Semi-Implicit Denoising Diffusion Models (SIDDMs)

Despite the proliferation of generative models, achieving fast sampling during inference without compromising sample diversity and quality remains challenging. Existing models such as Denoising Diffusion Probabilistic Models (DDPM) deliver high-quality, diverse samples but are slowed by an inherently high number of iterative steps. The Denoising Diffusion Generative Adversarial Networks (DDGAN) attempted to circumvent this limitation by integrating a GAN model for larger jumps in the diffusion process. However, DDGAN encountered scalability limitations when applied to large datasets. To address these limitations, we introduce a novel approach that tackles the problem by matching implicit and explicit factors. More specifically, our approach involves utilizing an implicit model to match the marginal distributions of noisy data and the explicit conditional distribution of the forward diffusion. This combination allows us to effectively match the joint denoising distributions. Unlike DDPM but similar to DDGAN, we do not enforce a parametric distribution for the reverse step, enabling us to take large steps during inference. Similar to the DDPM but unlike DDGAN, we take advantage of the exact form of the diffusion process. We demonstrate that our proposed method obtains comparable generative performance to diffusion-based models and vastly superior results to models with a small number of sampling steps. The code is available at https://github.com/xuyanwu/SIDDMs.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Peer review

Reviewer 6duC6/10 · confidence 4/52023-06-27

Summary

This paper proposes the Semi-Implicit Denoising Diffusion Model (SIDDM) to enable fast sampling while maintaining high-generation quality. Specifically, SIDDM applies an implicit model to match the marginal distributions of the reverse diffusion process. Meanwhile, SIDDM models explicit conditional distribution of the forward diffusion. Furthermore, a regularization method is proposed to enhance the performance. Experiments show that SIDDM has comparable performance to other diffusion models with fewer sampling steps.

Strengths

1. The authors analyse the reasons for the limitations of DDGAN, and decomposite the denoising distribution to improve the training objective. The idea is reasonable. 2. The experiments in the simulated Mixture of Gaussians and several popular public datasets demonstrate the effectiveness of the proposed method. 3. The authors also provide the code for results reproduction, which shows the solidness of the work.

Weaknesses

1. In Fig. 1, there are dashed lines of different colours in Figure 1. But the authors don't give some corresponding explanation. 2. The implementation details are not clarified, like the structure of the regression model, the discriminator regularizer, and the denoiser. Meanwhile, the training settings (e.g., training iterations) are not provided.

Questions

1. In Fig. 3 and Tab.5, the larger step (e.g., 16) performs worse than smaller step sizes (e.g., 2). This is different from the conventional situation in diffusion models. Please give some analysis. 2. It is recommended to give the training and inference strategy (algorithm) of SIDDM.

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

3 good

Limitations

The authors have discussed the societal impact in the supplementary material. It would be better if the authors discuss some limitations.

Reviewer 3HLp5/10 · confidence 3/52023-07-06

Summary

This method propose a way to achieve fast sampling during inference without compromising sample diversity and quality of diffusion model, and shows their effectiveness on both conditional and unconditional generation, qualitatively and quantitatively. A theoretical framework is also proposed which looks reasonable to me.

Strengths

This method basically proves that a very low FID can be obtained with very few sampling steps, compared to ADM and DDGAN. In addition, it claims that it can do both conditional and unconditional generation. Moreover, the theoretical framework looks solid to me, though I am not from the theory field so I cannot comment more on this.

Weaknesses

However, it seems that there are no experiments on conditional generation (both qualitative and quantitative). Perhaps more relevant experiments are welcome to consolidate this submission and prove the effectiveness of the new sampling strategy.

Questions

The overall story looks good to me. I wonder whether you would like to perfect your experiment with more qualitative/quantitative in the conditional setting.

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

3 good

Limitations

As mentioned above, more analysis on the conditional setting is welcome, e.g., using Stable Diffusion as the base model.

Reviewer SF8Z5/10 · confidence 4/52023-07-12

Summary

In this paper, the authors propose to use an implicit model to match the marginal distributions of noisy data and the explicit conditional distribution of the forward diffusion. Specifically, the adversarial loss is applied to the marginal distributions obtained by forward and backward process, and KL loss is used to regularize the different between the conditional distribution. Experiments show that the method works with small inference steps.

Strengths

* Using the adversarial loss is a good idea to learn the conditional distribution of the backward process with large step size. * The experimental results with small steps seem to be good, both for toy examples and real datasets.

Weaknesses

* Generally, the paper is hard to follow, the formulas should be presented in a better way: * How to obtain equation (6) through (5)? * Line 129-137 is really difficult to follow, since $q(x_t|x_{t-1})$ is Gaussian distribution, no matter the step size, why not directly use a Gaussian distribution to parameterize $p_\theta(x_t|x_{t-1})$? * It problematic to treat equation (11) and (12) as equivalent, since $\psi$ is fixed in (11). * In the experiments, with smaller steps, the method works better for both the toy and real datasets. It is not reasonable. Any explanation about this phenomenon?

Questions

Please see weakness part.

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Please see weakness part.

Reviewer g6aX5/10 · confidence 2/52023-07-27

Summary

The paper presents a method called Semi-Implicit Denoising Diffusion Model (SIDDM) aimed to accelerate the sampling process, enhance scalability to large datasets, and improve model performance. It improves upon the DDGAN model by reformulating the denoising distribution of diffusion models with explicit and implicit training objectives, which leverages an implicit GAN objective for the marginal distribution and an L2 reconstruction loss for the conditional distribution. To improve generative quality, the authors adopt a U-net-like structure for the discriminator and a new regularization technique involving an auxiliary denoising task. Experiments are conduct on CIFAR-10, CelebA-HQ-256, and ImageNet to demonstrate it effectiveness.

Strengths

(1) The experiment and comparison were comprehensive, and the results were good. Ablation was comprehensive. (2) The writing has a clear structure (just not sure if there are any issues with the derivation). (3) The method performs quite well on large datasets like ImageNet, whereas the previous ones didn't do well.

Weaknesses

(1) There lacks a graph showing the FID-sampling speed tradeoff on ImageNet. (2) In Table 1, the 'SIDDMs w/o AFD (ours)' entry is quite close to DDGAN in terms of setting, but why is there such a big difference in performance?

Questions

Please see the weakness part.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

NA

Reviewer azcW6/10 · confidence 4/52023-07-27

Summary

The paper addresses the issue of making diffusion models faster (faster sampling) while still yielding high quality, diverse samples, from a large phenomenological space (scaling to large datasets). This is achieved by a novel semi-implicit denoising diffusion model (SIDDM). The idea extends the approach followed in DDGANs by reformulating the objective, and a novel decomposition into two components: (a) a pair of marginal distributions over the denoised data (at each step) with an implicit form to be aligned with the JSD via a learnt critic, and (b) a pair of conditional, forward diffusion components having explicit forms to be aligned using the KL divergence. This mixed objective allows better scaling than DDGANs. A novel regularizer is also introduced for the discriminator which allows a more granular distribution matching. A proof of concept validation is performed using a synthetic, Mixture of Gaussians dataset. Ablations demonstrate the benefit of both the innovations. SIDDM is benchmarked against the art on the CIFAR10, CelebA-HQ-256, and ImageNet datasets demonstrating performance competitive with DDGANs, better scaling to ImageNet, and much higher sampling speeds than classical DDPMs.

Strengths

**Relevance** The paper addresses an important problem – speeding up diffusion based generative image models while not compromising on the image quality, with good scalability to large, complex datasets. This should be of relevance to the community working in these areas but also of interest to a much wider audience. **Originality** The proposed decomposition is interesting and novel, balancing both a non-parametric, population-level statistical alignment of distributions and a direct (simple, parametric) objective which decomposes into a sample level objective over Auxiliary Forward Diffusion (AFD). It stands to reason that this can have a significant impact on the training of the diffusion reversal DNN. **Technical Quality** - The technical approach mostly appears sound. It creates a new, direct and stronger learning signal which augments the critic based learning objective. This can lead to better learning outcomes over more complex distributions representing larger datasets like ImageNet. In addition, regularizing the discriminator via the UnetGAN formulation seems reasonable too. - The mathematical formalism used in the paper, including the upper bound in Theorem 1, add to the strength of the contribution, making it principled, and beyond the heuristic choices made. I haven’t verified the derivations in the paper entirely but have followed the structure of the derivations and the proof and find them reasonable. **Experimental Validation** There are several positives in the choices made for evaluation: - The choice of using a synthetic dataset based on Mixture of Gaussians (MoGs) allows for validating the core idea. - While the results on CIFAR10 and CelebA-HQ-256 are comparable with DDGANs, the FID scores on ImageNet demonstrate a performance comparable to SOTA diffusion models while DDGANs performance seems to demonstrate a failure to scale to ImageNet. This validates the core motivation of the paper. - The ablation study, with and without the AFD term, validates the benefit of the decomposition of the training objective that forms the basis of SIDDMs; as well as the benefit due to the UNetGAN regularizer in (14) **Significance** The approach has the potential for significant impact over an important area of research once the approach has been reproduced and ‘hardened’.

Weaknesses

**Experimental Validation** (a) The MoG experiment (Section 5.1) demonstrates that SIDDM achieves good results with very small number of time steps, and has better stability than DDGANs. However, - It is not discussed why it is unable to converge to the simple MoG data distribution, when the number of time steps increases (similar to DDGAN). - Table 1 shows the FID score. These trends don’t tally well with the behavior shown in Figure 3. Perhaps using other metrics (JSD, LPIPS, etc.) may explain the results better? - The implications of the above on the learning on real datasets, and the scalability thereof on larger, complex datasets like ImageNet is unclear. Such tasks may require more time steps (it is not clear whether this is the case) in which case, convergence to the data distribution becomes an issue. (b) The results on CelebA-HQ-256 are typically used to show performance on high-resolution data. The images (depictions) of the generated samples, both in the paper and in the supplementary, are too small. Similarly, the ImageNet depictions are very small. (c) While NFE (number of function evaluations) and the wall-clock time are shared, no details of the hardware are shared. **Clarity** (l. 230) It is not quite clear why the material on training with real datasets like CIFAR10, Celeb-HQ etc. is in Section 5.1 and not Section 5.2. Kindly fix. **Reproducibility** Implementation details are missing. The authors also don’t mention that the code will be released. Given this, I suspect it may be hard to reproduce results. **Discussion of Limitations** No discussion of limitations in the paper.

Questions

Kindly address the following: - Explain the behavior on the MoG dataset using a larger suite of metrics (see Borji, CVIU 2019; and its 2021 update for example). - Explain the behavior (non-convergence to data distribution) when using a larger number of steps and the scalability implications. - Did you investigate the tradeoffs between a smaller step size with larger number of steps, and smaller number of steps with a larger step size? How does one decide this? Does the theoretical formulation lend itself to such a guidance? - Kindly share details of the hardware, implementation, and other details which will aid reproducibility. - Will the code be shared? - Discuss limitations.

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

2 fair

Contribution

3 good

Limitations

Authors don’t address limitations of the paper. I don’t think there are any direct negative societal implications. Other limitations and opportunities for improvement are addressed in my responses to previous questions.

Reviewer 6duC2023-08-11

After rebuttal

Thanks for the rebuttal. First of all, I need to apologize. Due to my mistake, I mistype the method name in **Summary** and **Question-2**. **I have revised my comment.** For Question-2, I actually hope the authors provide the training and inference strategy of this method (SIDDM). I apologize again for the misunderstanding and inconvenience caused to the authors, area chairs, and other reviewers. Back to the rebuttal, the authors provide an explanation for Question-1. It is better if the authors further provide the method's training and inference strategies (algorithms).

Authorsrebuttal2023-08-14

Training and inference strategy of this method (SIDDM)

To reviewer 6duC: Hi, No worries, we are glad to provide the training and inference strategy of our proposed method. Training: 1. Sampling $x_0\sim q(x_0)), t-1\sim \text{Uniform}({0, . . . , T-1}), \epsilon_{t-1}\sim N(0,1), \epsilon_t\sim N(0,1)$. 2. $x_{t-1}=\sqrt{\bar \alpha}x_0 + \sqrt{1 - \bar \alpha}\epsilon_{t-1}, x_t=\sqrt{1-\beta_t}x_{t-1} + \sqrt{\beta_t}\epsilon_t$. 3. $x_0'=G_\theta(x_t,t), x_{t-1}' \sim q(x_{t-1}|x_0',x_t)$, where $q(x_{t-1}|x_0',x_t)$ is the posterior sampling from DDPM,.$x_t'\sim q(x_t|x_{t-1}')$, where q(x_t|x_{t-1}') is the distribution of forward diffusion. 4. D step: $\nabla_{\phi,\psi} (-\log(D_\phi(x_{t-1},t-1))-\log(1-D_\phi(x_{t-1}',t-1))$ $+ ||C_\psi(x_{t-1}', t-1))-x_t'||_2) $ 5. G step: $\nabla_\theta (-\log(D_\phi(x_{t-1},t-1)) + ||x_t-x_t'||_2$ $-||C_\psi(x_{t-1}', t-1))-x_t'||_2)$ Repeat 1-5 until model converge. Our inference follows the same strategy as DDGAN. Inference: 1. $x_T\sim N(0, 1)$ 2. for t = T,...,1 do $\epsilon \sim N(0,1)$ if $t>1$ else $\epsilon=0$, $x_0'=G_\theta(x_t,t)$, $x_{t-1}' \sim q(x_{t-1}|x_0',x_t)$ end for return $x_0$. Let us know if you have more concerns on our training or inference strategies, we are more than happy to help you address them.

Reviewer g6aX2023-08-17

Thanks for the reply

NA

Authorsrebuttal2023-08-20

Thanks again for the reviewing!

Reviewer azcW2023-08-20

Post rebuttal

Thanks for the detailed response. I have no further questions. I have also gone through the remaining reviews and author responses. I will finalize the rating accordingly.

Reviewer SF8Z2023-08-21

Thanks for the rebuttal, the new version of the paper looks better and clearer and I'll change my rating to 5. However, I am still not convinced about why smaller steps get better results and think the authors can do more work on this part.

Authorsrebuttal2023-08-21

Hi, Thanks for the thoughtful suggestions, we will make this explanation better and we are trying to analyze the reason behind the influence of training steps in our method. Thanks!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC