Diffusion Tuning: Transferring Diffusion Models via Chain of Forgetting

Diffusion models have significantly advanced the field of generative modeling. However, training a diffusion model is computationally expensive, creating a pressing need to adapt off-the-shelf diffusion models for downstream generation tasks. Current fine-tuning methods focus on parameter-efficient transfer learning but overlook the fundamental transfer characteristics of diffusion models. In this paper, we investigate the transferability of diffusion models and observe a monotonous chain of forgetting trend of transferability along the reverse process. Based on this observation and novel theoretical insights, we present Diff-Tuning, a frustratingly simple transfer approach that leverages the chain of forgetting tendency. Diff-Tuning encourages the fine-tuned model to retain the pre-trained knowledge at the end of the denoising chain close to the generated data while discarding the other noise side. We conduct comprehensive experiments to evaluate Diff-Tuning, including the transfer of pre-trained Diffusion Transformer models to eight downstream generations and the adaptation of Stable Diffusion to five control conditions with ControlNet. Diff-Tuning achieves a 26% improvement over standard fine-tuning and enhances the convergence speed of ControlNet by 24%. Notably, parameter-efficient transfer learning techniques for diffusion models can also benefit from Diff-Tuning.

Paper

References (58)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer x2jk5/10 · confidence 4/52024-06-20

Summary

This paper proposes Diff-Tuning method to encourage the fine-tuned model to retain the pre-trained knowledge. In this method, both pre-trained data and downstream data are used to train the diffusion model. Compared to standard fine-tuning methods, Diff-Tuning enhances the convergence speed and improves the performance. Diff-Tuning can also be used in Conditional Generation. Additionally, Diff-Tuning is compatible with existing parameter-efficient fine-tuning methods.

Strengths

+ The idea of using the pre-trained model to act as a universal denoiser for lightly corrupted data is interesting. + Diff-tuning achieved faster training and better performance than standard fine-tuning. + This paper provides novel theoretical insights to reveal the principle behind the chain of forgetting. Therefore, this paper is solid.

Weaknesses

+ The novelty of this paper is limited. Many previous works [1,2] find that gradients exist in conflict for diffusion models across timesteps even in the same dataset. But none of them are compared. **Since the core conclusions of this paper are very similar to them, I think the author needs to highlight the differences.** 1. Efficient Diffusion Training via Min-SNR Weighting Strategy ICCV 2023 2. Addressing Negative Transfer in Diffusion Models. NeurIPS 2023 + The robustness of Diff-Tuning to sampling algorithms has not been verified, including DDIM [1], DPM-Solver [2], and others. Meanwhile, More diffusion models need to be validated, including VP-SDE [3], Flow Matching [4], EDM [5], etc. 1. Denoising Diffusion Implicit Models ICLR 2021 2. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps Neurips 2022 O 3. Score-Based Generative Modeling through Stochastic Differential Equations ICLR2021 4. Flow Matching for Generative Modeling ICLR 2023 5. Elucidating the Design Space of Diffusion-Based Generative Models NeurIPS 2022 + In the training of the diffusion model, there are different time schedules (Linear, Cosine, etc.), so do I need to keep the same schedule in the Diff-Tuning?

Questions

Please see Weakness. If my question is answered, I will raise my score.

Rating

5

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Please see Weakness. If my question is answered, I will raise my score.

Reviewer y5ux6/10 · confidence 4/52024-07-14

Summary

This paper explores the fundamental transfer characteristics of diffusion models and observes the monotonous chain of forgetting trend of transferability of diffusion models in the reverse process. It then proposes a simple but effective transfer approach to make the fine-tuned model retain the denoising ability of the pre-trained model close to the generated image, while using domain-specific denoising ability at the beginning of the denoising process. Experimentally, Diff-Tuning can achieve effective improvement on some standard fine-tuning tasks, and also improve the convergence speed of ControlNet.

Strengths

1. Diff-Tuning is simple and easy to follow, and experimentally well realized. 2. the motivation is clear and it explores the transfer capability of the most popular diffusion model, which is novel. 3. this article should make me deepen my understanding of the diffusion model and notice its interesting phenomena in the reverse process.

Weaknesses

1. There are too few metrics on class-conditional generation, while FID is used to extract features with models trained on top of ImageNet, authors use DiT pre-trained on ImageNet to generate samples, which are then added to the training process, it is difficult to say the lower FID actually translates to improved image quality. can you use more metric to show your results?

Questions

1. the authors use a pre-trained diffusion model to preserve denoising ability, would it be useful to use pre-trained domains, i.e. images from ImageNet, instead of augmented images? Although I know that in most cases we can't fully know the pre-trained diffusion model's pre-training data 2. when using a pre-trained diffusion model to generate augmented data, do different sampling methods achieve similar results, and does the choice of sampling steps make a difference? 3. would using a pre-trained diffusion model on ImageNet and then fine-tuning it on small data achieve better results than a diffusion model trained directly on small data? For example, pre-training on ImageNet and then fine-tuning on CIFAR10/100.

Rating

6

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

yes

Reviewer WW5v5/10 · confidence 4/52024-07-17

Summary

The paper proposes a new method to fine-tune the pretrained large-scale diffusion models for new tasks. It finds that different time steps of the denoising process of diffusion models have varying transferability. Specifically, the paper finds that "low-noise" time steps close to the end of the denoising process have better transferability. In contrast, the "high-noise" steps responsible for generating the image layout and containing domain knowledge are less transferable. Based on these observations, the paper proposes a novel fine-tuning objective demonstrating better convergence speed than baselines.

Strengths

- The idea is simple without numerous hyperparameters or tricks to work. - The proposed idea is orthogonal to the main focus of the recent papers that aim to do parameter-efficient fine-tuning, and it can be combined with them as the paper demonstrates in the experiments.

Weaknesses

1. I could not follow the motivation behind the scheme that the paper claims to perform Knowledge Retention with. The main argument of the paper is that "the time steps of the denoising process close to the end of the sampling chain of diffusion models are more transferable." Then, why do we need to sample an augmented dataset to do the Knowledge Retention? A straightforward way can be using knowledge distillation from the original model to the fine-tuned one in the low noise time-steps. 2. Following the previous point, distillation may be better than an augmented dataset considering that the fine-tuning data may have characteristics that are different from the original domain. For instance, (this is just a hypothetical example) the fine-tuning dataset may be on flowers that are mostly red, but the pretaining dataset (ImageNet in the experiments) may not have red flowers. Therefore, for the fine-tuning process, we care more about red flowers, not other types. However, using an augmented dataset can bias the fine-tuning process and lower the model's quality on red flowers. In contrast, distillation on the low-noise time steps may readily preserve the prior knowledge of the model without biasing the fine-tuning process. 3. I appreciate the analysis in Fig. 1, and the formal experimentation of it is a novelty. Yet, some similar phenomena have been shown in the literature before that the paper does not discuss them. [1] shows that a diffusion model usually generates a high-level layout of the image until SNR gets between [0.01, 1], and then, starts to fill in the details. Similarly, eDiff-I shows that changing the text prompt at the low-noise stages of the denoising process process does not change the output image. These experiments in Fig. 1 also show that if we use the original model in the low-noise time steps, the performance does not decrease, which is similar to these papers. I think the paper should include these references in the related work section. [1] Perception Prioritized Training of Diffusion Models, CVPR 2022. [2] eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers, 2023.

Questions

I have the following questions from the paper: 1. When fine-tuning the ImageNet pretrained model on the new task, do you train the class embeddings (used for cross-attention) for the new task as well? 2. Regarding points 1 and 2 in the Weaknesses section, as far as I know, training on images generated by the generative models may degrade the model's performance [1]. Yet, using the whole ImageNet dataset performs worse than using 5k images that are sampled from the pretrained model. Is there any intuition about why this happens? [1] Self-consuming generative models go mad, ICLR 2024.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have pointed to using an augmented dataset as a potential limitation, but as I mentioned in the weakness section, the motivation and intuition of why it should work is not clear.

Reviewer a4Zm6/10 · confidence 3/52024-07-21

Summary

This paper focuses on transfer learning methods for diffusion models. It experimentally demonstrates and provides theoretical insights into how the forgetting trend varies with the diffusion timestep. Based on this observation, they proposes Diff-Tuning. The proposed method introduces objectives for knowledge retention and reconsolidation, ensuring that they are affected decreasingly and increasingly by the diffusion timestep, respectively. The superiority of the proposed method is demonstrated experimentally.

Strengths

The paper is well-written, and the method is well-motivated both experimentally and theoretically.

Weaknesses

* More discussion on how to set the hyperparameters $\xi(t)$ and $\psi(t)$ would be helpful. For example, these parameters perform an weighted sum of the retention and reconsolidation objectives, and combining this with a maximum likelihood analysis could potentially help find hyperparameter selection. * I am curious if applying the datasets used for pre-training to retention would be better (or worse) than using a pre-sampled dataset and the reason. * It would be beneficial to discuss how the method can be extended in scenarios where transfer learning is applied multiple times or where the data distribution changes online.

Questions

Please see the Weaknesses part.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

They provided in the Supplementary section D.

Reviewer x2jk2024-08-08

I thank the author for his reply, which resolved my main confusion. Consequently, I have raised my score.

Reviewer WW5v2024-08-13

Official Comment by Reviewer WW5v

I appreciate the authors' rebuttal. The rebuttal addressed most of my concerns, and I raised my score to 5. I did not use a higher score as I think the design choice of using an augmented dataset is not well-motivated. The rebuttal did not provide a compelling answer to my second comment in the weakness section of my initial review. Still, I raised my score as the framework is flexible such that one can replace the augmented dataset with distillation, as the provided experimental results in the rebuttal suggest.

Reviewer a4Zm2024-08-13

I appreciate the author's response. Most of my concerns have been addressed, and I raise the score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC