Spontaneous Symmetry Breaking in Generative Diffusion Models

Generative diffusion models have recently emerged as a leading approach for generating high-dimensional data. In this paper, we show that the dynamics of these models exhibit a spontaneous symmetry breaking that divides the generative dynamics into two distinct phases: 1) A linear steady-state dynamics around a central fixed-point and 2) an attractor dynamics directed towards the data manifold. These two "phases" are separated by the change in stability of the central fixed-point, with the resulting window of instability being responsible for the diversity of the generated samples. Using both theoretical and empirical evidence, we show that an accurate simulation of the early dynamics does not significantly contribute to the final generation, since early fluctuations are reverted to the central fixed point. To leverage this insight, we propose a Gaussian late initialization scheme, which significantly improves model performance, achieving up to 3x FID improvements on fast samplers, while also increasing sample diversity (e.g., racial composition of generated CelebA images). Our work offers a new way to understand the generative dynamics of diffusion models that has the potential to bring about higher performance and less biased fast-samplers.

Paper

Similar papers

Peer review

Reviewer Z6e75/10 · confidence 4/52023-07-05

Summary

The paper explores the generative dynamic of diffusion models. It proposes spontaneous symmetry breaking and Gaussian late initialization scheme and they achieve better fidelity and diversity on the generated images.

Strengths

[S1] The paper supports the claims in both theoretical and empirical. [S2] The paper successfully convince the proposed approach improves diversity by comparing.

Weaknesses

[W1] They do not discuss the limitations and drawbacks of their model. [W2] The other generative models such as VAE, GAN, regression models, etc. do not discuss. At least, it is expected to mention them in the related work section as they claim that the proposed approach has a broader impact on generative models.

Questions

[Q1] In Table 1, why the dataset and denoising step (n) is repeated? You can put them in the first column and it might be enough. So, instead of making subtables, can you make them a single table? [Q2] In Table 1 and Figure 7 why you didn't include other generative models such as GANs in comparison? [Q3] Is Table 1 an ablation study to show the effect of the gls? If so, where is the fidelity comparison of the proposed approach? [Q4] In the discussion of Table 1, you do not discuss the results so as to indicate the spontaneous symmetry breaking. What is the reason for this?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

[L1] It is only applicable to the diffusion models. [L2] The paper ignores the other type of generative models in both related work and experiments. It is required to compare the proposed approach with existing generative models. [L3] The paper does not discuss the drawback of the proposed approach.

Reviewer FFjV4/10 · confidence 4/52023-07-07

Summary

The authors approach the reverse diffusion process in score-based generative modeling from the perspective of spontaneous symmetry breaking, which may explain why generation is minimally affected by "late initializations" (e.g. initializing x_t at t=T, but simulating the reverse process starting at some t < T). Motivated by this framework, the authors propose late-start scheme and show improved performance compared to some baselines.

Strengths

**Exciting direction of study.** The authors propose an exciting perspective on the diffusion model in terms of spontaneous symmetry breaking in physical systems. **Initially promising empirical results.** The empirical results appear to be promising. (However, they do not presently provide much in the way of concrete proof that the symmetry breaking framework provides tangible improvements on generation speed or quality, see "Weaknesses").

Weaknesses

**Motivation.** The motivation of the work should be made more clear. If it is to improve generation speed, the model should be compared against fast generation models, such as Model Distillation, Consistency Models, and EDM. If it is to generate better samples, then the model should be compared against standard diffusion models such as [4] or [5]. **Empirical Results.** The empirical results are not very convincing. All models achieve FIDs that are far from state-of-the-art. It is not clear at this stage whether improvements come directly from the proposed approach, and whether the improvements will remain when models are tuned to be more competitive with respect to the state of the art. **Claims.** The claim (up to 3x FID improvements) feels overstated. The authors do not compare against state-of-the-art fast samplers, such as Model Distillation [1], Consistency Models [2], and EDM [3]. **Soundness.** I have some concerns over the soundness of the claims in the paper. - The existence of stable points is only shown for toy examples (e.g. Secs. 3.1 and 3.2). Even in these toy examples, it is not clear to me how the stable points will affect the result of the reverse diffusion process for most initial conditions of the reverse process. Showing the existence of stable points at the origin and the original data points implies only the existence of a finite set of stable points, which are essentially probability zero set of the entire space. Diffusion models are initialized from a standard normal, so x_T can be very far from the origin (or any of the stable points). So is it true that the spontaneous symmetry breaking model is relevant here, even in theory? - While the second toy example is more realistic (Sec 3.2), and the authors claim that the assumptions required are mild, I do not believe they hold for any usual image distributions (e.g. MNIST, CIFAR10, ImageNet, CelebFaces, etc.). - Finally, the authors connect the theory of symmetry breaking to diffusion modeling with a single piece of empirical evidence: The FID of samples drawn with late start diffusion is relatively stable (up to a critical) point in time t_c. I can see how this is necessary condition to show the existence of symmetry breaking. But how is it sufficient? In summary, while the theory of spontaneous symmetry breaking is very interesting and thought-provoking, I am not yet convinced that it is presently a useful perspective in diffusion-based generative modeling. **Minor edits.** 39: the theory weak and electromagnetic -> the theory of weak and electromagnetic 108: Hessian matrices (of what?) [1] Progressive Distillation for Fast Sampling of Diffusion Models. https://arxiv.org/abs/2202.00512 [2] Consistency Models. https://arxiv.org/abs/2303.01469 [3] Elucidating the Design Space of Diffusion-Based Generative Models. https://arxiv.org/abs/2206.00364 [4] Denoising Diffusion Probabilistic Models. https://arxiv.org/abs/2006.11239 [5] Score-Based Generative Modeling through Stochastic Differential Equations.https://arxiv.org/abs/2011.13456

Questions

See "Weaknesses" section.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

See "Weaknesses" section.

Reviewer kBdj6/10 · confidence 4/52023-07-09

Summary

The authors investigate theoretically and empirical when (in terms of diffusion timestep) certain symmetries in a diffusion process are broken, corresponding to when "choices" are made about which qualitative features a generated data point should have. Their findings suggest that there is no symmetry breaking occuring early in the process. They follow this up by showing that the early dynamics can be replaced by a sample from a multivariate Gaussian without meaningful degradation of the perceptual quality of generated data.

Strengths

- This perspective on the diffusion process is novel as far as I know. - Understanding when in the diffusion process different symmetries are broken is likeliy to be very helpful for future work designing better diffusion-based models. - Their technique for better fast sampling yields surprising large improvements. Although they acknowledge that it does not immediately scale to high-dimensional data, it provides good support of their analysis.

Weaknesses

- There is a fairly large gap between the simple example in Section 3.1, where we could reason analytically about when the symmetry is broken, to the realistic case in Section 3.2, where the only symmetries considered were the identity of each data point. Do the authors think it would be possible or meaningful to do this type of analysis on other features that could be extracted from the data (e.g. the identity of a person in an image)? - This analysis of symmetry breaking makes sense for "discrete" variables, like cluster index. It does not significantly help to provide insights about when values of "continuous variables" like e.g. background colour of an image are selected. - The proposed Gaussian initialization improves fast samplers in terms of FID and the authors show is helpful in terms of maintaining the identity of discrete variables, but it is important to note that replacing the early diffusion process with a Gaussian may still lead to large changes in the distribution.

Questions

See weaknesses (mainly the first one)

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

sufficiently addressed

Reviewer J7iR7/10 · confidence 4/52023-07-10

Summary

The authors of this work propose a new method to accelerate the sampling of diffusion models. First, the authors define as fixed points of the reverse process, points where the drift function is 0. Then, the authors claim that if two paths of stable fixed points interact, then it is the noise that can make the system jump from one path to the other. The authors argue that there are phase transitions happening in the sampling of diffusion models: until a certain time, the stable paths interact and hence the system jumps between potential paths. After a point in the reverse diffusion, there are no more intersections in the paths. Even if we don't initialize exactly in the path, the stability ensures that the differential equations are mean-reverting and will correct any introduced errors. Hence, the authors propose to start the sampling from a point in the middle of the diffusion and accelerate the sampling.

Strengths

The introduced framework is very interesting and novel. The example with the two Dirac functions is very pedagogical and its' generalization to a uniform distribution over many discrete points follows naturally. The experimental evidence supports that this phenomenon is indeed happening in trained diffusion models. The proposed method surpasses the baselines for a low number of function evaluations. The paper is well-written and the approach can be implemented easily. I expect that the paper will be of great interest to the research community and the audience of NeurIPS.

Weaknesses

The theoretical analysis has some simplifying assumptions that it is not clear to what extent they alter what happens in practice. Specifically, the authors analyze uniform distributions over Dirac functions. The distributions observed in practice have a more complicated structure and it is not clear how we can argue anything about the fixed points of a model given access to its (learned) score-function. A weakness of the approach is that to start the sampling from some intermediate diffusion time, one needs to estimate the parameters of a multivariate Gaussian. This can be a particularly challenging task and it doesn't scale well for high dimensions because of the quadratic parameters needed for the covariance. The authors acknowledge this limitation. Experimentally, the authors mostly consider what happens in the low NFEs regime. More detailed comparisons would be useful. If the method performs poorly (relative to the baselines) for higher NFEs, it is important to acknowledge it.

Questions

* It seems to me that the method can be used *together* with other approaches for accelerating sampling (since it is only changing the initial point). It would strengthen the paper if the authors included experiments to show this. * Can the authors comment on whether we can find fixed points given a trained diffusion model? * I am puzzled by what happens to the story that this paper builds once we think about the deterministic samplers. In such settings, stable points would correspond to no movement at all. If the system becomes mean-reverting at some point, wouldn't that mean that we stop moving once we reach a fixed point? * It would strengthen the paper to include comparisons with EDM and other more recent sampling methods. * It would also help to include results for higher NFEs.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations.

Reviewer sN4H7/10 · confidence 3/52023-07-14

Summary

The authors examine denoising diffusion models and their dynamics through the lense of spontaneous symmetry breaking. They characterize the dynamics of the denoising process into two stages: in the beginning, samples orbit a central fixed point and after a critical amount of time denoising of a ‘selected’ sample begins. They use insights gained from theory as well as practical examples to develop a novel fast sampling method they call “Gaussian late start” (GLS).

Strengths

This was a very well written paper. The authors motivate the issue very well, and they additionally illucidate the concept of spontaneous symmetry breaking in a way that is easy to understand. One of the biggest strengths, was the foundational example laid out in Sec. 3.1 — this example helped fill in a lot of gaps I had up until that point, and made me really appreciate the authors characterization of the denoising dynamics into two separate stages. The experiment of section 4.1 followed by its extension in section 4.2 do appear to give credence to thinking of the denoising dynamics as having a spontaneous symmetry breaking characteristic. This also leads in nicely to the Gaussian late start sampling method, which is both intuitive, practical, and appears to work well. Overall, the paper presents a compelling idea, motivates it through practical toy examples as well as theory, and then provide a novel efficient way of sampling.

Weaknesses

It seems like this is not a problem constrained solely to diffusion models used for image generation, but for any data modality. In that case, it might be good to empirically demonstrate this for another data modality.

Questions

This may be a silly question, but I was wondering if the authors thought if the datasets clustered and the sampler was initialized using the statistics of each cluster — do you expect the samples drawn to be similar to the cluster whose statistics were used to initialize the sampler?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

As noted, GLS as is, is slightly impractical, but there is room for future research to develop fast approximate methods in the same vain.

Reviewer sN4H2023-08-17

Thank you to the authors for answering my questions. I maintain my original score which is to accept.

Reviewer J7iR2023-08-12

Response to Rebuttal

Thank you for the rebuttal. I will keep my score which recommends acceptance. Please incorporate the rebuttal discussion in the main paper, if possible.

Reviewer Z6e72023-08-16

Response to Rebuttal

Thank you for the rebuttal. I think it is good to enlarge the discussion by adding other generative (types of) models to demonstrate a complete picture and allow others to have better insight into how the idea can be generalized. Besides, I carefully read the rebuttal and the other reviews and rebuttals and reconsidered my rating as borderline accept.

Reviewer kBdj2023-08-17

Thanks for the response, which addressed my concerns. I will keep my rating of 6.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC