acceptodds
Under review as a conference paper at ICLR 2027

Conditional Particle Matching: State-Conditioned Particle Fields for Few-step Generation

Abstract

Drifting Models train generators by moving their endpoint pushforwards toward the data distribution, while Drift Flow Matching extends this principle to time-indexed marginal transports. Marginal matching alone, however, does not uniquely determine how clean predictions depend on a particular noisy state: distinct reverse kernels can induce the same clean marginal. We introduce Conditional Particle Matching (CPM), a population conditional reverse-KL formulation that targets the corruption posterior with a stochastic reverse kernel . This stochastic family can, in principle, represent multiple clean explanations of the same corrupted state. At the population level, the conditional reverse-KL field decomposes into a data-prior term, an analytic likelihood term from the known corruption process, and a state-conditioned model self-field. Guided by this decomposition, we develop a practical feature-space particle semi-gradient built on the released Drifting implementation. CPM operates in clean-endpoint coordinates: frozen perceptual features act exclusively on clean-coordinate predictions and references, while the practical update incorporates explicit noisy-state dependence through the analytic corruption likelihood and an SNR-dependent normalized distance in native latent space. The formulation further motivates fixed-state sampling and observation-shuffle controls to probe conditional behavior that pooled distributional metrics alone cannot reveal.

Primary area:
generative models
Keywords:
Few-step Generation, Particle Methods, Generative Models

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

This paper proposes Conditional Particle Matching (CPM), a particle-based generative framework that targets clean explanations conditioned on each noisy input rather than only the pooled clean marginal. CPM combines a conditional reverse-KL objective, an analytic likelihood term from the known corruption law, and native-state particle affinities, and supports one-step and multistep sampling via clean prediction followed by explicit re-noising. The paper provides theoretical results on conditional versus marginal matching, pure-noise endpoints, and stochasticity requirements, and evaluates CPM on synthetic data, MNIST, FFHQ, and ImageNet-256.

Critical strengths

1.The paper identifies a well-motivated and conceptually important gap: matching pooled clean marginals does not identify the corruption posterior, and the conditional formulation is clearly formalized.

2.The CPM construction is coherent, combining particle interactions, an analytic likelihood term, and native-state affinities with a clean-prediction-then-re-noise sampling interface.

3.The empirical study is broad, and the ImageNet-256 result of FID 1.30 and IS 311.2 at four NFEs is strong. The paper is also transparent about negative or inconclusive evidence.

Critical weaknesses

1.The conditional mechanism is not fully isolated empirically. Shuffled affinity retains most of the endpoint gain, and the pairing diagnostic is significant only at log-SNR +2. The authors acknowledge this, but the evidence currently supports the composite recipe more than conditional recovery specifically.

2.The theory-to-implementation gap is not quantified. The analysis assumes exact conditional matching, while the implementation uses finite particles, kernel affinities, frozen features, and finite-step rollouts. A diagnostic or empirical sensitivity analysis would clarify when the practical method is close to the population object.

3.The comparison protocol is not sufficiently matched. FFHQ uses a local benchmark and local tuning budgets, while ImageNet baselines are largely quoted from prior work. Capacity-, compute-, and tuning-matched comparisons are needed to isolate the effect of conditional particle training.

4.Conditional evaluation remains thin. FID, EMD, and a single reassignment diagnostic do not directly measure per-input posterior coverage, conditional correctness, or input-output association across noise levels.

Questions for authors

1.Can the authors provide a direct test that isolates the conditional mechanism, comparing CPM-full, shuffled affinity, and marginal matching under identical architecture, training budget, and seeds, with conditional metrics beyond FID?

2.Can the authors quantify how finite particle count, bandwidth, feature dimension, and affinity approximation affect conditional KL or posterior recovery, even empirically?

3.Please provide matched capacity-, compute-, and tuning-budget comparisons against Drifting, DFM, MeanFlow, and marginal-matching variants, including the planned matched ImageNet comparisons.

Reviewer_s6nNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper identifies the issue of Drifting (Deng et al. 2026) that matching marginal distributions does not necessarily ensure that generated outputs are consistent with their corresponding noisy inputs. To address this issue, the authors propose Conditional Particle Matching (CPM), which extends Drifting with a conditional reverse-KL formulation and state-conditioned particle interactions. The experimental results show promising performance across several benchmarks, particularly for few-step generation.

Critical strengths
  • The paper identifies a limitation of marginal distribution matching and motivates the need for a conditional formulation.
  • The experimental results appear to be good, demonstrating competitive performance in few-step image generation.
Critical weaknesses

1. Unclear practical motivation for conditional matching.

All generation starts from pure Gaussian noise, and intermediate noisy states are obtained by re-corrupting the model's own outputs. It is therefore unclear why learning the conditional posterior is necessary or practically beneficial over marginal matching for the proposed generation task. The fact that conditional matching imposes a stronger theoretical constraint does not, by itself, justify its practical relevance.

2. Lack of evidence for meaningful refinement.

The proposed multi-step sampling repeatedly corrupts and regenerates its own outputs, but it is unclear whether this process genuinely refines samples or simply performs repeated resampling. The theoretical guarantee establishes marginal preservation, not sample-wise improvement. Controlled experiments are needed to demonstrate that conditional matching actually enables meaningful refinement beyond marginal resampling.

3. Lack of theoretical justification for the proposed algorithm.

There appears to be a significant gap between the theoretical derivation (Eq. 7 and Appendix B.2) and the actual training algorithm (Appendix F). The paper fails to clearly establish how the practical feature-space update follows from or approximates the derived conditional KL gradient.

4. Insufficient ablation of the conditional matching mechanism (Sec. 4.4).

The experiments do not convincingly establish the contribution of the proposed conditional particle interactions. Shuffling the state affinities barely changes FID (8.866 vs. 8.960), while the Marginal-PM baseline removes both the conditional particle mechanism and the likelihood term. Thus, it remains unclear whether the improvement comes from the proposed conditional particle matching or simply from the likelihood term. A likelihood-only baseline and direct evaluation of conditional matching quality are necessary to substantiate the paper's central contribution.

Questions for authors

The weakness above are closely tied to the paper's central motivation and contributions. In particular, I would like the authors to clarify the practical benefit of conditional matching over marginal matching, and justify the connection between the theoretical formulation and the practical algorithm. Without clearly addressing these fundamental concerns, I would find it difficult to support acceptance.

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

The paper proposes Conditional Particle Matching (CPM), which extends Drifting models to generate clean images consistent with the noisy input. The method combines particle interactions in perceptual feature space, a likelihood term based on the noise model, and weights determined by similarity between noisy inputs. It predicts clean images directly and supports additional sampling steps by renoising these predictions. Experiments use synthetic data, MNIST, FFHQ, and ImageNet. Experiments show an improvement from the combined training recipe over pooled marginal matching, but do not clearly establish the benefit of conditional weighting. The effect of additional sampling steps is also mixed across the reported metrics.

Critical strengths

none

Critical weaknesses
  1. The benefit of conditional particle weighting is not clearly established. In 4.4 authors report that full CPM achieves FID 8.866, compared with 8.960 after shuffling the conditional weights and 9.332 for pooled marginal matching (baseline). The combined additions improve over pooled matching, but correct conditional weights provide only a small additional gain. The pooled baseline also removes the likelihood term, so the comparison does not isolate which component helps.
  2. The evidence for improved quality with additional sampling steps is incomplete. The main ImageNet model is reported only at 4 NFEs, so it's unclear whether additional steps improve performance. On FFHQ, extra steps seem to worsen FID. The paper also doesn't show how additional steps influence input/output consistency.
Questions for authors
  1. Could the authors evaluate CPM ImageNet at more NFEs (for example: 1, 2, 4, 10 and maybe 20) to show whether additional steps improve quality?
  2. Could the authors extend Fig4 by reporting paired MAE feature errors and PSNR/SSIM across different NFEs for CPM and its control variants using the same noisy inputs and clean targets?

Comments on the paper

1 comment
Area_Chair_R2h5OpenReview ↗

Author-Reviewer Discussion ends in 1 day, 12 hours

Dear authors and reviewers,

Thank you to the authors for their detailed responses of Nov 12 and the updated manuscript. The discussion period closes in about a day and a half, so this is the last chance for a direct exchange before the reviewer-AC discussion starts on Nov 19.

Reviewers: I haven’t yet seen a reply from any of you. Please read the authors’ responses and the revision, which address several of your points directly. Please let us know whether your concerns have been addressed, and update your recommendation if your assessment has changed. Even a short acknowledgment helps.

Authors: please be ready to answer any follow-up questions in the time left, focusing on the points most likely to affect the decision.

Thank you all for keeping the discussion constructive and timely.

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.