- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes EDDY, a training-free particle guidance method for improving sample diversity in diffusion and flow matching models. The method aims to reduce the image quality degradation caused by conventional repulsive guidance. It introduces particle interactions based on anti-symmetric matrix fields and the Stein operator. The authors provide a theoretical analysis of how these interactions affect particle independence and marginal distributions. They also propose a negatively correlated initialization and efficient approximations to support high-dimensional image generation. Experiments on synthetic data and three text-to-image models show improved diversity and image quality compared with existing particle guidance methods.
Critical strengths
-
Well-motivated method design. The paper identifies an important limitation of existing particle guidance methods and proposes a principled approach based on the Fokker–Planck equation. The connection between distribution preservation and particle interactions provides a clear motivation for the method.
-
Meaningful theoretical analysis. The paper studies how particle interactions affect independence and marginal distributions. These results help explain the role of correlated initialization and provide useful insights into diversity-oriented sampling.
-
Strong experimental results. The method is evaluated on three text-to-image models, including FLUX.1, FLUX.2, and SDXL. The results show favorable trade-offs between diversity and image quality across multiple evaluation metrics. The experiments also include comparisons with existing methods and a synthetic study.
Critical weaknesses
-
A gap between the theoretical assumptions and the actual initialization. Claim 3 assumes that the joint particle distribution has a sufficiently smooth density. However, the negatively correlated initialization in Claim 4 produces a degenerate Gaussian distribution, since the particles always sum to zero. Under deterministic flow matching dynamics, this distribution remains supported on a lower-dimensional set. Therefore, the assumptions used to derive Claim 3 do not directly apply to the proposed initialization. This weakens the theoretical justification for the marginal bias analysis in the actual sampling setting.
-
Insufficient analysis of the sources of diversity improvement. The paper acknowledges that numerical discretization can introduce particle dependence even with independent initialization. However, the experiments do not fully separate the effects of correlated initialization, divergence-free guidance, and discretization in practical text-to-image generation. This makes it difficult to determine which components are responsible for the observed improvements. A more detailed ablation across different sampling steps and initialization strategies would strengthen the empirical support for the proposed design.
Questions for authors
Can the authors clarify whether Claim 3 applies to the negatively correlated initialization used in EDDY? This initialization produces a singular joint distribution, while the proof assumes a smooth joint density and uses conditional density scores. In particular, how can the marginal bias formula be justified for deterministic flow matching models, where the joint distribution may remain singular throughout sampling?
Could the authors clarify the FID evaluation protocol? Section 6.2 and Appendix B.2 indicate that FID is computed by comparing generated samples against samples from the I.I.D. baseline, rather than real images. Since standard FID measures the distance between generated and real data distributions, could the authors explain the motivation for this choice? How would the conclusions change if FID were computed against real MS-COCO images?
3 comments
Thanks to the author for the reply, I will keep my positive rating.
We thank the reviewer for the helpful comments and positive rating
We appreciate the reviewer stating the paper presents a well-motivated, empirically strong method while providing a meaningful theoretical analysis.
Note. we reordered claims 2 and 3 in the revised manuscript. Claim 2 is now concerning marginal bias, while Claim 3 is about the propagation of independence. Our responses below use the new numbering.
CW1/Q1 - marginal bias claim with degenerate coupling or deterministic flows. We appreciate you pointing it out, and fixed it in the new revision using smooth compactly supported test functions (see appendix for the proof) without changing the claim itself. Regardless, adding an arbitrarily small perturbation to the suggested prior coupling solves the problem.
CW2 - diversity attribution to method’s components. We extended our ablation study in a new section under “Experiments” exploring hyperparameter sensitivity and component-wise diversity attribution. It turns out a significant part of diversity can be attributed to time discretization and that the coupled prior has almost no effect.
Q2 - FID methodology. First, we measure FID under the same protocol we measure CMMD - we choose a single image for each prompt and form a set, and compare the set guided (EDDY, PG, etc.) to the set of unguided (i.i.d sampled). Since we generate 4 images per prompt, we create 4 non-intersecting sets to measure the uncertainty of the metric. That way, we measure how much bias we introduce to the marginal distribution. Second, we do not use MS-COCO or any other type of dataset since the target marginal distribution is the one generated by the unguided backbone.Furthermore, MS-COCO is composed of photographs which do not capture digital art that many backbones may generate (see figure 1).