EXPLORING FEW-SHOT IMAGE GENERATION WITH MINIMIZED RISK OF OVERFITTING

1. The experiments are mainly about face, which is a not very challenging task. Although a few results are provided in the supplementary, more expeirmental results about more scenes and categories should be provided.
2. There is no efficiency comparision with baselines.
3. The discussion on the related works of few-shot image generation is not comprehensive.
4. The authors only compare with three few-shot image generation baselines, which is far from enough. More recent methods should be compared.
5. There is no discussion on the failure cases.

1. My main concern is that the method proposed in this paper is highly sensitive to hyperparameters, specifically \gamma. From Figure 3, it can be observed that as $\gamma$ increases, the FID first decreases and then increases, and the optimal value varies across different datasets. I suspect that even within the same domain, the optimal value may differ between datasets, which would significantly limit the applicability of this method.
2. This paper does not discuss how the sharing mechanism works when \gamma is less than the number of images. This is worth exploring.
3. When \gamma is not large enough, the model underfits, which may be related to the number of learnable parameters or the capacity of the model. This part could explore other ways to alleviate the issue of underfitting.

1. The method's reliance on both IGM and sharing degree might complicate the implementation process. Users may face challenges in tuning these additional parameters.
2. The selection of the sharing degree is crucial, yet the paper does not provide comprehensive guidelines on how to choose this parameter effectively.
3. Although the method shows improvements in representation learning, the computational cost associated with training and tuning, especially with higher resolutions and batch sizes, may limit its practical application in resource-constrained environments.

(-) The concept of overfitting in this paper seems confusing. The term "overfitting" is commonly associated with discriminative tasks rather than generative tasks. Specifically, in the training GANs with limited data, existing approaches only demonstrate the overfitting of the discriminator (D) issue, with no mention of the overfitting in the generator (G). Instead, the generator usually suffers from the gradient vanishing or instability problem. Furthermore, in diffusion models, denoising score matching is used to update the parameters of the diffusion model, which indicates that the diffusion model also suffers from the gradient issue rather than the so-called overfitting problem. The effectiveness of the proposed gradient clipping in Line 94 also demonstrates that alleviating the gradient problem is useful. Thus, only using the term “overfitting” without clearly establishing its relevance to diffusion models is inappropriate.

(-) The authors repeatedly claim that their method can analyze the degree of overfitting in diffusion models. However, concrete evidence to support this cannot be found in the paper. Which specific metric does the author use to analyze the degree of overfitting in diffusion models? One clear example, in the ADA paper, the probability p is used to indicate the overfitting degree of the discriminator. I suggest that the authors provide more supporting evidence for their claim.

(-) The authors state that their proposed method also acts as the regularizer in the diffusion model. Given that applying a regularizer to alleviate overfitting is commonly used, the novelty of this paper is not strong enough. Furthermore, I cannot find the comparison experiments between the proposed regularizer and the existing regularizer to demonstrate that the proposed method is indeed effective.