acceptodds
Under review as a conference paper at ICLR 2027

When Simulator Targets Suppress Diversity in Conditional Diffusion: Evidence from ECG Generation

Abstract

Matching a diffusion model's clean-signal estimate to a deterministic class template penalises within-class variation. We derive this penalty and show that the population-optimal estimate's variance contracts by along eigenvectors of , where is the coupling operator, the coupling strength, and the corresponding eigenvalue. Template accuracy does not enter this factor. A reverse-chain covariance recursion, exact under Gaussian assumptions, predicts sampled variance ratios in synthetic experiments without fitted parameters. On patient-disjoint PTB-XL beats, with templates fitted without test patients, point coupling reduces pooled within-class standard deviation to of real data and increases raw-signal by under a shared kernel. Improving the simulator, using an oracle class mean, or drawing an independent random target does not recover the lost diversity. Training the morphology and inter-lead terms separately shows that each suppresses diversity: the inter-lead term contracts the limb leads it references without improving their amplitude-normalised consistency. A constructed beat satisfies SE-Diff's released inter-lead penalty while violating Einthoven's law among its generated leads, and the released paired evaluation ranks the most concentrated generator best. Moment matching, spread matching, and exact lead identities reduce the distributional penalty; spread matching with exact identities reaches the unregularised model's raw-signal . No tested coupling improves within-PTB-XL downstream classification over the unregularised baseline beyond seed variability. Experiments on MIMIC-IV-ECG, multi-beat windows, and a small latent model reproduce the contraction. These results support avoiding deterministic template matching when within-class variation matters. They do not establish that simulator priors or sample-dependent constraints are ineffective in general.

Primary area:
generative models
Keywords:
Conditional diffusion models, physics-informed generative model, generative modeling, diversity collapse, ECG generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CdmMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the impact of deterministic simulator template constraints on intra-class diversity within conditional diffusion models. The authors derive the variance-shrinking behavior of optimal denoising estimates and provide a recurrence relation for the covariance of the reverse sampling process under linear Gaussian assumptions. Through generative experiments using synthetic data as well as the PTB-XL and MIMIC-IV-ECG datasets, the study examines the effects of template accuracy, constraint strength, and various coupling strategies. The results demonstrate that fixed template constraints reduce generative diversity and increase distribution divergence; furthermore, neither improving template accuracy nor sampling templates independently and randomly is sufficient to mitigate these effects.

Critical strengths

The key conclusions are supported by robust controlled experiments. The authors examine the roles of template accuracy, constraint strength, and losses related to morphology and leads; for primary comparisons, they employ a mutually exclusive partitioning of patients and exclude the test patients from the template cohorts. Validation using both synthetic data and two ECG datasets further strengthens the credibility of these conclusions.

Critical weaknesses
  1. Theoretical analysis offers only limited quantitative explanation for practical diffusion models. The accuracy of multi-step covariance recursion depends on ideal conditions such as Gaussian distribution and overall optimal denoiser, while in real ECG experiments, this recursion significantly underestimates the degree of contraction (in Section 5.1). Consequently, current theoretical frameworks primarily account for the general trend of contraction, yet they remain insufficient for predicting the actual magnitude of contraction in trained models.
  2. This paper focuses on the fixed-template squared-error constraint used in SE-Diff, analyzing its impact on generation diversity. However, practical simulator-guided methods employ a variety of constraint forms, ranging from residuals of physical equations to adaptive constraints that depend on the generated samples themselves. Although the paper explicitly defines the scope of its conclusions, it lacks a systematic comparison of these diverse constraint types. Consequently, it fails to clarify which constraint designs can enhance physical consistency while preserving intra-class diversity, thereby limiting the paper’s utility as a guide for the development of practical methods.
Questions for authors
  1. In the real ECG experiments, the multi-step covariance recursion underestimates the observed contraction. What explains this discrepancy?

  2. Table 2 shows that improved diversity and reduced distributional discrepancy do not consistently translate into higher downstream classification AUC. How do the authors interpret this mismatch, and what does it imply about the task relevance of the variation recovered by alternative constraints?

  3. Beyond the SE-Diff-derived experimental settings, could the authors provide evidence from another diffusion method that incorporates simulator-based constraints to assess whether the observed contraction generalizes across methods?

1 comment

Thank the author for the additional experiments and clarifications. The residual and PPG experiments help clarify when contraction occurs. As the authors note in the abstract (last sentence), the results “do not establish that simulator priors or sample-dependent constraints are ineffective in general.” My concern remains how far the proposed design guidance extends to other forms of simulator guidance. I will therefore maintain my original score.

Reviewer_cp4SOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper looks at the issue of conditional generation and whether the loss is affected by pulling the signal towards a defined class. The argument is that this reduced within-class diversity and the authors describe this effect through theory and experiments. They then evaluate on PTB-XL and demonstrate the results through ablation to explain away other possible effects.

Critical strengths

The ablation study is good, variance is reported and error bars plotted. Alternative explanations for the effects are tested and ablated, isolating the probable effect.

Critical weaknesses

My biggest issue is that while SE diff is the benchmark, baseline and starting point, SE is not faithfully reproduced for direct comparison. The hyperparams do not match either (the smallest tested lambda value is orders of magnitude larger than in SE diff). The authors acknowledge this, but the difference from the actual SE diff operating values.

Also notable, the main text is very number-heavy and some passages in the main text are more like a table than prose. It would be easier to just reference the relevant tables and spend more of the space and attention on things the numbers don’t say.

I also wonder if the novelty is high enough here as the main theorem is ridge regression, but the characterisation itself is useful and will be of interest.

Questions for authors
  1. Can fig font size be increased and one bar stopped from overlapping with null line.
  2. What happens with the published SE diff hyperparams? Why were they not used?
  3. Why use only 4000 samples?
Reviewer_bWkZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies point coupling in conditional diffusion — matching the model's clean-signal estimate to a deterministic per-class template, as in SE-Diff's simulator losses for ECG. It shows the squared penalty decomposes into a target-mismatch term plus a within-class variance penalty, so the population-optimal estimate's variance contracts by (1+λμ_i)^-2 along eigendirections of AᵀA regardless of template accuracy, and gives a Gaussian reverse-chain recursion that predicts the contraction on synthetic data without fitted parameters. On patient-disjoint PTB-XL, point coupling reduces diversity to 0.66 of real data and increases raw MMD² 3.7x; better/oracle/random targets do not restore diversity; both the morphology and inter-lead terms suppress diversity separately, and the inter-lead term does not improve amplitude-normalised consistency — a constructed beat attains ~0 released inter-lead loss while violating Einthoven's law. Moment/spread matching and exact lead identities reduce the damage, but no tested coupling beats the unregularised baseline on downstream classification.

Critical strengths

(1) Exemplary controls for a negative result: template quality (Gabor/oracle/random targets), patient/template leakage (clean cohort excluding all test patients), and metric artefacts (shared kernel, matched draws, null calibration) are each closed experimentally, and a matched-seed factorial design separates the morphology and inter-lead harms. (2) A concrete, falsifiable counterexample against SE-Diff's released loss and paired evaluation, plus an actionable distinction: exact identities enforce consistency at no measured distributional cost, while deterministic templates tax variance.

Critical weaknesses

(1) Theoretical novelty is limited — the authors note the bias–variance identity is standard and Theorem 1 is essentially ridge shrinkage restated for the clean-signal estimate; the new recursion is exact only under Gaussian/population-optimal assumptions and underestimates real-ECG contraction by ~2x by the authors' own account, so it validates on the synthetic setting it was derived for. (2) Generality is not established: one coupling family, one headline weight, one small UNet, phase-aligned single beats, one domain — and SE-Diff itself is not reproduced, only its loss functions are audited, so the framing around "simulator targets" is broader than the evidence. (3) The paper offers no constructive improvement: the best fix (spread matching + exact identities) still trails the unregularised model on TSTR and feature MMD² and collapses on the rarest class, so the contribution is a well-supported warning rather than a method.

Questions for authors

(1) Can you demonstrate the contraction, even qualitatively, for one sample-dependent or PDE-residual simulator loss or one non-ECG modality — or will you scope the title/claims explicitly to deterministic point-template coupling? (2) What explains the ~2x gap between the Gaussian recursion and real-ECG contraction? (3) With a batch-size-corrected moment estimator, does any corrected coupling ever beat no coupling on a pre-specified metric?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.