Alleviating Distortion in Image Generation via Multi-Resolution Diffusion Models and Time-Dependent Layer Normalization

This paper presents innovative enhancements to diffusion models by integrating a novel multi-resolution network and time-dependent layer normalization. Diffusion models have gained prominence for their effectiveness in high-fidelity image generation. While conventional approaches rely on convolutional U-Net architectures, recent Transformer-based designs have demonstrated superior performance and scalability. However, Transformer architectures, which tokenize input data (via"patchification"), face a trade-off between visual fidelity and computational complexity due to the quadratic nature of self-attention operations concerning token length. While larger patch sizes enable attention computation efficiency, they struggle to capture fine-grained visual details, leading to image distortions. To address this challenge, we propose augmenting the Diffusion model with the Multi-Resolution network (DiMR), a framework that refines features across multiple resolutions, progressively enhancing detail from low to high resolution. Additionally, we introduce Time-Dependent Layer Normalization (TD-LN), a parameter-efficient approach that incorporates time-dependent parameters into layer normalization to inject time information and achieve superior performance. Our method's efficacy is demonstrated on the class-conditional ImageNet generation benchmark, where DiMR-XL variants outperform prior diffusion models, setting new state-of-the-art FID scores of 1.70 on ImageNet 256 x 256 and 2.89 on ImageNet 512 x 512. Project page: https://qihao067.github.io/projects/DiMR

Paper

References (67)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Wwps7/10 · confidence 5/52024-06-27

Summary

The paper proposes a multi-resolution network for diffusion model which emphasizes the learning features across resolutions. The method further introduces a time-dependent layer norm to boost the model performance with fewer parameters compared to AdaLN in DiT. This proposed network demonstrates a SoTA performance on ImageNet at 1.70 and 2.89 on ImageNet 256 & 512, respectively.

Strengths

- The paper addresses image distortion caused by varying patch sizes by integrating a multi-resolution structure into the network. The choice of patch sizes involves a trade-off between computational complexity and model performance: larger patch sizes reduce computational complexity but degrade performance. - The method finds a balance between compute complexity and performance by introducing a multi-branch architecture and multi-resolution loss to decompose the learning process from low to high resolution. The idea is a substantial contribution. Through quantitative and qualitative results, the method can alleviate distortion of image generation and achieve SoTA FID scores. - The time-dependent layer norm is a simplified form of AdaLN with less intensive parameters by removing the cumbersome MLP layer and rearranging class embedding and time embedding for parameter-efficient time conditioning.

Weaknesses

- The discussion between the method and cascaded methods like "Cascaded diffusion models" and "Matryoshka diffusion models" should be included to highlight the advantages of the method. - In table 1, it needs to be updated with some latest SoTa methods like MDTv2 with the best FID of 1.58 and PaGoDA with the best FID of 1.56 on ImageNet 256x256. Not to compete with them, but to get an overview picture of the latest methods. - The proposed multi-scale diffusion loss is also introduced in SimpleDiffusion. So, the authors need to mention the difference in the paper. - Equation 2: have the authors used concat instead of adding the upsampling features directly to the larger-resolution features of the next branch? - The current method only uses the output features of the previous branch and injects them into the start of the next branch. However, this design lacks interconnections between blocks from the low-resolution and high-resolution branches. Ideally, there should be skip-connections across branches. What is the motivation behind this? If authors have not tested this, it is encouraged to do so. - In table 2, what is the reason why using AdaLN-Zero with Multi-branch (row 3) causes a bad result? - The root problem of distortion boils down to the use of patch size which is the main motivation of the work. However, in the design of multiscale network, the authors also patchify the input image to the corresponding resolution for each branch. I wonder why the author do not use different-resolution noisy images as the input to each branch like Matryoshka Diffusion Models. I think this is more effective to completely remove distortion problem. As shown in Figure 5, the method still exhibits a certain level of distortion rates. - Sampling speed: it is valid to include a comparison of sampling time with baselines. - Figure 3: What is the red line meaning? Misc: L37, L130: the figure 7 of DiT paper is incorrectly linked. Ref: - Gao, Shanghua, et al. "MDTv2: Masked Diffusion Transformer is a Strong Image Synthesizer." arXiv preprint arXiv:2303.14389 (2023). - Kim, Dongjun, et al. "PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher." arXiv preprint arXiv:2405.14822 (2024). - Hoogeboom, Emiel, Jonathan Heek, and Tim Salimans. "simple diffusion: End-to-end diffusion for high resolution images." International Conference on Machine Learning. PMLR, 2023.

Questions

N/A

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

As pointed out in Weaknesses, the method still has a certain distortion rate which is due to the use of strided convolution (a form of patchification) at the start of each branch. I think the author should include this in the revised manuscript.

Authorsrebuttal2024-08-07

Part2 of our rebuttal to Reviewer Wwps: addressing Weakness 7-10 and Limitation 1

> W7: I wonder why the author do not use different-resolution noisy images as the input to each branch like Matryoshka Diffusion Models. I think this is more effective to completely remove distortion problems. We thank the reviewer for the suggestion. In DiMR, we have one specific branch that uses the original resolution without patchification, which is designed to capture the finer visual details. Using different-resolution noisy images is a very good idea, but we think that resizing the input images to a smaller resolution may also cause the loss of details. Additionally, the FID score of 6.62 reported by Matryoshka Diffusion Models (vs. DiMR 1.70) may also suggest that this would only provide marginal improvements. Nevertheless, we agree that designing a method without any sort of patchification would further reduce the distortion rate. We leave it for future work.   > W8: Sampling speed: it is valid to include a comparison of sampling time with baselines. Thanks for the suggestion! We provide the comparison of sampling time in the following table. We test both methods with batch size 1, and compare the sampling time to generate one image on an A100. We can see that our model also has a very high sampling speed. ***Table L**: Comparison of sampling speed. We report Sampling Speed (second/sample) here.* | | ImageNet-256 | ImageNet-512 | |----------------|--------------|--------------| | DiT-XL | 3.3 | 6.3 | | DiMR-XL (Ours) | 2.4 | 2.6 |   > W9: Figure 3: What is the red line meaning? Thanks for pointing it out. It means 0.1 threshold. We will make it clear in the revision.   > W10: L37, L130: the figure 7 of DiT paper is incorrectly linked. Thank you. We actually did intend to refer to the Figure 7 of DiT paper, which qualitatively shows that decreasing patch size reduces distortion. We note that they are not hyper-linked.   > L1: The method still has a certain distortion rate which is due to the use of strided convolution (a form of patchification) at the start of each branch. I think the author should include this in the revised manuscript. Thanks for the suggestion. We will include this in the limitation section. It is noteworthy that completely solving distortion in image/video generation is an extremely challenging problem. In this work, we have already significantly reduced the distortion rate compared to U-ViT and DiT, which is quite a step towards the goal.

Reviewer Wwps2024-08-08

Response

I appreciate the authors' effort in addressing all of my concerns and I am happy with these answers. I will raise my final score and vote for acceptance, as the contributions are significant for advancing the growth of diffusion models.

Reviewer RADT5/10 · confidence 4/52024-07-03

Summary

This paper proposes to replace the original pull transformer blocks in DiT with multi-resolution network. Detailly, transformer block is employed for low resolution and Conv blocks are used for the remaining higher resolution. An additional time-conditioning block is designed for Conv blocks. The effectiveness of this method is validated with numercial experiments.

Strengths

1. This paper proposes a multi-resolution network, consisting of both transformer and conv blocks, to boost the performamce of image generation. Such design is rational. To fit the conv block, this paper also introduce an additional time conditioning block. 2. This method achieves better performance than DiT.

Weaknesses

1. It is unclear what the performance gain comes from. intuitively, the multi-resolution design could already brings better performance. But this paper proposes the conbination of transformer and conv. I wonder if the design of transformer in low resolution really necessary. 2. The scalibility of this network may be limited due to the introduction of conv blocks. For example, for network parameters larger than 1B. 3. This paper mainly claims to solve the large computional complexityinduced by the long token length. But in the real long token length case (image 512x512), the perfomance gains of this method is marginal compared to DiT. This raises the concern of the real performance of this method on large image case, eg. >512x512.

Questions

1. It is suggested to identify the real core contribution for performance gain of this paper. The ablation study of replacing the transformer block with conv block is suggested. 2. The effectiveness of this method on larger image size and networks are concerened. The authors are suggested to give furhter discussions on these points if they wish a full surpass of the original DiT.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors mask well discussions on the limitaion of this paper.

Reviewer MPTv5/10 · confidence 3/52024-07-12

Summary

This paper introduces a diffusion model named DiMR, which incorporates a Multi-Resolution Network and Time-Dependent Layer Normalization to enhance image generation quality. Traditional diffusion models, often limited by a trade-off between computational efficiency and visual fidelity, struggle with image distortion due to the coarse resolution of input data processing in Transformer-based designs. DiMR addresses this by refining image details progressively across multiple resolutions, reducing distortions significantly. The new light-weight Time-Dependent Layer Normalization technique introduced in the model embeds time-based parameters into the normalization process efficiently. Demonstrated through benchmarks on ImageNet 256x256 and ImageNet 512x512, DiMR outperforms competing models.

Strengths

+ The paper is well-written and easy to follow, featuring clear diagrams and detailed captions that enhance understanding. + The introduction of Time-Dependent Layer Normalization (TD-LN) is particularly interesting. The use of PCA analysis to justify this approach provides a strong motivation and highlights its innovativeness. + The experimental section of the paper effectively demonstrates the effectiveness of the proposed method.

Weaknesses

- The model has a higher FLOPs count compared to a DiT of similar parameter size, raising concerns about slower training speeds. - It is unclear whether this cascaded approach can support multi-resolution training.

Questions

- This work appears to be a cascade diffusion model based on DiT. How does it perform compared to traditional UNet-based cascade diffusion models? - What are the training and convergence speeds of this model, particularly in comparison to the baselines? - The proposed method adopted the class token similar to that used in U-ViT for injecting class conditions. Is it possible to replace this with an approach similar to Time-Dependent Layer Normalization (TD-LN)?

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors have adequately addressed the limitations and potential negative societal impact of their work.

Reviewer 1gRs5/10 · confidence 3/52024-07-13

Summary

This paper works on efficient diffusion model backbone architecture by using transformer and ConvNeXt architecture respectively on small and large resolutions of the same inputs, to leverage the strength of both architectures, and alleviate the distortion problem.

Strengths

1. The writing is easy to read and follow. 2. The idea is simple and effective. 3. They conduct rich experiments to support the idea.

Weaknesses

1. This work utilizes two standard architectures, incorporating only minor design elements such as the TD-LN. It would be more advantageous to propose a new, general, and elegant architecture that effectively combines the strengths of the attention layer and the convolution layer.

Questions

Are there any ablation studies on the number of different resolutions used? For instance, what happens if we use one small resolution with DiT and one normal resolution with a larger conv net?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

See the weakness section.

Reviewer RADT2024-08-12

Thanks for the response. I decide to raise the score to 5.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC