AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising

Diffusion models have garnered significant interest from the community for their great generative ability across various applications. However, their typical multi-step sequential-denoising nature gives rise to high cumulative latency, thereby precluding the possibilities of parallel computation. To address this, we introduce AsyncDiff, a universal and plug-and-play acceleration scheme that enables model parallelism across multiple devices. Our approach divides the cumbersome noise prediction model into multiple components, assigning each to a different device. To break the dependency chain between these components, it transforms the conventional sequential denoising into an asynchronous process by exploiting the high similarity between hidden states in consecutive diffusion steps. Consequently, each component is facilitated to compute in parallel on separate devices. The proposed strategy significantly reduces inference latency while minimally impacting the generative quality. Specifically, for the Stable Diffusion v2.1, AsyncDiff achieves a 2.7x speedup with negligible degradation and a 4.0x speedup with only a slight reduction of 0.38 in CLIP Score, on four NVIDIA A5000 GPUs. Our experiments also demonstrate that AsyncDiff can be readily applied to video diffusion models with encouraging performances. The code is available at https://github.com/czg1225/AsyncDiff.

Paper

Similar papers

Peer review

Reviewer 773c6/10 · confidence 4/52024-07-09

Summary

This paper introduces AsyncDiff, an acceleration framework for diffusion models that transforms the traditional sequential denoising process into an asynchronous process. The key insight is that hidden state features in consecutive sampling steps exhibit high similarity. Therefore, feeding the output of the preceding component at time step t-1 as an approximation of the original input to each U-Net component has a negligible impact on performance. This approach allows the parallel denoising processes of AsyncDiff to operate fully asynchronously. Experiments on various versions of Stable Diffusion demonstrate that AsyncDiff significantly speeds up the denoising process for text-to-image generation. Additionally, AsyncDiff is applicable to video generation models such as AnimateDiff and SVD, further showcasing its versatility.

Strengths

1. The writing is overall clear and well-structured. The paper effectively explains the limitations of previous methods that use patch parallelism and introduces a novel approach to "asynchronous" denoising to address these issues. 2. The paper provides thorough comparisons with baseline methods, such as Distrifusion, demonstrating clear improvements in both generation quality and efficiency. 3. The versatility of AsyncDiff is evidenced by experiments conducted on different versions of Stable Diffusion as well as video generation models. 4. The authors address concerns regarding the overhead of communication costs across multiple GPUs, showing that these costs are significantly lower than the model execution time.

Weaknesses

1. Although the main contribution relies on the observation that the hidden states of consecutive steps are similar, the analysis of this phenomenon lacks details. Several key aspects need clearer explanation: (a) Can the similarity of the hidden states be quantitatively measured? For instance, does a low MSE between hidden states indicate that the two states are “similar”? (b) Is this phenomenon specific to the U-Net architecture, or is it agnostic to the backbone of the denoising model (e.g., Diffusion Transformer)? (c) Is it only applicable to DDIM sampling? Does the phenomenon also hold for other fast ODE solvers, such as the DPM solver [1]? 2. While the experiments were conducted using 50 DDIM steps, a comparison with the original model using a number of DDIM steps that achieve a similar speedup would strengthen the argument for AsyncDiff. For instance, in a setup where AsyncDiff achieves a 2.7x speedup, comparing the FID score with the original model using 50 / 2.7 = 19 DDIM steps would clearly demonstrate the necessity of parallelizing the diffusion model. [1] DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps, Lu et al., NeurIPS 2022

Questions

Most questions are included in the “Weaknesses” section. Would a similar asynchronous denoising approach be applicable in a conditional generation setup as well? For instance, achieving faster inference for diverse conditional image generation models, such as ControlNet [1] and Zero-1-to-3 [2], could be practically useful. [1] Adding Conditional Control to Text-to-Image Diffusion Models, Zhang et al., ICCV 2023 [2] Zero-1-to-3: Zero-shot One Image to 3D Object, Liu et al., ICCV 2023

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors discuss the limitations of AsyncDiff (communication cost, dependency on the base model) and the potential societal impacts.

Reviewer 7eQm7/10 · confidence 4/52024-07-10

Summary

This paper proposes AsyncDiff, a plug-and-play acceleration scheme that enables model parallelism across multiple devices. The core method involves dividing the diffusion model into multiple components and executing the inference in parallel. This is facilitated by the high similarity between hidden states in consecutive diffusion steps. With AsyncDiff, it is claimed that off-the-shelf diffusion models can be effectively accelerated with negligible degradation. Therefore, the key contribution of this paper lies in providing a simple and effective method to accelerate the diffusion process through parallelism.

Strengths

1. The proposed method is well-motivated, and the effectiveness of AsyncDiff has been evaluated on multiple models, making the results compelling. 2. This method should be useful for video generation, which can take minutes or hours to produce a single video. The paper also supports this with experiments conducted on Stable Video Diffusion. 3. As shown in Table 5, it seems that the communication cost can be covered by the inference cost, which is a favorable feature for model acceleration.

Weaknesses

1. Some concepts, such as the “dependency chain” in ABS, are not well explained. It would be beneficial if the author could provide a minimal explanation for these concepts. 2. The paper appears to share a related idea with Distrifusion [1]. Could the author provide a more detailed discussion about the key differences between them? For example, the main difference seems to be the splitting schemes, where Distrifusion splits the data, and this approach splits the model. What is the main advantage of AsyncDiff? 3. Are there any failure cases for the proposed AsyncDiff? [1] DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models

Questions

Additional Questions: 1. Is this method helpful for large batch sizes? 2. Is this method applicable to more GPUs?

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

I do not find any potential negative societal impact of this work.

Reviewer iHdZ7/10 · confidence 5/52024-07-10

Summary

This work introduces an acceleration method for diffusion models by distributing the model blocks across multiple GPUs and running different blocks asynchronously. The core motivation of this paper lies in the observation that the hidden states of a block exhibit high similarity across consecutive diffusion steps. This enables the execution of different blocks based on pre-computed results from earlier steps. The proposed method was evaluated on Stable Diffusion v2.1, v1.5, and XL, achieving a 1.8x to 4x acceleration with minimal performance loss on CLIP score and FID.

Strengths

1) The idea of parallelizing the inference of different blocks is quite interesting. The method is general and can be applied to several diffusion models like Stable Diffusion, Stable Video Diffusion, and AnimateDiff. The method shows robustness to different speed-up ratios as shown in Tables 3, 4, and 5. 2) Extensive experiments were conducted in this paper, encompassing both image generation and video generation. And the results are good: the proposed method as able to achieve significant acceleration with slight performance degradation.

Weaknesses

1) Is the communication cost huge compared to the inference cost? As shown in Table 3, the number of GPUs used in this paper is primarily 2 or 3. Will there be issues if this is extended to 8 or 16 GPUs? Is this method still competitive when the number of devices exceeds 8, where cross-node communication becomes substantial? 2) The experimental section mainly focuses on quality metrics such as FID and CLIP Score. It would be beneficial if the authors could provide more analytical results about the method, such as the similarity of hidden states. 3) Figure 5: “Quantitative” should be “Qualitative”.

Questions

1) What happens if the total number of timesteps is not divisible by the number of GPUs? Is the proposed method still applicable in this scenario? 2) Is the proposed method still effective on low-end GPUs or edge devices?

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

4

Limitations

N/A.

Reviewer E7zL6/10 · confidence 3/52024-07-13

Summary

This paper proposes AsyncDiff to enable diffusion model parallelism across multiple devices and achieve a very impressive speedup with negligible degradation. Specifically, the denoising model is split into several components, each assigned to a different device. The conventional sequential denoising process is transformed into an asynchronous one by exploiting the high similarity of hidden states between consecutive time steps, enabling each component to compute in parallel.

Strengths

1. AsyncDiff is a universal and plug-and-play acceleration scheme and can enable model parallelism across multiple devices. The acceleration that AsyncDiff achieves on image and video diffusion models is impressive 2. The motivation is clear and the experiments are comprehensive.

Weaknesses

1. The discussion and comparison with the existing diffusion sampler is not sufficient, whether AsyncDiff can be combined with other diffusion samplers such as DPM-Solver. 2. In the absence of an explanation of the GPU, and AsyncDiff's ablation experiments on different GPU devices, can other GPUs achieve a good acceleration ratio?

Questions

I do not have other questions about this paper.

Rating

6

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

As stated by the author, AsyncDiff necessitates frequent communication between devices throughout the denoising process. Therefore, if the devices lack the capability to communicate effectively, AsyncDiff may not perform optimally.

Authorsrebuttal2024-08-06

Dear Reviewers, Chairs, We sincerely appreciate the time and effort you have spent evaluating our submission, and we look forward to the discussion stage. We will include the review stage results in the appendix of our next version.

Reviewer E7zL2024-08-08

The authors have posted an effective rebuttal that resolves my concerns. I still vote for a weak acceptance of this work.

Reviewer 773c2024-08-10

After thoroughly reviewing the authors' rebuttal, I am satisfied that my original concerns have been addressed. I particularly appreciate the additional experiments with Diffusion Transformers (SD3), and the clear evidence supporting the similarity of the hidden states. In light of this, I have decided to raise my rating to "Weak Accept."

Area Chair 71QC2024-08-10

Dear reviewers, Could you please respond to the rebuttal, discuss with authors and finalize your score?

Reviewer iHdZ2024-08-13

Final Rating

The rebuttal from the author has addressed all my concerns, thus I would like to keep my rating as accept.

Authorsrebuttal2024-08-13

Thank you for the insightful and constructive comment. Your expertise and time are greatly appreciated in helping improve my work's quality.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC