • R7 → EDM with all noise levels (R8) changes the aggregation and the objective (axes 3+4);
    • R8 → EDM at a single noise level (R9) removes the aggregation (axis 3).

    This is where we isolate "aggregation" as an important contributor.

  3. We then test whether training the ResNet-50 at 64 px with DISC (R7 → R18) has an effect similar to the aggregation in EDM (R9 → R8). The ablations stay in this controlled setting (ImageNet-1k, 64 px): the DISC row of Tables 2 and 5 is the same model as R18.

  4. To strengthen the control of axis 3, we already had Stable Diffusion at a single noise level (R11) → Stable Diffusion with all noise levels (R4), where axes 1, 2 and 4 are exactly the same. We now run the same test on DiT-XL/2 and added it to Table 1 (single noise level, R21 → all noise levels, R10; in purple).

We agree that we need to follow the same line of decomposition on more than one setting to strengthen the conclusions. We have only done it on ImageNet-1k at 64 px with EDM and ResNet-50; nevertheless, we do all of the ablations and most of the experiments in this controlled setting, and we hope that this makes up for this weakness.

Q2. The 224-px setting is outside this controlled line: the curves come from a 64-px diffusion model, and we only rescale them with a heuristic for 224 px. The only aim of this setting was the comparison with existing augmentation methods. We stopped once we had the sufficient results and did not tune the recipe much, keeping most of the hyperparameters of the 64-px setting, which seems to hurt at 224 px: the matched baseline without noise (R13) is below the standard ResNet-50 (R1). Against this matched baseline, DISC at 224 px (R19) recovers +3.4 in error consistency (see R13->R19 row in Table 1b).

[1] Jaini, P., Clark, K., and Geirhos, R. Intriguing Properties of Generative Classifiers. In International Conference on Learning Representations (ICLR), 2024.

AuthorsOpenReview ↗

Reminder to Reviewer VL77

Dear Reviewer VL77,

As the discussion period ends today, we would like to kindly remind you that the system still shows your rating as 2. It seems that the increase to "Weak Acceptance" was not saved in the system, so we would be grateful if you could check it.

Thank you again for your time and consideration.

Reviewer_SVCBOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Given a pre-trained diffusion model on a dataset (it can be pixel or latent-based), the paper estimates a curve of two quantities (Fig 2):

  • wgenw_{gen} the diffusion classifier’s normalized attribution of its decision to each noise level
  • ρ\rho, which is the cumulative of the one above.

Following prior work on JEM, a hybrid discriminative-generative classifier, the method proposes to sample training data, noise them following the curve of Fig. (2), and also perturb the label in a similar way.

Using this simple method (no change in the arch of. classifier, no conditioning on the time step, etc.), the method DISC appears to be more performing in OOD, robust to noise, while keeping similar accuracy to the standard classifier. Alternative experiments are also proposed where the two quantities are used to try to improve previous work. Experiments are on ImageNet-1k images at 64p resolution (mainly) and 1 only at 224p resolution. There are also more convincing 2D experiments on point clouds.

Critical strengths
  • This is one of the papers in my batch where I felt that I was learning something.
  • The paper is felt to be constructed in a scientific way, starting from a small toy example (2D), then 64p, then 224p.
  • The outcome of the method makes sense: it is in my expectation (and also in the prior paper [A]) that the diffusion model has a sense of where the data is starting to be formed, and thus class features are starting to emerge.
  • One clear contribution

We isolate “aggregation” as an important contributor (more than training data or the generative objective alone) to this alignment.

which may help future research, i.e., marginalize across noise levels. The paper turns this into modeling KL of softmax distribution but probably treating the predicted eps from the network as score function. There is also a way to link to the score $\nabla_{\mathbf{x}_t} \log p_t(\mathbf{x}_t) $ as

−ϵθ(xt,t)1−αˉt-\frac{\epsilon_\theta(\mathbf{x}_t, t)}{\sqrt{1 - \bar{\alpha}_t}}

Note that alpha is related to the variance scheduler. I would appreciate an analysis of this. smt like check if there is a correlation between the norm of \eps and your curve or smt similar or the curve by [A] and your curve. I suspect that the proposed curve must be connected to when in the noisy steps the data mainly commits to one class. This is why I assume that parametrizing your curve per class may benefit the method.

[A] Choi, Jooyoung, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. "Perception prioritized training of diffusion models." In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11462-11471. IEEE, 2022.

Critical weaknesses

Paper quality and its appearance

  • The paper is a bit fragmented and imprecise sometimes. The text is fluent but the way it is organized it made me put effort to parse everything.

Limitations and Missing Papers and comparison

  • The experiments are all within the same dataset. ImageNet-1k to ImageNet-1k. Train the diffusion model on ImageNet, then test the classifier on ImageNet. This is reported as a strength, but actually measuring cross-transfer of curves in Fig. 2 would improve the contribution.

  • Assume first you need a diffusion model on a dataset, before training the classifier? Maybe there is a better way to do this without diffusion models. i.e., noising the data should give an idea of the local curvature of the loss of the classifier at each training point.

  • Not many experiments are high resolution (224p)

  • What about references on randomized smoothing?

  • The proposed curve in Fig. 2 has the same overlap with where the content is created as reported by [A]. Why is [A] not able to?


Expt

  • Modest experiments: All of the experiments are ImageNet-1k (besides the 2D clouds) or ImageNet-based.
  • The improvement in some cases is marginal.
  • Missing decomposing Fig. 2 and the proposed empirical scheduler per class. I suspect classes with less confusion relative to other classes have a very different shape than classes that are more confusing (e.g., MNIST 0 vs 1 is easy to discriminate, but 7 vs 1 is hard to discriminate).
  • Table 3 does not play in favor of the proposed approach.
  • Table 4 plays against the paper to show DISC + IPMix. It feels that the method is not enough to compete, so the method has to be combined. Note this expt counts more because we have those at 224p resolution. DISC is similar to MixUp (Zhang et al., 2018).
Questions for authors
  • given that their classifier is trained with noise, why they do not marginalize across noise levels there too?

a single prediction costs a forward pass for every class at every noise level, on the order of 105 passes per image on ImageNet-1

Where is this coming from? Please report precise citations or data.

  • The paper did not notice, but it makes Eq. (7) somewhat classifier equivariant. The noise added is propagated into the label.

  • L313 please define first what is ADM.

  • L353 > The matched control is the identical recipe with no noise.

not clear what it means.

  • why imagene in Tab. 1(a) takes as input 64 pixels?

  • in Tab1(a) why not adding the clean accuracy of generative model (their prediction with no noise.)

The 2×2 ablation

What is the meaning of 2x2? not clear. This sounds like a suggestion by an LLM.


I want to be positive on this paper for now: the rating is around borderline, but given that I have to choose, I lean towards weak acceptance.

4 comments

Thank you for the constructive review. Your pointers to the score function led us to add a new Appendix B. The curve is not correlated with the norm of the score, but it follows the gap between the class-conditional and the unconditional denoising error, which prior work links to mutual information.

Our responses are as follows:

Q1: Why not marginalize across noise levels? We added DISC with test-time aggregation over noise levels to Table 1 (R20) and to Table 4. We missed this experiment in the first version because the point of DISC was to be single-pass. As a side note, L100-L137 of the revised paper explains why DISC could approximate the aggregated decision of the reference diffusion classifier in single-pass.

Q2: Where does 10510^5 come from? See L41 and the citation of Li et al.

Q3: Noise added to the input propagates into the label. Yes, this is intended. We clarify it and give the rationale in L246–254, which is empirically tested in L864-869.

Q4: Definition of ADM. ADM is now defined and cited in L146.

Q5: "Matched control". We now call it matched baseline and define it in Sec. 4: the same training recipe (architecture, optimizer, schedule, augmentation), but without input noise and with one-hot labels.

Q6: Why does Imagen take 64-pixel inputs in Table 1(a)? This row is from Jaini et al. (2024), "Intriguing properties of generative classifiers", which cites Clark and Jaini, "Text-to-image diffusion models are zero-shot classifiers", and states explicitly that Imagen operates at 64×64 resolution.

Q7: Clean accuracy of generative classifiers. We filled the "clean" column of Table 1(a) from prior work. A diffusion classifier takes minutes to classify one image over 1,000 classes, so prior work uses small subsets of ImageNet. We give these details in the footnote of Table 1.

Q8: Meaning of "2×2". We now call it the "sampler × label ablation".

W1. The paper is fragmented and imprecise in places. We fixed the imprecise places you pointed out (Q2, Q4, Q5, Q8). We hope that you find the revised version more coherent.

W2. Cross-transfer of curves? We measured the curves on MNIST, trained on MNIST, and evaluated on MNIST-M as the OOD dataset (Appendix E). The MNIST and ImageNet-1k curves differ in shape, and the decision is formed earlier on MNIST. The MNIST-M results were similar nevertheless. We also used CIFAR-10 curves on ImageNet-64 (Table 5); clean accuracy drops by 1.2 points.

W3. Is there a way without a diffusion model? This is a thought-provoking question that we cannot answer yet. [A] attempts to do this: we now cite it and compare our curves with its measurements (App. B) and analyze it on toy examples (App. F).

W4. Few experiments at 224 px. We evaluated the 224-px DISC model of Table 4 on the human-alignment benchmark and added it to Table 1 (R19), next to its matched baseline (R13).

W5. References on randomized smoothing. We now added these references (see L180).

W6. Why not [A] on toy datasets? We experimented with this idea (see App. F); the curve of [A] does not overlap with our ρ\rho on these toy datasets. At first, we expected it to work on the checkerboard in the 2D input space (not even in the perceptual space of the MLP), but once we visualized it, we saw that the two points of a pair can come from non-adjacent cells that are far apart. As discussed in W3, this is an interesting topic.

W7. All experiments are on ImageNet. We now added MNIST (see App. E).

W8. Some gains are marginal. We tested different combinations of the curves proposed in the literature in Tables 2, 3 and 5, and compared their shapes and means with our curves in Figures 6-8 (App. B and C). The gains may be marginal, but we should not ignore that there is a gain in every setting that one of our curves touches, with no other change to the competitor's setting. More importantly, our curves have no free parameters to tune. Although prior work tuned its heuristics well and saturated the benchmarks, our curves can still match or improve on their benchmarks, even if only marginally. Our curves look close to some curves in prior art, but with parameters that one would hardly include in a grid search, e.g., see L910–919.

W9. The curve per class. We now computed the curve separately for each of the 1,000 classes (App. B, Fig. 6a). Every class curve decreases monotonically. We also trained a classifier with per-class curves (Table 5). We couldn't understand what you mean by "Missing decomposing Fig. 2".

W10. Table 3 does not favor the method. The rows our curves touch are generally better than the matched baselines for multiple metrics. We can't see why it doesn't favor our method. Also, the response to W8 applies here too.

W11. DISC Has to be Combined. We already acknowledged this in L430–431, L471-473. IPMix is a tuned combination of several augmentation methods, and we do not claim that DISC, a single ingredient, beats it.

Thank you for the rebuttal and the updated PDF, they both help.

 it follows the gap between the class-conditional and the unconditional denoising error, which prior work links to mutual information.

Citing this work and this phenomenon may improve the paper. OK, I saw they are in App. F. You may consider pushing up to the paper to motivate your findings (very briefly).

Why not marginalize across noise levels? Ok fair enough. This is convincing both for Tab 4 and Tab. 1 that they numbers do not change much when you augment the test, which makes a point for your single pass. Is it interesting, though, that SB in Tab. 1 increases a lot compared to DISC, probably inducing more PAG gradients? (assuming SB means more PAG gradients?)

Noise added to the input propagates to the label. My point here was not a criticism. I got that was intended. It was more to say that you train the classifier more like a diffusion model-like loss (equivariance) and less like classifier invariance. I suggest expanding on these and emphasizing them in the paper.

Transfering the results from CIFAR to ImageNet-64* Are those in Tab 5? OK, good. The clean accuracy does not drop much, but what about the other metrics like OOD and adv. robustness ? Can you give a clear comparison on ImageNet-64-> ImageNet-64 and CIFAR10->ImageNet-64 with a table with just those 2 rows? Not only on clean accuracy but on the other metrics like Tab. 1. Basically, the idea is to understand how much you curves are universal and what the impact is when you cross-transfer. Currently, the manuscript is very dense, and it is hard to extrapolate those.

Diffusion model: yes or no I also suspect, as a reviewer VL77 that your method may also work without training a DM first, but it is hard to say. The "class-conditional and the unconditional denoising error " may hint at this; the problem is how to estimate without the DM.

I am also waiting for the response from the other reviewers.

Thank you very much for your suggestions. Following them, the main text now

  • summarizes the link to score based alternatives, and comparison to curves obtained from [A] (L251–254), and
  • states after Eq. 7 that the label follows the noise added to the input, as in the loss of a diffusion model (L308–309).

Both additions are in purple in the new PDF.

Why not marginalize across noise levels? Thank you for the interesting observation, we agree with your point. One small note: SB in Table 1 is the shape bias on the cue-conflict images of the model-vs-human benchmark, so it does not directly measure perceptually aligned gradients.

Transferring the curves from CIFAR-10 to ImageNet-64. Here is the table you asked for (columns as in Table 2):

curves used on ImageNet-64cleanmCAAASBOODEC
measured on ImageNet-64 (DISC, 3 seeds)73.64 ± 0.0650.89 ± 0.1418.2 ± 1.460.5 ± 0.156.7 ± 0.323.0 ± 0.1
measured on CIFAR-10 (1 seed)72.4049.9726.165.556.822.5

Curves that transfer between datasets, or a universal curve, were never our aim. Still, the table shows that switching to the curves of a different dataset (CIFAR-10) does not change the results drastically. This suggests that some universality and cross-transfer might be achieved in the image domain.

Diffusion model: yes or no. Thank you for the inspiring comment. Yes, the method also works without a diffusion model, but not optimally. The curves of Heinonen et al. (2025) already give good results, and the curves obtained with the method of [A], which you shared with us, have parallels with ours (Appendix B) and give good results too. However, DISC, whose curves come from a diffusion model, outperforms them on most metrics:

curves used on ImageNet-64cleanmCAAASBOODEC
measured on ImageNet-64 (DISC, 3 seeds)73.64 ± 0.0650.89 ± 0.1418.2 ± 1.460.5 ± 0.156.7 ± 0.323.0 ± 0.1
Heinonen et al. (2025) (1 seed)72.4749.9525.759.055.621.3
method of [A], VGG features (1 seed)72.8150.4625.660.556.221.6

So these alternatives to a diffusion model could open the door to a more practical application of our idea of incorporating aggregation into the training of a discriminative classifier while keeping inference single-pass.

Dear Authors,

Thank you for all these details. I am no longer on the border, and I confirm my rating as Rating: 3: Weak acceptance

Comments on the paper

2 comments

We thank all three reviewers for their careful and constructive reviews. We replied to each review separately, and the paper is better thanks to their advice. We uploaded a revision in which the modified parts are written in blue. Our replies refer to its line numbers, so we kindly ask the reviewers to use this revision when reading our replies. Below are the changes, in the order of the paper. The method and the results of DISC are unchanged.

1. Where the alignment comes from (Introduction, Table 1, new Appendix G). Table 1 compares models that differ in training data, resolution, aggregation over noise levels, and training objective. There were concerns that the comparison was not controlled enough, so we modified it to be more tightly controlled.

  • Training data (R1 → R6). Before: R6 was CLIP ResNet-50, which also differs from R1 in the training objective. Now: R6 is the SWSL ResNet-50 (Yalniz et al., 2019), with the same architecture and loss as R1, trained with an additional 64M training images extending IN-1k, preserving the one-hot label format. Although these labels are obtained via weak supervision, the final training has the same CE loss as in our settings.
  • Aggregation (R9 → R8, R11 → R4). Before: the diffusion classifiers at a single noise level used one noise draw. Now: they use 250 draws, the same number of evaluations as aggregation over 250 levels. Aggregation still improves shape bias, OOD accuracy and error consistency.
  • Resolution (R1 → R12). Before: the only pair was R1 → R7, which also changes the stem of the network. Now: the new row R12 is the same 224-px network as R1, evaluated on images reduced to 64 px and enlarged again.
  • Other rows. We added the clean accuracies of the diffusion classifiers, DISC at 224 px, and DISC averaged over 250 noise levels at test time (also in Table 4). Appendix G describes how we read the rows of Ortega et al. from their figures.

2. The basis of the method (Introduction, Sec. 3.1, new Appendices B and F). Before: the paper only defined the two curves and the training recipe.

  • We added the reasoning for using the curves as the noise sampler and the label confidence (Introduction, "How an ordinary classifier can have it").
  • We added the two assumptions under which ρ\rho is the expected log-probability of the true class (Sec. 3.1).
  • We added an analysis of the curves (Appendix B). We recomputed them with gains in the log-probability of the ground-truth label instead of reductions in KL to the full prediction; the recomputed curves are close to ours (Fig. 6b, 6c). We computed a curve for each of the 1,000 classes. We related the curves to the score: the gap between the class-conditional and the unconditional denoising error overlaps wgenw_{\mathrm{gen}} (Fig. 6b).
  • We compared the curves with the measurement of Choi et al. (2022) on ImageNet (Appendix B) and on our toy datasets (Appendix F).

3. Other samplers and label schedules (Tables 2 and 3, new Appendix C).

  • Table 2 (formerly "2×2", now "sampler × label ablation"). Before: our sampler or a uniform one, with our labels or hard labels. We now added the SNR-based labels of Heinonen et al. (2025). We also added an FID column, in which each model guides ADM.
  • Table 3 (our curves in the pipeline of Heinonen et al.). There were concerns about the ablations being limited. We now added their curves in the 60% noise (xnnnb) setting and extended the "noise-level sampler x label schedule" ablation in this setting.
  • Appendix C (Table 5, Figs. 7 and 8). We added the log-normal noise level sampler of EDM with our labels. And further "noise-level sampler x label schedule" ablation is done. We also fitted a log-normal to our curves: our curves fit well to a log-normal distribution, but with parameters that no one would include in a grid search.

4. Guidance (Sec. 3.3, Fig. 3, new Appendix D with Table 6). Before: DISC, ADM's own classifier, and a ResNet-50 trained without noise as guides. Added: FID numbers for every variant compared in Table 2. We also ran the full sweep with two additional guides (one trained with uniform noise levels and hard labels, one with the recipe of Heinonen et al.). Moreover, in the CFG setting, we compare wgenw_{\mathrm{gen}} with other schedules and combine them.

5. A second dataset (new Appendix E). Before: ImageNet-based datasets and 2D toy data. Added: MNIST. We measured the curves on MNIST and trained with them. Then we compare MNIST curves to ImageNet curves and report DISC accuracy on MNIST-M.

6. Clarifications. We added the source of the 10510^5 cost and the definition of ADM. We renamed the "matched control" to "matched baseline" and defined it in Sec. 4. We added references on randomized smoothing and on guidance schedules.

Area_Chair_jXJiOpenReview ↗

Review discussion

Dear reviewers,

the authors have responded to your reviews. Can you please check their responses and engage in a discussion with them to clarify any open questions or help them address your concerns. I also strongly recommend you check the reviews and answers of your fellow reviewers and take them into consideration during the discussion so that we can preferably reach a consensus at the end of the review period.

Thank you, your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.