acceptodds
Withdrawn from ICLR 2027

Martingale Residuals for Evaluating and Training Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models are typically trained with noise or score targets. We introduce a martingale residual (MR) framework for evaluation and learning from sampled reference paths, without requiring these targets. A martingale residual is the observed change along a reference path minus the change predicted by the model. We prove that, under suitable assumptions, MR orthogonality implies that the learned and reference processes have the same path law. We combine weighted residual expectations into an MR score. When the learned process uses the reference process's known covariance, the MR score formed from all coordinate test functions converges to the squared drift error as the time interval shrinks, provided the instruments form a complete orthonormal basis. Finite instruments give a projection approximation. For DDPMs, suitable residual scaling connects this limit to denoising MSE minus a constant independent of the model parameters. We evaluate the MR score as a diagnostic, a training regularizer, and a sole training objective. On CIFAR-10, MR scores track denoising training, with richer instrument sets capturing a larger fraction of the denoising MSE decrease. MR regularization yields modest improvements in sample quality during continuation of pretrained CIFAR-10 and ImageNet-64 models. Using the MR score alone, we learn stochastic transport from a Gaussian mixture to a two-moons distribution in a Schrödinger bridge task.

Primary area:
generative models
Keywords:
diffusion models, martingale residuals, conditional moment matching, stochastic differential equations, Schrödinger bridges

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZykWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript proposes a metric, martingale score, for optimizing, regularizing, or evaluating denoising diffusion models. The primary distinction from standard diffusion model setup is, it is assuming the trajectory data of the noisification process is given and used for parametrizing the neural models, contrast to the typical diffusion model setup that an analytical score or reverse-time drift is provided as regression target.

The theoretical setup of the method can be roughly summarized into the followings. The authors consider the reference process - the process of noisification - and a modeling process, characterized by their drifts and diffusions. If the drift of the modeling process is the same as the reference process, subtracting the model-predicted drift displacement from the reference displacement - here integrated to some finite horizon hh - results in a process that is a martingale with respect to reference process-induced filtration. Conversely, if the drift-offsetted process is a martingale with respect to the reference process-induced filtration, then the modeling process and the reference process have the same path laws. This is true under the following conditions (1) both processes are of the same Itô process form, with shared covariance matrix driving the Wiener noise, (2) both processes have the same initial law, (3) the relevant martingale problem is well-posed. With these asumptions, the martigale property becomes a metric that can be quantified, optimized, and evaluated through the defined martingale residual (MR), without knowing the score or reverse-time drift.

More technical nuances include the standard weak-form formulation for the stochastic process (on (x,t)(x,t) domain) and instrument functions. As most weak-form methods, the optimization in the end has to be carried out with a finite set of instrument functions, leading to Galerkin-type projection. The manuscript went through the technical nuances and provide a concrete numerical method to estimate the MR. The authors finally used the quantity to perform various tasks, including learning a diffusion model directly from MR, using MR to regularize or evaluate a diffusion model.

Critical strengths
  • [S1] The reviewer likes the alternative construction of the MR metric stemming from weak-form and martingale theories. These analytic tools provide a principled approach that defines the MR score.
    • [S2] Although an elementary fact, transferring the requirement of a complete set of test function φ\varphi to only the linear φ(x):=x\varphi(x):=x to a complete set of instrument function ZZ is very elegant.
    • [S3] It is pleasing to see that under the assumption of a complete instrument function set and the infinitesimal-horizon (h↓0h \downarrow 0 limit), the MR score recovers the Lμ2L^2_\mu of the drift discrepancy between the reference process and the modeling process, establishing the connection to the conventional DDPM training objective.
Critical weaknesses
  • [W1.1] It is challenging for the reviewer to identify a scenario where the reference path is given and not the typical DDPM or score-based approaches. The typical workflow of diffusion models is to define the forward corruption process, which is usually simple zero-drift Wiener process (VE) or Ornstein-Uhlenbeck process (VP) and use Denoising Score Matching (Vincent, 2011) to use initial sample-dependent score, which is analytically tractable for the choices of these processes as the regression target to solve for the population-level score function. It appears to the reviewer that the manuscript still operates on these analytically tractable processes, raising the question that what advantages the MR score framework brings.

    • [W1.2] Although the theory seems to be consistent, the resulting method seems to be extremely similar to some form of flow, generator, or score-matching methods. In addition, the proposed method in this manuscript critically depends on the choice of the instrument function ZZ and the time discretization (hh). While the reviewer understood the setup of the problem is tied to only trajectories with finite-time (h>0h>0) separation are given and most of these flow/generator/score-matching methods are using instantaneous samples, the manuscript can benefit from stating what the advantages of the proposed MR framework over the instantaneous matching framework.

      • [Instantaneous samples] A special case is DDPM, which is in discrete-time nature, that the targets are derived from consecutive snapshots (Xk,Xk+1)(X_k, X_{k+1}); this case is particularly relevant as one can argue the DDPM's setup is also "trajectory-based", "with finite hh time seaparation", of an underlying Ornstein-Uhlenbeck process (arxiv:2311.17673).
    • [W2] In practice,

      • a finite set of instruments instead of the theoretical complete set of basis function
      • a fixed hh finite time grid and a scalarization of the strict martingale condition (The classical martingale condition imposes an infinite family of conditional constraints for every pair of times s<ts<t.)

      It is unclear to the reviewer if this construct really reveal the path-law discrepancy, especially over existing path-law based methods (e.g., "Improved Sampling via Learned Diffusions", ICLR, 2024). Can the authors establish a quantitative relationship between their finite-horizon (h>0h>0) and scalrized MR score and a standard path-space divergence, rather than merely identifying a common short-time limit? More concretely, does a small {\rm MR}= \Vert P_\cal{S} \mathbb{E}_\cal{Q} \left[ r_{t,h}^\theta \mid t, X_t \right] \Vert^2_{L^2 \left( \mu \right)} imples a small path-law divergence from Girsanov change of measure with the standard Kullback-Leibler divergence: DKL(Q∥Pθ)=12∫0tEQ[(b−bθ)T⋅(σσT)−1(b−bθ)]dt.D_{\rm KL} \left ( Q \Vert P_\theta \right ) = \frac{1}{2} \int_0^t \mathbb{E}_\mathcal{Q} \left[ (b-b_\theta)^T \cdot \left( \sigma \sigma^T \right)^{-1} (b-b_\theta) \right] {\rm d}t.

      It is unclear to the reviewer if the MR score is more statistically efficient, robust, brings more computational advantages.

    • [W3] The paper’s framing as a diagnostic, regularizer, and standalone training method is stronger than the evidence supports. As a diagnostic, MR mainly tracks denoising MSE, without demonstrating clear added value over standard metrics. As a regularizer, it is compared only against denoising alone, so the reported FID gains are not shown to be specific to the MR construction. As a standalone objective, convincing from-scratch results are largely limited to low-dimensional synthetic problems; at image scale, the strongest result uses pretrained initialization, while from-scratch CIFAR training is unstable or substantially worse than denoising. The authors should either strengthen these claims with targeted diagnostic examples, meaningful regularizer baselines, and competitive image-scale from-scratch experiments, or reframe the paper as an exploratory study of MR objectives with strong population theory but mixed practical performance.

    • [W4] The reviewer is skeptical about numerically solving the min\rm min-max\rm max formulation given in Eq. (11). These saddle-point problems are notoriously challenging to optimize reliably, often exhibiting oscillation, instability, and hyper-sensitivity to optimization parameters. In addition, a learned finite set of instrument functions does not guarantee to recover the completeness of instrument functions.

Questions for authors
  • [Q1] The manuscript establishes a connection between martingale-residual minimization and the conventional squared drift-matching objective. The reviewer wonders about the distinction of the proposed framework, particularly given existing interpretations of diffusion training as L2L^2 velocity-field matching, including Wasserstein gradient-flow matching (Vuong et al., "Are We Really Learning the Score Function? Reinterpreting Diffusion Models Through Wasserstein Gradient Flow Matching", TMLR, 2026). Could the authors clarify what contribution of this manuscript is a new theoretical characterization, an alternative estimator/parametrization for learning, or an improved diagnostic, given the existing path-law based methods and flow/velocity-matching approaches?

    • [Q2] What is the intended practical information regime in which MR is preferable to existing trajectory-based SDE inference methods? In particular, can the authors give a realistic example where large numbers of reference trajectories and the covariance are available, but direct drift/score supervision is unavailable and standard drift-estimation, stochastic-force-inference, weak-form, or likelihood-based approaches are inadequate? A convincing answer would materially strengthen the significance of the method.

    • [Q3] The finite-set instrument results are closely related to standard Galerkin-type Extended Dynamic Mode Decomposition (EDMD, Williams et al., "Extended Dynamic Mode Decomposition with Application to the Navier-Stokes Equations", SIAM Journal on Applied Dynamical Systems, 2015) of stochastic generators. The matrix construction G†CG^\dagger C corresponds to the least squares projection of the martingale residual onto a finite-dimensional observable dictionary (standard notation in EDMD/DMD literature; here, "instrument"). Could the authors clarify the relationship to standard EDMD, Dynkin formula-based generator matching, and Petrov-Galerkin approaches and illustrate the necessity/advantages of the MR formulation?

    • [Q4] In Theorem 3.1, the admissible test-function class is defined relative to the reference law QQ, but the conclusion is that QQ solves the martingale problem for LθL_\theta. What conditions ensure that this admissible class is a sufficiently rich domain/core for the learned generator so that the stated martingale identities determine MP(Lθ,μ0)MP(L_\theta,\mu_0)? If an additional core/domain assumption is needed, please state it explicitly.

    • [Q5] Can the authors compare MR regularization against at least one meaningful alternative auxiliary regularizer or matched-compute/matched-gradient control? The current denoising v. denoising+MR comparison does not establish that the observed FID gains are specific to martingale residuals.

    • [Q6] Can the authors provide a diagnostic example in which MR reveals a meaningful model defect or improvement that is not already apparent from denoising loss, FID, or another standard metric? The current diagnostic experiment mainly establishes that MR can track denoising MSE when the instrument family is sufficiently rich.

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

This paper presents a method for evaluating and improving diffusion models based on martingale residuals, which allows for the analysis of how well the generation process aligns with a reference process without direct access to the score or noise targets. The authors derive mathematical conditions under which martingale residuals allow for the identification of trajectory distributions, and then construct an additional loss function based on them for training diffusion models. Experiments show that this approach can improve the quality of generated images, but its effectiveness depends on the configuration of the test functions and instruments, as well as the dataset.

Critical strengths

The theoretical framework appears solid and well justified mathematically. The authors use stochastic process theory and martingale residuals to formally characterize the agreement between a reference process and a learned model. Particularly interesting is the result showing that drift agreement can be assessed without directly knowing the reference drift.

Critical weaknesses

The main methodological concern is the unclear description of the reference process. The proposed algorithm requires samples of reference trajectories from Q, but the paper does not clearly explain where these samples come from or what information is assumed to be available in practice. The paper is difficult to follow and lacks a clear narrative. The introduction is not very accessible, and the related work section is rather limited. Furthermore, the authors introduce numerous abbreviations and cross-references, which make the presentation harder to follow. The proposed method introduces an additional, relatively complex loss function whose practical effectiveness remains uncertain. The experiments are fairly limited and do not establish whether the approach generalizes to a broader range of settings. The results also show that its effectiveness depends on the choice of test functions and instruments.

Questions for authors
  1. What are the potential practical applications of the proposed method, and in which settings could it offer an advantage over standard diffusion model training?
  2. How should test functions and instruments be selected in practice? The results suggest that using more DCT coefficients does not necessarily improve FID. Are there theoretical or empirical guidelines for predicting which configurations will be effective without costly hyperparameter tuning?
  3. Where do the reference trajectories used by the algorithm come from? Does the method require samples from a reference process whose drift depends on the unknown score? If so, how do the authors justify the practical availability of such samples?
Reviewer_sdMpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a trajectory-level framework for evaluating and training diffusion models based on Martingale Residuals (MR). By projecting finite-step coordinate drift residuals onto a dictionary of causal instrument functions with cross-fitted sample splitting, the authors show that the population MR score converges, matching the projected excess DDPM denoising risk under appropriate weighting. Experiments test MR as a diagnostic on 1D mixtures, continue pretrained CIFAR-10 and ImageNet-64 checkpoints, and train a 2D Schrödinger bridge drift network.

Critical strengths
  1. Clean theoretical formulation: Connecting generator discrepancies, coordinate martingale residuals, L2(μ)L^2(\mu) subspace projections, and excess DDPM denoising risk (Theorems 3.1–3.3) is mathematically clean and well-structured.
  2. Transparent reporting of negative results: The authors openly document failure modes in the text and appendices (e.g., FID degradation under large instrument sets in Table 10 and the failure of from-scratch CIFAR-10 training in Appendix E.7).
Critical weaknesses
  1. Null-space blindness and bias–variance dilemma as a high-dimensional training objective: With a finite set of J instruments, the MR objective only penalizes the subspace projection ||P(b* - b)||², leaving the infinite-dimensional orthogonal complement ||(I - P)(b* - b)||² completely unpenalized. When training from scratch on CIFAR-10 (Appendix E.7), the network exploits this null space: it reduces the targeted MR score by 99.6%, yet its true denoising MSE (0.1333–0.1348) remains more than 2× worse than standard denoising training (0.0609). Conversely, increasing J to cover more directions inflates finite-sample covariance estimation variance, degrading CIFAR-10 FID from 2.0128 to 3.8723 at J = 9,219 (Table 10). Due to this dilemma, all image experiments (Section 4.2–4.3) are restricted to continuing already converged checkpoints for 500–5,000 steps with tiny learning rates (1e-7 on CIFAR-10 and 2e-8 on ImageNet-64), yielding negligible FID changes (1.9916 to 1.9818).
  2. Unjustified need for instrument projections in training and missing direct trajectory-regression baseline: In trajectory-only regimes (such as Schrödinger bridges, Section 4.3), the conditional orthogonality property used in Appendix B.4 implies that the Brownian martingale increment in r(t, h) = (X(t+h) - X(t)) / h - b(t, X(t)) has zero conditional mean given (t, X(t)). Consequently, the expected squared error E[ ||(X(t+h) - X(t)) / h - b(t, X(t))||² ] equals the full unprojected drift error ||E[r(t, h) | t, X(t)]||² plus a Brownian variance constant that does not depend on the model parameters. Thus, direct least-squares regression of the drift network b(t, X(t)) onto the finite-difference velocity (X(t+h) - X(t)) / h (as in DSBM, Shi et al., 2023) directly optimizes the full unprojected objective without instrument truncation bias, Gram matrix inversion, or min-max auxiliary networks, yet this standard baseline is omitted from the 2D Schrödinger bridge comparison (Table 3).

References

  • Jonathan Ho et al., Denoising Diffusion Probabilistic Models, NeurIPS 2020.
  • Yang Song et al., Score-Based Generative Modeling through Stochastic Differential Equations, ICLR 2021.
  • Yuyang Shi et al., Diffusion Schrödinger Bridge Matching, NeurIPS 2023.
Questions for authors
  1. Since the parameter gradient of the direct velocity regression loss E[ ||(X(t+h) - X(t)) / h - b(t, X(t))||² ] equals the gradient of the full unprojected objective ||E[r(t, h) | t, X(t)]||² without instruments, why is projecting onto a finite instrument set S(J) (or training min-max auxiliary networks) preferable to direct L2 velocity regression for training the drift network, and how does direct velocity regression perform in Table 3?
  2. Given the from-scratch failure on CIFAR-10 (Appendix E.7) and the FID degradation at large J (Table 10), under what realistic high-dimensional regime can MR serve as a practical training objective rather than purely an evaluation diagnostic?

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

Related papers

Discussion (0)

Sign in to comment.