Slight Corruption in Pre-training Data Makes Better Diffusion Models

Diffusion models (DMs) have shown remarkable capabilities in generating realistic high-quality images, audios, and videos. They benefit significantly from extensive pre-training on large-scale datasets, including web-crawled data with paired data and conditions, such as image-text and image-class pairs. Despite rigorous filtering, these pre-training datasets often inevitably contain corrupted pairs where conditions do not accurately describe the data. This paper presents the first comprehensive study on the impact of such corruption in pre-training data of DMs. We synthetically corrupt ImageNet-1K and CC3M to pre-train and evaluate over 50 conditional DMs. Our empirical findings reveal that various types of slight corruption in pre-training can significantly enhance the quality, diversity, and fidelity of the generated images across different DMs, both during pre-training and downstream adaptation stages. Theoretically, we consider a Gaussian mixture model and prove that slight corruption in the condition leads to higher entropy and a reduced 2-Wasserstein distance to the ground truth of the data distribution generated by the corruptly trained DMs. Inspired by our analysis, we propose a simple method to improve the training of DMs on practical datasets by adding condition embedding perturbations (CEP). CEP significantly improves the performance of various DMs in both pre-training and downstream tasks. We hope that our study provides new insights into understanding the data and pre-training processes of DMs and all models are released at https://huggingface.co/DiffusionNoise.

Paper

Similar papers

Peer review

Reviewer XQzf7/10 · confidence 4/52024-07-08

Summary

The authors find that slight corruption in the conditioning of text-to-image diffusion models improves the performance. The authors theoretically analyze this empirical finding in a toy model where the goal is to learn to sample from a Gaussian Mixture and the network is piece-wise linear function. Inspired by their experimental and theoretical findings, the authors propose a technique to improve the performance of text-conditional diffusion models by adding noise to the embedding of the conditioning vector.

Strengths

* The research topic is timely. State-of-the-art diffusion models are trained on billions of image-text pairs. Many of those are potentially noisy. * The authors propose a very simple modification to the training of text-to-image diffusion models that leads to improved performance. * The finding that slight corruption in the conditioning helps is interesting. * The authors validate their findings through an extensive experimental evaluation.

Weaknesses

* I believe that the authors should emphasize in the title of their paper and in the main body that they study corruption in the conditioning. Numerous works study corruption in the images themselves, e.g. see: 1) Ambient Diffusion: Learning Clean Distributions from Corrupted Data 2) Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data 3) GSURE-Based Diffusion Model Training with Corrupted Data 4) Solving inverse problems with score-based generative priors learned from noisy data In these works, it has been observed that data corruption decreases the performance of diffusion models. I believe the authors should acknowledge these relevant works and clarify that their paper analyzes corruption in the conditioning. * The theoretical model is not very relevant, at least not very relevant to Section 3 of the paper. In the experiments of Section 3, the authors mislabeled some of the training examples. In the theoretical model, it is assumed that the true class is given to the model since the model is using the parameters that correspond to the true class. In a sense, the theoretical model of Section 3 is more related to the proposed algorithm in Section 5. Also, the existence of multiple centers in Section 4 does not seem relevant, unless I am missing something. This is because the model has separate parameters for each center and the centers are given for each example the model sees. I went over the proof and it looks like this is exploited in the proof as well, to obtain a closed-form solution for the optimal network. * For the reasons mentioned above, it seems to me that the reason this works is some sort of regularization. I wonder if a similar effect could be obtained by increased dropout in the conditioning or some other form of regularization. Overall, I believe that the reader of the paper doesn't develop much intuition about the origins of the experimental finding.

Questions

Apart from the weaknesses mentioned above, I have one more important question regarding this paper. Could the authors please clarify whether the corruption in Sections 3 and Sections 5 happens once for each data sample before the training or whether a new corruption takes place whenever we encounter the same data point across different epochs? To clarify, I want to understand which one of the two is happening: 1) We take a dataset, we corrupt it once, we train with the corrupted dataset 2) We use the clean dataset and every time we see a sample we corrupt it (by adding a different noise each time to its embedding or by changing each label/text to something else each time we see the same example across epochs). If 2) is happening, it seems to me that the proposed method is closely related to regularization. I think this needs to be clarified and the effect of doing 1) or 2) should be studied as an ablation.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

The authors have adequately discussed the limitations of their work.

Reviewer 45qi7/10 · confidence 3/52024-07-12

Summary

The paper investigates the impact of slight corruption of conditioning information in pre-training data on the performance of diffusion models (DMs). By introducing synthetic corruption to ImageNet-1K and CC3M datasets, the study evaluates over 50 conditional DMs. Empirical and theoretical analyses reveal that slight corruption enhances the quality, diversity, and fidelity of generated images. Based on these insights, the work proposes Conditional Embedding Perturbations (CEP) as a method to improve DM training which shows significant improvements in both pre-training and downstream tasks.

Strengths

- **Clarity, very well written:** The paper is clearly structured and builds up the motivation for the proposed method step by step from experiments and theoretical analysis. - **Extensive experiments:** The study examines many conditions training multiple diffusion models with different noise conditions from scratch. This gives a detailed perspective on the performance benefits of using slight conditioning corruption during training and is a valuable addition to the field. These experiments are evaluated both qualitatively and quantitatively. - **Unexpected insight:** The study well describes a generally unexpected phenomenon and shows how it can be leveraged to develop an new method.

Weaknesses

- **Notion of significance:** The paper claims that slight pre-training corruption yield significantly better performance in terms of FID and IS metrics. Nevertheless, this notion of significance is never formally tested. For this at least a experiment with multiple repetitions should be performed to determine statistical significance. - **Typos:**, Page 8, lines 257 and 259 rrecision -> precision, Page 9, lines 299 dataset -> data - **Link between theoretical analysis and metrics** should be made explicit. I.e. pointing out that the FID is essentially a Gaussian approximation based estimated for the 2-Wasserstein distance of inception features (preferably at page 7, line 223).

Questions

None

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

4

Limitations

The authors correctly reflect the limitations of the study. In particular, the theoretical analysis is limited due to the assumptions made, but can still shed some light on the underlying mechanism. Further, the general issue of evaluating image generative models quantitatively is pointed out and a limitation of the work.

Reviewer 6StB6/10 · confidence 4/52024-07-16

Summary

In this paper, the authors study the effect of slight corruptions to the training data during pretraining of conditioned diffusion models. They introduce the perturbations to the "condition" in the diffusion models and show that this leads to a better model (via FID and other metrics). They also provide an intuition on why corruptions help by theoretically studying this case on a GMM.

Strengths

- Well-written paper - All the experiments are well thought out, especially the downstream application experiments. - Multiple metrics are evaluated and shown that perturbation helps, which makes the claim more sound.

Weaknesses

- While authors showed extensively that adding noise can help, they missed to show "when does it help"? For example, if I'm a practitioner who wants to train a large model from scratch, I usually will not have the budget to train multiple models to determine the amount of noise that needs to be added. That insight is missing in this paper.

Questions

- Does the optimal noise perturbation level transfer between the models? If 10% is the best noise level for LDM on one dataset, would it be the best noise for the same-size LDM on another? - How do metrics look like throughout training for models trained with different levels of noise? - CC3M corruption details are not clear. I went to Appendix B1, you mentioned 5 levels of corruption, but it is still vague! What are these 5 levels? - How is the number of training iterations determined for the models? I looked at the appendix and made some rough calculations, it seems like you trained the IN model for 91 epochs and CC3M for 67 epochs (Pls correct me if I'm wrong). On what basis are these numbers decided? This again ties back to the question, when is "adding noise" helpful? Are we in an overtraining regime? Is that the reason why adding corruption acts as "augmentation" and hence the metrics improve? - (Minor) A few relevant citations are missing. For example, adding perturbations to the conditioning both in text space and latent space is explored in the context of memorization in last year's Neurips paper [1]. [1] - "Understanding and mitigating copying in diffusion models." Advances in Neural Information Processing Systems 36 (2023): 47783-47803.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

n/a

Reviewer mxHS6/10 · confidence 4/52024-07-21

Summary

This paper claims that slight corruption of the condition is beneficial for conditional diffusion models. Experimentally, they conducted experiments with label flips or text swaps in the dataset and observed the performance improvement over zero noise at noise levels of 2.5% to 5%. Theoretically, they used a Gaussian mixture model to show that generation diversity and quality, as measured by entropy and Wasserstein distance, can be improved with a small amount of state corruption. Based on this, they proposed a method for injecting small amounts of Gaussian noise into the condition embeddings and experimentally demonstrated its superiority.

Strengths

Their claim offers a new perspective that may not be intuitive, but has been partially mentioned in the existing literature. They provide a solid experimental and theoretical explanation for this view.

Weaknesses

* From my understanding, in Theorem 2, $\gamma$ is $O(1/\sqrt{n})$, which is the scenario for small corruption. It would be beneficial to address this explicitly in the main text. * It would be beneficial if the method that determines the noise level could be linked to the theory like using the $\gamma=O(1/\sqrt{n})$,. * They briefly mentioned the various assumptions in their theory as limitations. It would be helpful to provide a detailed explanation of these assumptions and the scenarios in which they hold.

Questions

Please see the Weaknesses part.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

They mentioned in the last section.

Reviewer XQzf2024-08-09

Increased my rating

Thanks for your rebuttal. Most of my concerns are now addressed. Please make the appropriate changes in the camera-ready version of your work, as promised. I increased my rating to 7.

Authorsrebuttal2024-08-09

We thank the reviewer's feedback and efforts reviewing this work again. The changes and additional ablation studies will be included our revised work.

Reviewer 45qi2024-08-12

Thank you for the comments and additional points. While they do clarify some of my concerns, I think my previous score is still appropriate.

Reviewer 6StB2024-08-13

Thank you for the response

While I still disagree with the justification for the paper's insights, I appreciate the authors' effort in addressing most of my other questions. I will increase my score.

Authorsrebuttal2024-08-13

Thank you for your feedback and for acknowledging our efforts to address your questions. We appreciate for the increased score.

Reviewer mxHS2024-08-13

Thank you for the author's response. After checking the author's response and other reviews, most of my concerns have been addressed, so I raise the score.

Authorsrebuttal2024-08-13

Thank you for your feedback and for revisiting our paper. We're glad our response addressed your concerns, and we appreciate the updated score.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC