Self-Refining Diffusion Samplers: Enabling Parallelization via Parareal Iterations

In diffusion models, samples are generated through an iterative refinement process, requiring hundreds of sequential model evaluations. Several recent methods have introduced approximations (fewer discretization steps or distillation) to trade off speed at the cost of sample quality. In contrast, we introduce Self-Refining Diffusion Samplers (SRDS) that retain sample quality and can improve latency at the cost of additional parallel compute. We take inspiration from the Parareal algorithm, a popular numerical method for parallel-in-time integration of differential equations. In SRDS, a quick but rough estimate of a sample is first created and then iteratively refined in parallel through Parareal iterations. SRDS is not only guaranteed to accurately solve the ODE and converge to the serial solution but also benefits from parallelization across the diffusion trajectory, enabling batched inference and pipelining. As we demonstrate for pre-trained diffusion models, the early convergence of this refinement procedure drastically reduces the number of steps required to produce a sample, speeding up generation for instance by up to 1.7x on a 25-step StableDiffusion-v2 benchmark and up to 4.3x on longer trajectories.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Peer review

Reviewer KhBd6/10 · confidence 5/52024-07-09

Summary

This paper proposes a new method for parallelizing the sampling process of diffusion models, offering a novel way to trade compute for speed. In particular, it relies on the parareal iteration in this setting, where the main idea is to split the solving timesteps into multiple groups and then update them in parallel. To ensure that the output from these timestep groups matches the exact solution, parareal performs predictor-corrector steps after each group is updated. Experiments demonstrate the efficacy of applying parareal iteration in speeding up the sampling process.

Strengths

1. This paper demonstrates a very interesting direction by introducing the parareal iteration into diffusion models. 2. The experiments indeed demonstrate the efficacy of this approach.

Weaknesses

1. I think the author misses a rigorous discussion comparing the parareal algorithm with ParaDiGMS [1] and also completely ignores a more recent reference [2]. It is clear that all these works, including parareal, focus on parallelizing the sampling process via certain forms of fixed-point iteration, although the computing graphs of these algorithms differ. The only discussion in the current manuscript is that ParaDiGMS is highly memory-intensive and must use a sliding window. While this statement is certainly true, I believe it is not a disadvantage of ParaDiGMS and does not imply that ParaDiGMS is inferior to parareal. I do not see why using a sliding window should be considered a disadvantage. 2. As far as I can see, the Eff. Serial Evals in Table 1 and Table 2 should correspond to the Parallel Iters in [1] and Steps in [2], am I correct? Because Eff. Serial Evals essentially represent the number of batch evaluations on the models. If this is the case, based on the current numbers in Table 1, it seems quite inferior to the results in [1] and [2]. Of course, I understand that the batch size may be different in this work, but I still wish the author would provide a detailed discussion on this point. 3. Proposition 1 in the manuscript is somewhat trivial and exists in many existing works. Note that each refinement consists of $\sqrt n$ sequential evaluation, so the parareal will also require at most $\sqrt n \times \sqrt n=n$ sequential evaluation on the models to converge. I think the author should mention that this result is fundamentally equivalent to Proposition 1 in [1], and a more general result is found in Theorem 3.6 in [2]. The key point is that when performing fixed-point iteration for an $n$-order triangular system, it will certainly converge within $n$ steps, and this bound can not be further improved without extra assumption. [1] Shih, Andy, et al. "Parallel sampling of diffusion models." Advances in Neural Information Processing Systems 36 (2024). [2] Tang, Zhiwei, et al. "Accelerating parallel sampling of diffusion models." Forty-first International Conference on Machine Learning. 2024.

Questions

1. In my understanding, the technique in Parareal have a different spirit in parallelization with [1][2]. For example, consider that we wish to do n-step sampling with $\sqrt n$ batchsize. Given a initial sequence $[x_1,...,x_n]$, parareal first group them into $g_1=[x_1,...,x_{\sqrt n}],...,g_{\sqrt n}=[x_{n-\sqrt n},x_n]$. The parallelism is that to update $g_1$, ..., $g_n$ in parallel, while the update of each $g_i$ consists of $\sqrt n$ sequential solving. On the other hand, [1][2] consider update $[x_{i},...,x_{i+\sqrt n}]$ in parallel until convergence. It is not clear to me which one is better from the theoretical aspects, as they have the same worst-case convergence bound. I suspect that it should depends on the form of ODE and also the batchsize? I wonder if the authors could provide some thought and discussion on this point? 2. Do you think parareal can be applied to the SDE sampler as [1] and [2] did? [1] Shih, Andy, et al. "Parallel sampling of diffusion models." Advances in Neural Information Processing Systems 36 (2024). [2] Tang, Zhiwei, et al. "Accelerating parallel sampling of diffusion models." Forty-first International Conference on Machine Learning. 2024.

Rating

6

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

See my comments on weaknesses and questions above. This work lacks a correct and thorough discussion comparing it to existing works.

Reviewer pK7d6/10 · confidence 3/52024-07-10

Summary

This paper proposes Self-Refining Diffusion Samplers (SRDS), which draws inspiration from the Parareal algorithm, aims to solve the reverse process accurately without retraining models, balance the tradeoff between sample quality and sampling speed. lower latency for requiring fewer sampling steps to reach convergence. The experiment section shows that the proposed algorithm speeds up the convergence without degrading sample quality in image generation tasks.

Strengths

1. This paper is motivated by Parareal algorithm to speed up sampling process without lowering sample quality which seems to work well with image generation. 2. This paper provides theoretical analysis for sampling convergence and latency investigation. 3. This paper observes the pipelined SRDS, which further speeds up the sampling process.

Weaknesses

1. In related work, the authors mentioned that ParaDiGMS proposed by Shih et al [1] is another parallel-in-time integration method, and this paper also works on image generation. In this case, I believe their results could be baselines that this paper should compare against. [1] Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, and Nima Anari. Parallel sampling of diffusion models. Advances in Neural Information Processing Systems, 36, 2024.

Questions

1. For the experiment section, are the numbers intentionally matched in ``FID Score`` in Table 1 and ``CLIP Score`` in Table 2? If that's true, it would be also interesting to see how the performance changes when you have different SRDS Iters and their corresponding Total Evals.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

This paper discussed the limitations and pointed out the potential positive or negative social impacts will not be the direct consequence of this work in the checklist guidelines.

Reviewer XqLH3/10 · confidence 4/52024-07-10

Summary

Inspired by the parallel-in-time ODE integration literature, especially Parareal method, this paper introduces SRDS as a fast sampler for diffusion models enabling efficient parallelization. Experimental results demonstrate that the proposed SRDS reduces the number of steps required to synthesize samples.

Strengths

It is interesting to use parallel sampling algorithms to accelerate the sampling of diffusion models, which will benefit practical applications such as real-time image editing.

Weaknesses

My concerns are mostly about the limited experimental results: 1. This paper lacks of clear discussions about the differences between the proposed method and the previous work “23NeurIPS-Parallel sampling of diffusion models.”, and sufficient experimental evidence to demonstrate the effectiveness of the proposed method. Specifically, there is no quantitative comparison with the most related method ParaDiGMs. It seems that the existing method ParaDiGMs can provide better speedup according to their results. 2. This paper lacks of a sufficient comparison to the existing fast samplers (such as Heun, DEIS, DPM-Solver). In practice, we can already synthesize samples with those advanced samplers with only around 10 steps. While this paper only presents experiments on several hundreds or one thousand steps at Table 1, which makes the method proposed in this paper unappealing. 3. It is claimed in Lines 38-40 and 196-198 that SRDS outperforms ParaDiGMs in terms of memory requirement. However, there is also no quantitative support from experimental results. 4. It is claimed in Line 62-64 that SRDS is compatible with off-the-shelf solvers. However, the speedup is only shown on DDIM. The speedup offered by SRDS is only explicitly shown in Table 2, making it hard to fully assess the effectiveness of the proposed method. 5. There is no quantitative comparison between SRDS and pipelined SRDS.

Questions

1. Could the pipelined SRDS be easily deployed in practice? 2. Why the SRDS Iters shown in Table 1 are not integers? 3. Why the first and second rows of Figure 6 are almost the same? 4. Minor points: (1) In line 10 of Algorithm 1, $x_{i-1}^p$ is undefined for p == 1. (2) It is redundant to use Figure 2, Figure 3b and Algorithm 1 to describe the same thing in the main text. They together occupy a whole page. Adding more experiments to support the effectiveness of the proposed method is more valuable.

Rating

3

Confidence

4

Soundness

2

Presentation

1

Contribution

2

Limitations

Please refer to the weakness and question sections above. This manuscript needs an overhaul to make it meet the bar of a top conference.

Reviewer MmLD6/10 · confidence 4/52024-07-11

Summary

This work proposes SRDS, a sampler for diffusion models that applies the Parareal algorithm, to reduce the overall sampling latency by introducing extra but parallelizable network evaluations compared to the fully sequential fashion. With higher device utilization or device parallelism through batched inference and pipelining, SRDS successfully reduces the overall sampling latency of existing diffusion models without sacrificing the sample quality.

Strengths

- This paper is well written. - The proposed SRDS sampling pipeline is flexible and extendable, providing a neat baseline and rich directions for future technical improvements. - The experiment successfully demonstrates SRDS's practical value of reducing the sampling latency compared to full sequential sampling. - The ablation studies provide a nice practical guidance for SRDS.

Weaknesses

- As the authors addressed in the limitation part, SRDS relies on parallelizing extra computations in exchange for a reduction in total latency, which may not be applicable to some scenario that performs full batch sequential sampling. However, I believe SRDS could still work in a fairly broad range of cases. - The important ParaDiGMs baseline is not compared in the experiment part. ### Minor - The meaning of the abbreviation "IVP" in line `1:` of Algorithm.1 is unclear.

Questions

- In the current SRDS pipeline, the intermediate results of the fine-grained solver are only used once. Do you have any thoughts on utilizing them in subsequent steps for better sample quality or faster convergence?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations of their work.

Reviewer dCML6/10 · confidence 3/52024-07-11

Summary

The paper presents a new approach to speed up (improve latency) the generation of samples from diffusion models. The approach is orthogonal to many other approaches present in the literature for the same task. It leverages Parareal algorithm for the task by getting a quick coarse approximate of the sample and then refining it iteratively in parallel, thus reducing the latency while maintaining sample quality. The authors present results for pre-trained pixel space diffusion models and Stable Diffusion and find upto 2x speed ups for the latter.

Strengths

Improving the latency of diffusion models is an active and important area of research. The paper presents a new strategy that relies on leveraging parallel computation of modern hardware. This approach can in theory be combined with other strategies to lead to further improvement, and does not require re-training. The presented approach seems to have guarantees on the quality of the solution and some control on trading off the speed and quality. There is also thought given to efficient batching and pipelining tasks for further potential improvements, which is good. Overall, the paper is also well written and easy to follow.

Weaknesses

The main weakness of the paper is lack of strong and diverse results. The authors only test one diffusion model task. Furthermore the best speed up is only 2.3x, while another experiment results in 0.7x speed up which is concerning. The authors argue on theoretical grounds that this approach can be combined with other approaches to reduce latency of diffusion models. This is fair, but they do not show any empirical results around it. Finally, the approach relies completely on leveraging parallel refinement and the actual number of model evaluations are much larger (often by a factor of 3x) which can limit the applicability in some cases.

Questions

- The authors claim that the worst case sampling latency is no worse than generating a sample through sequential sampling. Yet one of their results (last on in Table 2) has speed up of 0.7x. Is this all due to GPU overhead? - Is is possible to present empirical results with pipeline parallelism? It seems like the authors expect 2x gains with it, but I assume it will require some computational overhead and based on the previous question, the degradation can be non-trivial. - While I understand the claim regarding this approach being orthogonal to other works reducing, showing some experiments around the same will make for a stronger paper.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

There are a few limitations discussed towards the end of the paper.

Authorsrebuttal2024-08-07

Thanks for the reviews!

We sincerely hope that the additional experimental results/clarifications in the rebuttal clearly demonstrate/validate the usefulness of our proposed method (SRDS), thereby sufficiently answering all the reviewers’ concerns and significantly improving the strength of the paper. We once again thank the reviewers for their time and feedback, and we are happy to answer any further questions that may remain in the discussion period! --- Author Rebuttal References: - [1] Shih, Andy, et al. "Parallel sampling of diffusion models." Advances in Neural Information Processing Systems 36 (2023). - [2] Tang, Zhiwei, et al. "Accelerating parallel sampling of diffusion models." Forty-first International Conference on Machine Learning. 2024.

Reviewer dCML2024-08-09

Response to rebuttal

I thank the authors for their responses and new experiments. I am satisfied with their rebuttal and will maintain my score.

Reviewer pK7d2024-08-11

I have read through the responses and decide to increase the score from 5 to 6.

Reviewer MmLD2024-08-12

I would like to thank the authors for the rebuttal, especially the additional results. I will keep my current recommendation unchanged.

Authorsrebuttal2024-08-14

Thank you for your response and for raising the score! We are glad that your concerns have been addressed.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC