acceptodds
Under review as a conference paper at ICLR 2027

Equal Path Cost, Unequal Output Effects: Understanding Perturbation Propagation in Diffusion Models

Abstract

Diffusion models have achieved remarkable success in generative modeling, with their sampling procedures routinely modified to control generation and improve efficiency. These modifications introduce perturbations along the sampling trajectory, raising a central question: how do such perturbations affect generated output? To address this question, we develop a theoretical framework to investigate perturbation propagation that combines dynamical analysis of the sampling process with an information-theoretic characterization of output responses. Within this framework, we quantify perturbation strength using the Kullback–Leibler (KL) divergence between perturbed and reference trajectory distributions and show that this path cost bounds, but does not determine, changes in the output distribution. Building on this analysis, we derive a response identity that connects the propagation and accumulation of local perturbations with the information captured by a selected feature mean, explaining why changes in the output distribution can remain undetected by its first-order response. We test our theoretical analysis through controlled interventions at equal path cost in pretrained diffusion models, revealing distinct patterns of output sensitivity across sampling stages and spatial frequencies. To assess whether our framework can diagnose perturbations arising from practical approximations, we apply it to cache-based acceleration and show that our propagation analysis reliably identifies sampling intervals where caching causes larger image errors.

Primary area:
generative models
Keywords:
Diffusion Models, Information Theory, Perturbation Propagation, Sensitivity Analysis, Error Accumulation

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

This paper focuses on how perturbations introduced during diffusion sampling propagate through the remaining sampling trajectory and how it affects the final output. The main argument mentioned in the paper is that perturbations with the same path cost can still lead to substantially different output effects depending on when and how they are introduced. The authors propose a theoretical framework that decomposes path-level changes into output and conditional-trajectory changes and derives a first-order propagation analysis for understanding how local perturbations accumulate or cancel. They validate the framework using controlled equal-path-cost perturbations across JiT, ADM, and DiT, showing that output sensitivity varies across sampling stages and spatial frequencies. The paper further applies the propagation analysis to cache-based diffusion acceleration, where it is able to identify sampling windows associated with larger final-image errors.

Critical strengths
  1. The paper provides a well motivated theoretical framework for explaining why perturbations with the same path cost can lead to different final output effects. The decomposition between path level and output level changes, together with the propagation analysis, provides a useful way to understand how perturbations can accumulate or cancel throughout the sampling trajectory.

  2. The experimental validation is reasonably broad across JiT, ADM, and DiT, and shows that sensitivity to perturbations depends on both the sampling stage and spatial frequency. The cache-based acceleration experiment further shows a practical use of the framework, where the propagation-based predictor is able to identify caching windows associated with larger final-image errors.

Critical weaknesses
  1. The propagation analysis is based on a local first-order approximation, while the practical experiments use finite perturbations. Although the analysis captures the general temporal sensitivity patterns, there are substantial differences between the predicted and observed response magnitudes in the DiT experiments. This limits how directly the local analysis can be used to predict the magnitude of finite output changes.

  2. The practical application of the proposed propagation analysis is primarily demonstrated through cache-based acceleration, with only eight candidate caching windows per model. While the reported correlations are strong, broader evaluation across other practical perturbations or acceleration settings would provide stronger evidence that the proposed propagation analysis generalizes beyond the caching setup.

Questions for authors

The first-order propagation analysis appears to capture the temporal sensitivity patterns, but the finite perturbation experiments show substantial differences between predicted and observed response magnitudes, particularly for some DiT settings. Could the authors clarify how they expect the proposed analysis to behave as the perturbation magnitude increases, and whether there is a practical regime in which the first-order approximation can reliably predict both the location and magnitude of output sensitivity?

3 comments
AuthorsOpenReview ↗

Response to Reviewer QSrU (Part I)

Thank you for recognizing the theoretical framework and its validation across model families. To address your concern about applying the local theory to finite perturbations, we have expanded the analysis of how prediction accuracy changes with perturbation strength. We also clarify the purpose of the caching experiments and how other interventions enter the framework.

W1&Q1: Prediction accuracy as perturbation strength increases

As perturbation strength increases, higher-order terms can lead to differences between the first-order prediction and the observed output change. Our additional strength sweeps examine both the size of these differences and whether the predictions capture how output errors vary across sampling windows. Under the differentiability assumptions of Proposition 2, the output displacement satisfies Yα−Y0=αV+o(α)Y^\alpha-Y^0=\alpha V+o(\alpha). With bounded second derivatives, the displacement remainder is O(α2)O(\alpha^2) near zero. The size of this remainder depends on the sampling dynamics and the intervention, so we assess prediction accuracy at specified strengths for each model.

In Appendix D.3, we vary the caching interpolation:

oiα=oi(Xiα)+α(oa(Xaα)−oi(Xiα)),o_i^\alpha=o_i(X_i^\alpha)+\alpha\bigl(o_a(X_a^\alpha)-o_i(X_i^\alpha)\bigr),

where aa is the window's first step, α=0\alpha=0 gives fresh predictions, and α=1\alpha=1 gives direct caching. We pair classes and randomness across strengths and windows. We measure relative displacement error by ∥(Yα−Y0)−αV∥/∥Yα−Y0∥\|(Y^\alpha-Y^0)-\alpha V\|/\|Y^\alpha-Y^0\|, taking the median over 64 samples and eight windows. This measures errors in both the direction and magnitude of the predicted image displacement. We separately evaluate the ordering of windows using the Spearman correlation between their sample-mean predicted and observed image MSEs. Representative results from Table 6 are:

ModelStrengthSpearmanMedian error (%)
JiT-B0.251.000011.50
JiT-B10.952426.35
ADM0.0250.904810.56
ADM10.9762136.58
DiT0.010.976216.15
DiT0.11.000070.01
DiT10.9762123.40

At the tested strengths of 0.250.25 for JiT-B, 0.0250.025 for ADM, and 0.010.01 for DiT, we obtain median relative displacement errors of approximately 11%–16%, together with high correlations between predicted and observed window errors. These settings provide evidence of both relatively accurate displacement predictions and agreement in which sampling windows produce larger errors. At full caching strength, ADM and DiT retain high ranking correlations despite median displacement errors exceeding 100%. Thus, predicting the ordering of window errors can remain accurate even when predicting the image displacement incurs large errors.

This distinction is consistent with the local analysis: the ordering of two response magnitudes is preserved when the sum of their prediction errors is smaller than the gap between the predicted magnitudes. We also find that the error need not decrease monotonically over the sampled strengths: for JiT-B/L, median errors are lower at 0.250.25 than at 0.10.1. In Appendix D.3, we additionally compare symmetric finite differences with the local predictions. Under sufficient smoothness, the symmetric difference cancels even-order terms in the output expansion.

The DiT experiment in Section 4.3 examines the complete propagation through latent sampling states and the nonlinear image decoder. For the controlled DCT interventions at ε=0.8\varepsilon=0.8, we recover the same peak-response windows for the local predictions and finite responses. The finite/local peak amplitude ratios are 36.8%, 47.7%, and 87.9% for DC, DCT-4, and DCT-31, respectively. These results show how the propagation analysis can capture the temporal locations of peak responses while overestimating their magnitudes at the tested strength.

The exact constructions in Appendix B further show why finite-strength accuracy depends on the structure of the intervention and dynamics. In the additive mean-control family of B.1, the output displacement equals εV\varepsilon V at every strength. In the covariance family of B.2, higher-order effects appear as strength increases, and Figure 9 shows convergence to the local predictions as strength decreases. Together, these examples explain why the range of accurate first-order prediction varies across perturbation families.

Revision locations: Section 4.3, pp. 8–9, and Figure 5, p. 8; Appendices B.1–B.2, pp. 18–21, and Figure 9, p. 20; Appendix D.3, pp. 28–29, and Table 6, p. 29.

AuthorsOpenReview ↗

Response to Reviewer QSrU (Part II)

W2: Experimental design and generalization beyond the caching setup

We use caching-induced errors to empirically validate the framework under perturbations arising from a practical sampling approximation, complementing the controlled interventions in Section 4. We have revised Section 5 to make this motivation explicit. The eight equal-length windows keep the reuse rule and number of reused predictions fixed while varying when the errors enter the sampling trajectory. This design lets us compare their downstream effects under a consistent caching protocol across four pretrained generators.

Within this setting, we examine both the observed output changes and the contributions that produce them. The strength sweep discussed above assesses finite-displacement accuracy; the contribution analysis in Appendix D.4 separates individual propagated magnitudes from cross-time interactions. We find that the summed cross-time contribution increases the mean squared first-order response in every evaluated model and window. This provides evidence for how errors from a practical sampling approximation propagate and accumulate, alongside the controlled temporal and directional comparisons in Section 4.

For other differentiable interventions, we specify their modified sampling updates, differentiate them to obtain the local defects, and propagate those defects through the reference dynamics and output map. Appendix C.7 makes this extension concrete by deriving the guidance-scale update defect and describing how to incorporate history-dependent feature caching. We leave systematic comparisons across additional caching methods, window layouts, and practical interventions to future work, using this formulation to examine how their different error structures affect propagation and finite-strength accuracy.

Revision locations: Section 5, pp. 9–10; Appendix C.7, p. 25; Appendix D.3, pp. 28–29; Appendix D.4, pp. 29–30.

Thank you to the authors for the detailed responses

About my concern with the first-order approximation, the additional experiments and results in Appendix D.3 help address prediction accuracy under finite perturbation strengths, although magnitude prediction still remains less reliable when it comes to larger strengths. The caching explanations and cross-time analysis strengthen the experimental evidence, but broader validation across other practical perturbations remains limited.

I would maintain my rating

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

The paper studies how much the final image changes when an intermediate step of diffusion sampling is perturbed. The size of the perturbation is measured by the KL divergence between the original trajectory distribution and the perturbed one, which the paper calls path cost. The central claim is that path cost bounds the change in the output but does not determine it. This is because a perturbation of the same size produces different outcomes depending on which step it is applied at and in which direction, and because perturbations at different steps can cancel at the output. The paper describes, to first order, how a perturbation propagates through the remaining steps, and verifies this by injecting perturbations at several steps and spatial patterns while holding the path cost equal. It then applies the analysis to prediction caching in order to predict in which interval caching leads to larger image error.

Critical strengths
  1. Figure 1 presents the central claim in a verifiable form.
  2. The caching experiment in Section 5 tests the claim on a practical problem. The predictor that computes propagation attains Spearman correlations of 0.9524-0.9762 with the observed error, while the measure that only gauges the magnitude of the perturbation scatters from -0.9048 to 0.5000 and path cost is negative for all four models (-0.9048 to -0.3095). This pattern holds consistently across four models spanning pixel space and latent space (JiT-B, JiT-L, ADM, DiT).
Critical weaknesses
  1. Only a single caching protocol is considered. Section 5 covers only the scheme in which the prediction from the first step of a window is reused for the following seven steps. There is no analysis of whether this transfers to the other caching methods mentioned in Section 2.
  2. Which window is risky for caching can also be determined by directly applying caching at each window and measuring the image error. The paper does not report which of the two is cheaper, the direct measurement or the proposed predictor.
Questions for authors
  1. What is the computational cost of the predictor in Section 5? Compared with directly applying caching at each window and measuring the error, which is cheaper?
4 comments
AuthorsOpenReview ↗

Response to Reviewer WUcv (Part I)

Thank you for highlighting the distinction between local perturbation magnitude and output error, and for your questions about transfer across caching methods and computational cost. We clarify below how the framework accommodates different caching mechanisms, why we chose the current experimental protocol, and what propagation analysis reveals about caching-induced errors. We have also improved the presentation to clarify the theoretical quantities and experimental motivations.

W1: Extension to other caching mechanisms

We extend the analysis to other differentiable caching mechanisms through the local update perturbations they induce. For each mechanism, we differentiate its modified sampling update along the reference trajectory and propagate the resulting defect through subsequent updates and the output map. In Appendix C.7, we describe how an augmented state or an explicit history-dependent formulation incorporates cached features that depend on past states. This gives a common propagation identity while allowing each mechanism to introduce its own update defects and history dependence.

We chose a fixed-window caching protocol to examine propagation under a consistent reuse rule across four pretrained generators. Each candidate window uses one fresh prediction followed by seven reused predictions. By holding the window length and reuse pattern fixed while varying the intervention window, we study how the timing of model- and trajectory-dependent errors affects their downstream transformation and accumulation. This supplies perturbations from a practical sampling approximation alongside the prescribed interventions in Section 4. Comparing different caching methods would additionally vary the reused quantities, reuse conditions, or update rules.

Within this setting, the propagated response recovers the temporal ordering of observed window errors with Spearman correlations of 0.95240.9524–0.97620.9762, exceeding the local-defect and path-cost baselines. We thereby substantiate the role of downstream dynamics in shaping output effects beyond the strength of the injected errors. In Appendices D.3–D.4, we further show how this temporal agreement persists across strengths with different displacement accuracy, and how cross-time interactions contribute to the combined response. These analyses deepen the mechanism-level understanding of the tested protocol. We leave systematic comparisons across caching mechanisms to future work, using their induced update defects to examine how different reuse strategies affect perturbation propagation.

Revision locations: Section 5, pp. 9–10; Appendix C.7, p. 25; Appendix D.3, pp. 28–29; Appendix D.4, pp. 29–30.

AuthorsOpenReview ↗

Response to Reviewer WUcv (Part II)

W2&Q1: Computational cost and the purpose of propagation analysis

At the level of operation counts, direct measurement with prefix sharing requires fewer denoiser forward evaluations than our current propagation implementation and avoids its directional-derivative work. Both methods scale as O(KM)O(KM) in the number of candidate windows KK and sampling steps MM. We compare their computational work by counting forward evaluations and Jacobian–vector products (JVPs) separately; these counts specify the computation required rather than a measured wall-clock ratio.

For each window, our implementation propagates one tangent through all MM sampling updates and the terminal output map. Each JVP evaluates the reference forward computation together with its directional derivative, also supplying the reference trajectory. Direct measurement computes one shared reference output and one cached output per window. At full caching strength, a window of length LL uses one fresh prediction followed by L−1L-1 cached predictions, saving L−1L-1 denoiser evaluations on that candidate trajectory. It can additionally share the reference prefix before the intervention.

For M=64M=64, K=8K=8, and L=8L=8, the resulting counts per image are:

ComputationDenoiser evaluations, including the terminal prediction
Propagation branch of the current implementation8(64+1)=5208(64+1)=520 JVP evaluations, each including a forward evaluation and its directional derivative
Direct measurement with one shared reference and candidates run from the initial state, skipping cached predictions65+8(64−7+1)=52965+8(64-7+1)=529 ordinary forward evaluations
Direct measurement additionally sharing the reference prefix before each intervention65+(57+49+41+33+25+17+9+1)=29765+(57+49+41+33+25+17+9+1)=297 ordinary forward evaluations

Prefix sharing retains reference states and predictions at window starts and uses the same sampling noise. The propagation count includes all steps in our current implementation. DiT additionally requires eight decoder JVPs for propagation versus nine decoder forward evaluations for direct measurement, and its local-defect calculation differentiates the mean/variance update with respect to the supplied prediction.

Our purpose in Section 5 is to empirically validate the proposed perturbation propagation analysis using errors arising from a practical sampling approximation, complementing the controlled interventions in Section 4. We have revised Section 5 to make this motivation explicit. In this setting, propagation analysis allows us to examine how local caching errors are transformed downstream and combine across time, while direct interventions provide the observed output changes against which we assess the analysis. For example, in the separate 32-sample analysis in Appendix D.4, cross-time terms account for 59.6%–84.2% of the mean squared first-order output response. This decomposition, together with the comparison against local error measures, supplies evidence for the propagation and accumulation mechanisms studied in the paper.

Revision locations: Section 5, pp. 9–10; Appendix C.7, p. 25; Appendix D.4, pp. 29–30.

AuthorsOpenReview ↗

Response to Reviewer WUcv (Part III)

Presentation Improvement: clarifying the theoretical quantities and experimental motivations

We have improved the presentation to make the theoretical quantities, their interpretation, and the questions addressed by the experiments easier to follow. In Sections 3.3–3.4, we define an observable as a scalar output statistic f(Y)f(Y) and its mean response as the derivative of its expectation with respect to perturbation strength. We then explain the two stages of information loss—from the trajectory to the output distribution, and from that distribution to a selected observable’s mean—before introducing their exact factorization in Theorem 1.

We use a common Gaussian construction to connect the theoretical results visually: Figure 2 accompanies the accumulation identity, while Figure 3 accompanies Theorem 1 and illustrates a zero mean response under a locally changing output distribution. We define the control and observable angles in the captions and give the construction in Appendix B.1. Distinct line styles, markers, and bar patterns make the quantities distinguishable in Figures 2–3 and 8–9.

We also state the question motivating each experiment: how perturbation timing and spatial direction affect output responses at equal path cost in Section 4.2; how propagation through sampling and decoding predicts image responses in Section 4.3; and how the choice of output statistic affects the mean response to a fixed perturbation in Section 4.4. In Section 5, we explain how the propagated image displacement defines RpredR_{\mathrm{pred}}, how direct caching defines RobsR_{\mathrm{obs}}, and how their comparison examines propagation under practical sampling errors. We distinguish magnitude agreement from temporal ordering and connect the strength sweep and contribution analysis to finite-response accuracy and cross-time accumulation. These revisions give each experiment an explicit question, measurement, and theoretical interpretation.

Revision locations: Sections 3.3–3.4, pp. 5–7; Figures 2–3, p. 6; Appendices B.1–B.2, pp. 18–21; Sections 4.2–4.4, pp. 8–9, and Figure 6, p. 9; Section 5, pp. 9–10; Appendix C.1, p. 22; Appendix D.3, pp. 28–29; Appendix D.4, pp. 29–30.

Thank you for your response. My concerns are addressed, so I have raised my score.

Reviewer_1pjKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Many diffusion-model sampling strategy included purtubating the trajectory using external guiding forces such as classifier free guidances, quantization, and feature caching. How does such perturbation change the outcome? This work sheds a light on this question from an information-theoretic perspective. A straightforward idea is to measure the extent of path changes caused by the perturbation using KL divergence, termed path cost, but this does not fully explain the generation change because a same-cost trajectory may fall into different terminals. Within the eaqual cost trajectories, a purtubation that changes the feature mean is analyzed to cause output changes in a mean-variance decomposition analysis. In experiments, discrete cosine transform (DCT)-based blockwise perturbations are applied to three pretrained models, JiT, ADM, and DiT, showing that larger-block perturbations cause larger effects on the output, supporting the proposed theory.

Critical strengths
  • Insensitivity/robustness of diffusion models against perturbation is empirically known but to my knowledge it is analyzed in a general framework. This work presents a simple but widely applicable idea.
  • I am interested in the DCT-based perturbing strategy, which may serve as a diversification method to provide non-contracting perturbation.
  • The theory is supported using real-image large models beyond toy experiments.
Critical weaknesses
  • Theoretical contribution: I have a bit of concern about the significance of the theoretical results: the relationship between mean-reverting stochastic processes and diffusion generative models is well established [a,b]. Difference / novelty / contrubution seem somewhat weak in the context.

    [a]: Contractive Contracting Diffusion Probabilistic Models, SIAM Journal on Imaging Sciences, 2026 [b]: On the Robustness of Distribution Support under Diffusion Guidance arXiv 2026 (preprint)

  • From an applicational view, generative image quality and its relationship to classifier-free guidance should be discussed. For example, Fig. 2 shows large-block DCTs have large impacts; it is not clear whether large-block DCT perturbations are still feasible for maintaining image quality.

Questions for authors
  • Minor suggestion: Fig. 1 looks a bit uninformative. Adding axis labels and descriptions of which aspects of the trajectories represent output KL and bridge costs would be beneficial.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 1pjK (Part I)

Thank you for recognizing the relevance of a general framework and the validation on pretrained image models. Your comments helped us clarify both the theoretical contribution and the interpretation of the perturbation experiments. We address your constructive comments below.

W1: Theoretical contribution and relation to mean reversion, contraction, and support robustness

Our framework characterizes how local perturbations propagate into output-distribution changes and how those changes appear in the means of selected output statistics. We derive an exact decomposition of path cost into output-distribution changes and conditional-trajectory changes, and a propagation law that expresses the output derivative as a sum of transported local perturbations. Theorem 1 links the dynamically propagated response to an exact factorization into output visibility and observable–score alignment. This distinguishes whether perturbations change the output distribution locally from whether that change appears in a selected statistic’s mean. In particular, the output distribution can change locally while the statistic has zero first-order mean response. The factorization identifies this case through orthogonality to the output likelihood score, defined with respect to perturbation strength. Together, the results explain both unequal output effects at equal path cost and local distributional changes invisible to a selected mean.

This question complements the two works you suggest, which we now discuss in Section 2. Contractive Diffusion Probabilistic Models develops backward-contracting sampling dynamics to control score estimation and discretization errors. We characterize how individual perturbations are transformed and combined under locally differentiable sampling dynamics, including reinforcement and cancellation across time. On the Robustness of Distribution Support under Diffusion Guidance studies proximity of guided samples to the target support under its exact-score and distributional assumptions. We characterize changes in output probabilities and the expectations of selected output statistics, including changes that preserve the support. In this way, we connect changes in the generated distribution to their dynamical origins and measured responses.

We illustrate these theoretical distinctions in Figure 1: two controls with path cost 2 nats produce output KL approximately 1.957579 under reinforcement and zero under cancellation. Scaling the cancelling control gives arbitrary positive path cost with unchanged output distribution. In Appendix B.1, we independently rotate the temporal control and the measured observable, separating changes in output visibility from changes in observable alignment. Figures 2–3 illustrate these dependencies alongside the corresponding theoretical results. In Figure 3, the perturbation and its positive output Fisher information remain fixed while the mean response decreases to zero at the orthogonal measurement direction.

Section 4.4 provides empirical support for this distinction by holding the perturbation fixed and comparing different output statistics. The same intervention produces clear mean responses in some statistics and much weaker responses in others, complementing the exact zero-response constructions in Appendix B.

In Appendix B.2, we change endpoint covariance while preserving the means of all linear observables; the normalized squared mean responses of appropriate quadratic observables attain the output Fisher information. This construction distinguishes an output change invisible to the mean of a chosen observable from cancellation that preserves the output distribution. We further examine cross-time accumulation in pretrained generators through the contribution-level analysis in Appendix D.4, which isolates reinforcement among propagated caching errors.

Revision locations: Section 2, pp. 2–3; Figure 1, p. 4; Sections 3.3–3.4, pp. 5–7, and Figures 2–3, p. 6; Section 4.4 and Figure 6, p. 9; Appendices B.1–B.2, pp. 18–21; Appendix C.1, p. 22; Appendix D.2, pp. 27–28; Appendix D.4, pp. 29–30.

AuthorsOpenReview ↗

Response to Reviewer 1pjK (Part II)

W2: DCT probes, generation quality, and classifier-free guidance

We use DCT perturbations as controlled probes of how perturbation timing and spatial direction affect output responses at matched path cost. DCT-kk denotes the frequency of a global cosine pattern on the pixel or latent grid with higher indices corresponding to higher frequencies. Figure 4 normalizes each frequency row separately to display its temporal profile, with the unnormalized response maximum shown on the right. Thus, the comparison concerns frequency and timing at matched intervention strength.

We chose the measurements in Section 4 to directly assess the theoretical response quantities: mean responses of DCT coefficients in JiT and ADM, and paired decoded-image displacements in DiT. At matched path cost, these measurements reveal how sampling dynamics transform perturbations introduced at different times and in different spatial directions into unequal output effects. This experimental objective motivated our focus on responses relative to the unperturbed sampler rather than generation-quality evaluation. The larger responses in Figure 4 indicate a stronger effect on the measured outputs, not necessarily a deterioration in image quality.

Appendix D.1 provides paired reference and perturbed DiT images, together with difference maps, offering a qualitative view of how equal-cost interventions at different sampling stages affect generated images. We leave a systematic assessment of image quality and diversity across perturbation frequencies and strengths to future work, particularly for exploring your suggestion of using DCT interventions as a generation-diversification strategy.

For classifier-free guidance, we use conditional predictions without additional guidance amplification to provide a common conditional-sampling setting across the models while varying the timing and direction of the controlled perturbations. A guidance-scale sweep would additionally vary the reference trajectories and downstream dynamics through which these perturbations propagate. We therefore keep the guidance setting fixed in this study and leave the interaction between guidance strength and DCT perturbations, including its effects on image quality and diversity, to future work.

We make the connection to our framework explicit in Appendix C.7: a change in guidance strength introduces a state-dependent perturbation through the conditional–unconditional prediction difference. With prediction os=ouncond+s(ocond−ouncond)o_s=o_{\mathrm{uncond}}+s(o_{\mathrm{cond}}-o_{\mathrm{uncond}}), a change s↦s+εs\mapsto s+\varepsilon induces the local update defect ei=DoHi[ocond−ouncond]e_i=D_o\mathcal H_i[o_{\mathrm{cond}}-o_{\mathrm{uncond}}]. Propagating this defect through subsequent sampling updates and the output map gives the local output response to guidance strength. The associated path cost follows from the modified transition law, with the mean-shift formula applying when covariance is fixed. This formulation specifies how the same propagation analysis can be used to study guidance itself as an intervention.

Revision locations: Section 4.1, pp. 7–8, and Figure 4, p. 8; Appendix C.7, p. 25; Appendix D.1, p. 26.

Q1: Figure 1 presentation

Thank you for this suggestion. We have made the noise-to-output direction explicit in Figure 1's horizontal-axis label, added a density-axis label to the output panels, and specified nats next to the KL values. The expanded caption explains the state coordinate, the perturbed and reference means, and bridge cost as the average divergence between path laws conditioned on the same output. The reinforcing/cancelling comparison makes the difference in output visibility explicit.

Revision location: Figure 1 and caption, p. 4.

I appreciate the authors' response. My concerns are half-solved and half-remaining. I checked the added Figure 10 and 11, but the resulted perturbations on the generated images are rather subtle and I still have some concern whether the presented DCT-based perturbation is relevant to the quality-impactful perturbations such as classifier-based and classifier-free guidances used in application. I will keep my initial rating, for I have remaining concerns on the practical significance on the presented phenomena.

AuthorsOpenReview ↗

Thanks for your response

Thank you for the follow-up. We appreciate your concern about the practical significance and would like to clarify the role of the DCT experiments.

We use DCT perturbations as controlled probes to validate our analysis in pretrained generators. Matching path cost while varying timing and spatial direction allows us to isolate how sampling dynamics shape output responses. These experiments are intended to test the propagation analysis, rather than establish DCT perturbations themselves as a practical generation method.

Regading the the practical implications of our theoretical analysis, one application is allocating computation when accelerating sampling. Our caching experiments in Section 5 show why local approximation errors alone can be misleading: their final effects depend on downstream propagation and accumulation across steps. This suggests designing caching schedules that retain fresh evaluations in windows with larger propagated output errors and permit reuse in less sensitive windows. Another potential application is choosing when and how strongly to apply guidance. Using the guidance perturbation derived in Appendix C.7, the framework could assess how changes in guidance strength at different stages affect a selected output objective, such as conditional alignment. These sensitivities could inform adjustments to guidance intervals and weights.

We will expand the revised manuscript’s discussion to make these applications explicit and distinguish the existing empirical evidence from prospective algorithm design.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers for their careful reading and constructive comments. We appreciate the recognition of the relevance of a general framework for perturbation propagation [1pjK, QSrU], the validation using pretrained image generators [1pjK, QSrU], and the evidence connecting propagated caching errors to observed output changes [WUcv, QSrU].

Our paper develops a theoretical framework for understanding how perturbations propagate through diffusion sampling and affect generated outputs. Its central contributions are:

  • Perturbation propagation and accumulation. We decompose path cost into output-distribution and conditional-trajectory changes, and derive how local perturbations are transported through the remaining sampling dynamics and combine at the output. Together, these results explain why equal path costs can produce unequal output effects.
  • Distributional changes and measured responses. We connect the propagated perturbations to the mean response of a selected output statistic and establish an exact factorization separating output visibility from observable–score alignment. This identifies when the output distribution changes locally while the selected statistic has zero first-order mean response.
  • Empirical validation through controlled and practical perturbations. We validate the analysis through controlled interventions at equal path cost and through errors introduced by caching. The latter provides a practical sampling setting in which to examine downstream propagation and cross-time accumulation.

The reviews helped us clarify these contributions and strengthen their theoretical and empirical support. We summarize the main revisions and additional analyses below:

  • [1pjK] Theoretical positioning. We clarify the distinction from contraction and support-robustness analyses in Section 2. In Appendix B, we add exact constructions that distinguish cancellation preserving the output distribution from distributional changes invisible to selected means.
  • [1pjK, WUcv] Interpretation and presentation. We define observables and mean responses explicitly, improve Figure 1, and connect Figures 2–3 to the corresponding theoretical results. In Section 4.4, we compare different output statistics under a fixed perturbation to strengthen the empirical connection to the mean-response analysis.
  • [QSrU] Finite perturbation strength. We add caching-strength sweeps in Appendix D.3, measuring image-displacement errors separately from agreement in the ordering of window errors. We identify tested settings with relatively accurate displacement predictions and explain how window ordering can remain accurate at strengths with larger displacement errors.
  • [WUcv, QSrU] Purpose and design of the caching experiments. We clarify that Section 5 uses errors from a practical sampling approximation to validate perturbation propagation. We explain the fixed-window design and add a contribution-level analysis in Appendix D.4 to examine how errors introduced at different steps combine at the output.
  • [1pjK, WUcv, QSrU] Other interventions and evaluation scope. In Appendix C.7, we derive the guidance-scale update defect and describe how to incorporate history-dependent feature caching. In our responses, we explain the measurement choices for the controlled experiments, discuss the qualitative image examples in Appendix D.1, and identify systematic quality, diversity, and cross-method evaluations as directions for future research.
  • [WUcv] Computational cost. In our response, we compare the forward evaluations and directional-derivative computations required by propagation and direct measurement, including reference-prefix sharing. We explain how the additional propagation calculations serve the paper's goal of examining the mechanisms behind observed output changes.

Manuscript revisions are highlighted in blue. We provide detailed answers and revision locations in the individual responses below, and welcome further questions and suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.