Anti-Exposure Bias in Diffusion Models

1. while the authors use a lot of analysis to show that the gap between the true Gaussian kernel and its modeling counterpart does exist, there is a lack of visualization to further illustrate this point.

2. It is suggested to move Table 9 in the appendix to the main paper, this is important to illustrate the efficiency of the algorithm.

3. The authors need to conduct further analysis of the effect of the number of parameters of the Anti-Bias Prompt Model on inference performance.

4. The authors do not provide code, which is important for the reproduction of the algorithm, even if it is simple code on CIFAR-10.

1. More visualization comparision should be provided.
2. Based on the visual results, the improvement of the proposed method is not very significant.

Overall, from my perspective, I think this paper is valuable and should be accepted. However, there still remains some weaknesses. If the authors can respond my puzzles, I will raise my scores.

For presentation:

1. The title of this paper ``$**Anti-Exposure Bias in Diffusion Models via Prompt Learning**$''. Through the whole paper, I think it is more like noise learning instead of prompt learning? What is the relationship between the proposed methods and prompt learning?

2. This paper only provides the improved images in Fig.7. Can the authors provide more images for initutive comparsions?

For method:

1. In eqn. 7, the authors claim $\Phi(\hat{x}_t,x_t)$ represents the exposure bias. However, based on my understanding, I only find the $**discretization error**$, rather than both $**discretization error**$ and $**score estimation error**$.

2. This paper utilizes VE SDE to get $x_{t+1}$, in order to simulate exposure bias. However, from my perspective, I think the $\hat{x}_t$ from prompt prediction model is not the Gaussian noise. This approach can be considered as a way to mitigate the training-sampling discrepancy of diffusion model such as [1]. Any comparsions between these methods?

[1] Common Diffusion Noise Schedules and Sample Steps are Flawed

For experiments:

1. All of these experiments are conducted in traditional benchmark, such as CIFAR, CelebA, and ImageNet 256x256. Moreover, the DMs used in this paper are not the popular T2I diffusion models, such as SDXL. Although the authors leave this as the future work, I think without these experiments, the practial value of this paper is limited.

2. Moreover, in this paper, the authors state the batch size of all experiments is $**1024**$, and the iteration of EDM in Tab. 8 requires 150k. That means approximate 100M data samples. The $**prompt prediction model**$ is lightweight and the parameters of the DMs are fixed. Why such a huge amount of data samples are needed?

3. Based on 2, the size of image in all experiments are not bigger than 256x256. How about generating the images with 1024x1024, should it require more data samples and computional resources?

1.The architecture of the prompt learning network is unclear. It is recommended that the authors include the proposed module architecture or network detail description in the supplementary materials to aid readers in understanding the method.

2.While the experiments indicate some improvement from the proposed module, the selection of baseline models is relatively simple. Will the issue of exposure bias become less significant in practical applications after updating the model architecture and expanding the training data scale?

3.The improvements depicted in Figures 4, 5, and 6 (in the supplementary materials) do not appear to be particularly pronounced.

4.There are missing details regarding the training data used for the prompt learning framework.

5.The paper lacks references or calculation methods for NEFs.

While the experiments focus on improving image quality, is there a possibility of trade-offs such as loss of diversity that could be caused by applying the prompt adjustment? I think additional experiments that take into account possible trade-offs could further strengthen the method.