acceptodds
Withdrawn from ICLR 2027

Generative Modeling with Probability-Preserving Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern generative modeling usually produces samples by transporting a simple noise prior to the data distribution. We study a complementary setting: once a useful state already exists, can a pretrained flow be reused as a stochastic transition rather than restarted from noise? We construct two reusable kernels directly from pretrained flow checkpoints, without retraining: a denoiser-based Metropolis kernel (dMALA), whose ideal target is the Gaussian-smoothed law , and a predictor–corrector kernel whose ideal target is . We also characterize how one-step model and numerical errors accumulate under a local contraction assumption. On ImageNet-256, dMALA repairs deliberately weak generated states from FID to while retaining high similarity to the input (DINO ). The same dynamics also support controlled variation: as the chain length increases, samples move progressively farther from their inputs, providing a controllable transition from local variation to broader exploration. Experiments on a synthetic Swiss-roll target, ImageNet-256, and Oxford Flowers-102 show that these dynamics can reuse existing flow models when a useful initial state is already available.

Primary area:
generative models
Keywords:
flow matching, probability-preserving dynamics, predictor–corrector methods

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iNV5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Flow-based generative methods usually find a mapping from a simple prior (often a Gaussian distribution) to the data distribution. This paper proposes to reuse pretrained flow models to define probability-preserving stochastic dynamics, allowing the mapping of points sampled from or near the learned data distribution. To do so, they propose two schemes: 1) dMALA, which uses denoising and Metropolis-adjusted Langevin dynamics, and 2) a predictor–corrector scheme consisting of a prediction step, i.e., noising via the bridge (interpolation) used by the underlying flow model, followed by a correction step: running the pretrained flow from the noised version back to a data sample, either via numerical integration of the ODE or using a flow map model. This allows pretrained models to be used to refine corrupted data samples or explore the data manifold from predefined states.

They establish invariance properties for the idealized kernels and provide a Wasserstein perturbation bound to study approximation errors. The theoretical development seems partially supported by experiments on synthetic distributions, Flowers-102, and ImageNet demonstrating promising results

Critical strengths
  • The problem seems well motivated, and the paper is well written, making the theory easy to follow.
  • The theoretical results appear correct under the stated assumptions.
  • The experiments cover several settings and demonstrate promising results for sample refinement.
Critical weaknesses
  • It is hard to judge the novelty and advantages of the method. I think it would help to move the related work section to the main paper to better motivate and contextualize the problem. The experiment section would benefit from comparing with related methods, such as SDEdit and Restart Sampling (which is currently not cited). This can be discussed, but as things stand, it is not clear to me how the proposed method differs from or improves upon these existing methods (although I agree the problems aren't exactly the same). Is the main contribution the probability-preservation interpretation, or does the proposed method also offer practical advantages in terms of sample quality, diversity, or computational cost? A more direct comparison would help clarify this point.
  • Although mentioned in passing in Appendix C.4.3, it isn't clear to me that the theory holds in the ImageNet case, nor that the algorithm instantiated in this setting corresponds to the method developed theoretically in Section 2. Given the effort and emphasis placed on the theory, this disconnect seems like a potential issue. Is there a way to verify the assumptions, even empirically, in this setting? Or to justify the instantiated method and its acceptance rule?
Questions for authors
  • Could you clarify the main differences between the problem at hand and related methods and what is the advantage of the proposed method ?
  • Could you touch on the difference between the method employed in the ImageNet case and the one described in Section 2. Does the theory hold in this case? Can it be shown empirically?
  • How do you explain the deterioration of FID over longer dMALA chains, intuitively, in the limit where we evaluate FID on many samples, shouldn't we roughly preserve FID if the data distribution is conserved ?
Reviewer_KWSrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new problem setting: how to move the data within the data distribution and preserving it at every step. The authors proposes two methods to achieve this: a denoiser-based Metropolis kernel (dMALA) whose ideal target is the Gaussian smoothed law pσp_{\sigma} and the predictor–corrector kernel whose target is the original distribution pp. In the experiments, the authors show that with the proposed method, the FIDs remain stable as the baseline method. The Swiss roll example shows that the distribution remain stable as the baseline method.

Critical strengths

In the experiments, the authors use both toy data and image data to show the effectiveness of the method. On the Swiss roll dataset, the authors show the distribution of the initial state and the states after the dMALA and the predictor–corrector method. Also, on the ImageNet dataset, the authors use FIDs, Inception scores to show the method preserves the distributions in Table 1. The authors also show qualitative results when the number of steps increases in Figure 4.

Critical weaknesses

I think this paper lacks a real concrete application. What real application requires us to move the data points and preserve it at every step? Is it significant? Can the authors give an experiment for it? If it is to generate images, I can use a pre-trained flow model with random sampling to draw samples that fall in the same distribution.

Proposition 1 seems to be trivial. x^\hat{x} and x+x^{+} were constructed in the way showing in Eq. 24 and 25. It's obvious that x+∼p1x^{+} \sim p_1

The example in Fig. 2 is not reasonable. The authors chose s(x)=−xs(x) = -x, and claimed that it does not preserve the target law exactly for finite step size. In fact, s(x) needs to be trained, and also the hh can be chosen vanishing as number of steps grow.

Questions for authors

I feel Sec. 2.4 APPROXIMATE KERNELS is disconnected to Sec. 2.2 and Sec. 2.3. Can the authors clarify the relationship between Sec. 2.4 and Sec. 2.2 and 2.3 and how the APPROXIMATE KERNELS are used?

Reviewer_4KuzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes dMALA, which treats pretrained flow-matching models as a Langevin-like diffusion kernel. This enables the generation of new samples based on an existing sample.

Critical strengths
  • The paper is clearly written and easy to follow.
  • The derivation of the re-noise Langevin kernel form is solid.
Critical weaknesses
  • The proposed method is very similar to an inversion process [1] combined with stochastic inference [2]; there are also some papers that adopt Langevin-like diffusion in generative models [3] to enable controllable generation. The novelty is substantially limited for me.
  • The experimental results show limited importance in the CV field (20 NFE is expensive for large models); I encourage the authors to try in the AI4Science field, where controllable generation is more meaningful.

[1] Meng, Chenlin, et al. "Sdedit: Guided image synthesis and editing with stochastic differential equations." arXiv preprint arXiv:2108.01073 (2021). [2] Zhang, Qinsheng, and Yongxin Chen. "Fast sampling of diffusion models with exponential integrator." arXiv preprint arXiv:2204.13902 (2022). [3] Kim, B., & Ye, J. C. (2022). Denoising MCMC for accelerating diffusion-based generative models. arXiv preprint arXiv:2209.14593.

Questions for authors

see weakness

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

Related papers

Discussion (0)

Sign in to comment.