SA-Solver: Stochastic Adams Solver for Fast Sampling of Diffusion Models

Diffusion Probabilistic Models (DPMs) have achieved considerable success in generation tasks. As sampling from DPMs is equivalent to solving diffusion SDE or ODE which is time-consuming, numerous fast sampling methods built upon improved differential equation solvers are proposed. The majority of such techniques consider solving the diffusion ODE due to its superior efficiency. However, stochastic sampling could offer additional advantages in generating diverse and high-quality data. In this work, we engage in a comprehensive analysis of stochastic sampling from two aspects: variance-controlled diffusion SDE and linear multi-step SDE solver. Based on our analysis, we propose \textit{SA-Solver}, which is an improved efficient stochastic Adams method for solving diffusion SDE to generate data with high quality. Our experiments show that \textit{SA-Solver} achieves: 1) improved or comparable performance compared with the existing state-of-the-art (SOTA) sampling methods for few-step sampling; 2) SOTA FID on substantial benchmark datasets under a suitable number of function evaluations (NFEs). Code is available at https://github.com/scxue/SA-Solver.

Paper

Similar papers

Peer review

Reviewer vQfu6/10 · confidence 3/52023-06-17

Summary

The authors propose a sampling method for diffusion probabilistic models by solving an alternative SDE with the same marginal distribution. Approximation techniques are applied in constructing the computationally efficient solver and corrector . The results show a superior FID with less function evaluations.

Strengths

1. The construction of the alternative SDE and the approximation techniques, including score function approximation, change of variable and Lagrange interpolation, provide a complete framework for efficient DFM generation. 2. The proposed method can achieve lower FID scores with less function evaluations which is promising.

Weaknesses

1. The paper content seems not well organized yet. 2. In the appendix, for other methods, with different NFEs, the image content is consistent. But for the proposed method, the image content varies a lot. It might be caused by the inaccurate approximations.

Questions

1. Could you elaborate on why the images are not consistent for different NFEs in your method?

Rating

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors did not address the limitations. There is no potential negative societal impact of their work.

Reviewer zg3j7/10 · confidence 5/52023-07-04

Summary

This paper extends UniPC to a stochastic manner to obtain a faster sampler called SA-Solver for diffusion models. Specifically, the authors start from the formulation of diffusion SDE and derive the SA-Predictor and SA-Corrector. Extensive experiments on CIFAR10, ImageNet, LSUN, etc demonstrate the effectiveness of the proposed SA-Solver.

Strengths

1. It is interesting to consider sampling from the SDE instead of ODE. This opens up a new direction for the fast sampling of DPMs. I am surprised that this simple practice can improve the sampling quality so much. 2. The writing is clear and the derivation of SA-Solver is easy to understand. I especially appreciate Section 5.3, where the relationship with UniPC helps me to understand the proposed method at a glance. 3. The experiments are thorough and convincing. Most required experiments are included, such as unconditional/conditional sampling with data/noise prediction with different image resolutions.

Weaknesses

1. It is questionable whether the novelty of SA-Solver is enough. As said in L228, SA-Solver can be viewed as an extension of UniPC with a non-zero $\tau(t)$. However, I think the idea to introduce stochasticity back to the sampling of the diffusion models also contributes to the novelty. I just feel worried that this issue might be raised by other reviewers. 2. Some experiments are lacking, such as comparing the convergence of different methods on COCO using a pre-trained text-to-image diffusion model. For now, only some qualitative results are provided. It would be better to also demonstrate some quantitative comparisons. 3. Some minor format issues: too much space below Table 1. Maybe the authors can try to make it looks better.

Questions

See the weaknesses.

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitations have been fully discussed.

Reviewer VoWY5/10 · confidence 3/52023-07-12

Summary

The paper proposed a stochastic Adam solver for solving diffusion SDEs in an efficient way with a convergence guarantee. Authors adapt the stochastic Adam from numerical literature and use Lagrange interpolation to predict unknown terms. They show strong convergence for both predictor and corrector. Numerical experiments are done to verify their claims.

Strengths

1. By deriving an explicit form of the solution of the SDE, authors are able to utilize stochastic Adam with Lagrange interpolation to find an efficient diffusion SDE solver, 2. Theoretical analysis is well-written. 3. Numerical experiments show improvements compared to other solvers, especially the diffusion ODE solvers. 4. It sort of unifies previous methods in some sense.

Weaknesses

1. My main concern is, from Figure 2, it seems SA-solver only outperforms other methods within a certain range of NFE. This "optimal" range looks very different for different data sets. In reality, if I use SA-solver for the sake of doing fewer function evals, how do I know when it outperforms other methods? 2. I don't think I fully understand how to choose the parameter tau(t) in the experiments. Please comment on this.

Questions

See weaknesses.

Rating

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

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

yes

Reviewer txgZ6/10 · confidence 4/52023-07-15

Summary

The paper proposes a new solver for diffusion SDEs, termed SA-Solver, combining the ideas of a predictor-corrector scheme and the stochastic Adam solvers. The predictor and corrector utilize the Lagrange polynomials for extrapolation to lower the approximation error at future time stamps. Experimentally, SA -Solver improves over previous ODE and SDE across a wide range of image generation benchmarks.

Strengths

- The considered problem of accelerating SDE solvers is of great practical value in the field, as SDE solver often delivers better sample quality than ODE solvers, but are hindered by their slow sampling speed. - Adapting the idea in Stochastic Adams is interesting, offering higher-order convergence with Lagrange polynomial extrapolation. - The fusion of the PC sampler and Lagrange polynomial yields superior empirical results on diffusion SDE, tested on a range of dataset resolutions from 32x32 to 256x256. It's nice to see the methods scale to different resolutions and architectures.

Weaknesses

- Throughout the paper, the authors mention several times that empirically the quality of data generated by SDE has a better upper limit. I think some theoretical analysis in the main text to make the paper more self-contained. A concurrent [1] provides some theoretical arguments and comparisons between SDE and ODE, showing that the stochasticity in SDE can reduce overall sample errors. I think it would be beneficial to discuss them to better motivate the idea of the paper. - The authors didn't provide detailed discussions of the benefit of introducing $\tau(t)$ (at the end the author set $\tau(t)=1$ in experiments). Since the authors observe the benefit of larger $\tau$ on larger NFE, why not use a varying $\tau$ in Fig.2 based on NFE? In addition, the benefit of a larger $\tau$ on larger NFE seems predictable based on the theory in [1]. - The authors did some comparisons of epsilon-prediction and data prediction models. It seems that EDM [2] uses interpolation between these two by pre-conditioning. Does it offer better results? - Unlike the deterministic path in ODE, the randomness in SDE can change the sample trajectories quite a bit. For example, the trajectories of generating a dog can stray into generating a cat. I wonder if this is an issue for using Lagrange polynomials in SDE, since the preceding predictions could be misleading. If not, could the author provide some intuitions? In addition, the authors mention that "predictor step 2 and corrector step 1 is the most stable setup". such small steps could be caused by the issue I raise. - The paper separately gives the convergence order of predictor and corrector. Is it possible to give the theoretical order after combing the predictor and corrector? - I think the author didn't use the "EDM VE" model for CIFAR-10. In Table 2 of EDM [2], the unconditional VE (in config F, EDM VE) can achieve an FID of 1.98 using only 35 NFE, as opposed to the >3 FID in Fig 2 in the current paper. So I guess the author use their VE model in config A (baseline VE). Could the author compare different methods using the VP model (config A), as in [1] and [2], because the VP model is more commonly used in practice and offers better results? - The stochastic sampler in the concurrent work [1] obtains an FID ~ 1.8 using less than 100 NFE on ImageNet-64 with Pixel DPM, as opposed to the > 2.1 FID in Fig 2 of the current paper. I wonder if one could combine the multi-step idea in this paper with [1] to obtain further improvements. - Could the author also provide quantitative results on Stable Diffusion experiments, to better showcase the advantage of SA-Solver? Current visualized images do not seem compelling. ### Minors - line 225, DPM-solver++ and UniPC are *ODE* solvers which would not naturally be special cases of the proposed *SDE* solver. [1] Restart Sampling for Improving Generative Processes, Xu et al, https://arxiv.org/abs/2306.14878 [2] Elucidating the Design Space of Diffusion-Based Generative Models, Karras et al, https://arxiv.org/abs/2206.00364 # Post rebuttal Thanks the authors for the rebuttal. After another pass of the paper, it occurred to me that that paper is a direct application of the ideas (variation-of-parameter, or exponential integrator, and Langraian polynomial for approximation) in [1] to SDE samplers. The same formula of the exponential integrated version of SDE (Eq.5) has been proposed in prior works [1] (see their eq.17, arXiv version 1). In addition, [1] also uses a Lagrange polynomial to extrapolate. Due to it's straightforward extension of prior works, I will keep my score unchanged. [1] Zhang, Qinsheng, and Yongxin Chen. "Fast sampling of diffusion models with exponential integrator." arXiv preprint arXiv:2204.13902 (2022).

Questions

I copy-and-paste some questions from above: - Since the authors observe the benefit of larger $\tau$ on larger NFE, why not use a varying $\tau$ in Fig.2 based on NFE? - I wonder if this is an issue for using Lagrange polynomials in SDE, since the preceding predictions could be misleading. If not, could the author provide some intuitions? In addition, the authors mention that "predictor step 2 and corrector step 1 is the most stable setup". such small steps could be caused by the issue I raise. - Is it possible to give the theoretical order after combing the predictor and corrector? - Could the author compare different methods using the VP model (config A), as in [1] and [2], because the VP model is more commonly used in practice and offers better results? - Could the author also provide quantitative results on Stable Diffusion experiments, to better showcase the advantage of SA-Solver?

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes

Reviewer UcRr7/10 · confidence 4/52023-07-26

Summary

This paper studies the problem of fast sampling of diffusion models. Since standard DDPM sampling is slow, this is a very important topic of late, with many competing methods. Many methods reformulate as a ODE solving problem, which makes it easier to do few-step sampling. However, it has been noted that sampling from the original SDE formulation (as opposed to ODE) can lead to better samples if there is budget for taking many diffusion steps. Therefore, this paper studies the problem of accelerating sampling of the SDE. The method, named SA-Solver is a predictor corrector method. It incorporates some recent findings, such as semi-linearity, but mainly uses the method of Lagrange interpolations. More specifically, the predictor step involves (uniquely) fitting a polynomial of degree s-1 using s predictions and reading off the polynomial value at desired points. Finally, a correction step is incorporated that plug in the predicted value at the new timestep (t+1) back into the formula to choose a new predicted value (similar in spirit to heun's method). The paper draws connections to DDIM (deterministic version of 1-step SA-Predictor), DPM-Solver (deterministic version of 2-step SA-Predictor), and UniPC (deterministic version of SA_Solver) The paper shows sampling results on cifar10, imagenet64, and imagenet256 (latent), demonstrating best FID scores (conditioned on the same number of NFEs) when using more than 30 NFEs.

Strengths

The problem being studied is an important one, as sampling speed is arguably the biggest weakness of diffusion models (and diffusion models are being used in a huge number of applications right now). The experimental results seem quite promising, showing better FID scores than DDIM/DPMSolver, which are very popular baselines.

Weaknesses

Seems to be missing quantitative evaluation (e.g. CLIP score on stable diffusion) for text-to-image tasks. Also, evaluating on different domains besides image would improve the empirical results.

Questions

How much of the performance is due to the predictor/corrector approach, versus the lagrange interpolation? An ablation study in this regard would improve the paper and our understanding of the method.

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

4 excellent

Limitations

N/A

Reviewer tWFq5/10 · confidence 4/52023-08-01

Summary

The paper presents a multistep SDE solver for diffusion models instead of ODE solvers. The main goal is to have diverse and high-quality samples while reducing the number of solver step required. To do this, the paper proposes a new SDE that includes an additional term, ensuring the marginal distribution unchanged. Based on this new SDE, the paper adopts stochastic Adam method, introducing two key component: SA-predictor, and SA-corrector. The convergences of these two are provided. Experiments are conducted in different scenario including varying stochastic noise scale, different models. The results support the claims on the reduction in NFEs.

Strengths

The paper has the following strong points: - The experiment shows state-of-the-art result in term of FID given a limited NFEs. This highlights the effectiveness of the proposed method in achieving impressive results with efficient resource utilization. - It provides theoretical studies on the convergence of proposed method, offering valuable insights

Weaknesses

The main weaknesses I find in this paper are: - In term of the contribution to difussion model research, the proposed method may be considered incremental. This is because the key factor leading to good result, such as predictor-corrector, multi-step have been well-established in the existing literature. - The paper presentation can be improved, in particular the exposure of stochastic Adam methods. I would be good to dedicate a small paragraph or section explicitly discussing ths method to provide readers with clear understand of how it is employed in the proposed approach. - The paper lacks clarity in establishing the connection between the variance-control SDE and the use of stochastic Adam methods. The discussion should be more thorough and elaborate, especially in addressing why setting $\tau=1$ (falling back to DDPM) is the best choice in many cases. Then, stochastic Adam significantly contributes to the model's overall performance.

Questions

How many of predictor step $s_p$ and corrector step $s_c$ is used? Minor points, - Line 199, please add the notation stepsize $h$ - Eq (13), please add the description for the Lagrange basis $l(t)$. - Caption of Fig. 1, "scholastic" -> "stochastic" - There are some typos in Appendix

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

2 fair

Contribution

3 good

Limitations

The paper mentioned a current limitation is that the optimal $\tau(t)$ is unknown, which I agree. There many factors affecting this parameters including NFE, dataset and approximated score function.

Reviewer UcRr2023-08-10

Response

Thank you for the response, the ablation study is excellent. I stand by my recommendation of acceptance. A1: for text-to-image tasks, papers on fast samplers rarely report the FID and CLIP scores, as all solvers achieve similar FID results [1][2]. These works attribute this to the powerful pretrained decoder, which can map a non-converged latent code to a good image sample Having a powerful decoder is a property of latent diffusion (as opposed to pixel diffusion), and is not an issue of text-to-image vs unconditional generation right? In other words, this statement seems to also imply Imagenet 256x256 with latent DPM is easy?

Authorsrebuttal2023-08-13

Thank you for your comment. We believe that two factors are present, influencing the FID of sd v1.5 on MS-COCO. One is the encoder/decoder architecture and the other is the text-to-image tasks. The rationale behind the latter aspect has not been definitively established yet. However, we hypothesize that improved sample quality doesn't necessarily guarantee that the sample will resemble the COCO sample under the same prompt. Please let us know if you have any other questions. Thanks again for your valuable review!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC