Fair Image Generation from Pre-trained Models by Probabilistic Modeling

For the contributions of this work, introducing PCs and introducing the sensitive attributes in PC training are the main improvements. However, there could be shortcomings in both contributions:

1. The former one may not provide a very overwhelming improvement since there are previous methods tried to introduce other models, as described in Sec. 2, while the PCs are also an existing method.

2. The latter contribution is not detailly described. The formal definition of the sensitive attributes $S$ is not clarified. Since it is determined by the sampling algorithm, it is not clear how to ensure that the PCs can correctly learn the attributes. In the experiments, the dataset is divided by the gender attribute. But among images of one gender, the other attributes could also be not fair.

For the experiments, it can be observed that the method works well on CelebA for gender fairness. However, there could be many other cases not evaluated:

1. The performance for some other attributes is not evaluated. For CelebA, attributes such as race, age, or hair color are not considered. The performance on different datasets, such as CIFAR-10 and LSUN, which contain more attributes is not evaluated. It may be necessary because the results can demonstrate that the PCs can process different attributes' features.

2. The performance for different models is not evaluated. Since one advantage of the method is that it is easy to apply without the requirement of retraining, experiments only on VQ-GAN may not be enough to prove its effectiveness. On the one hand, widely utilized GAN models, such as StyleGAN, should be considered. On the other hand, different generative models, such as VAEs, flow-based models, and diffusion models should also be considered.

1. The introduction is not well organized
   1. Need some background of probabilistic circuits in the introduction before mentioning in the contribution sections.
   2. Use reference when making some claims, i.e.
      1. On page 1, line 47, "Rather, the users and ML practitioners for downstream tasks may be interested in the distribution of the generated samples and ensuring that it is not biased with respect to some sensitive attributes", - add reference.
2. In the proposed method, as the authors tried to manipulate the latent space using both latent space and sensitive attributes, a comparison should made with [1]; they also manipulate the latent space by minimizing the correlation distance between the sensitive and non-sensitive attributes. Besides this, the authors should also compare their work with more recent work [2].
3. The overall writing should be improved
   1. In the Related Works section, some examples of probabilistic circuit-based generative models should be given
   2. The author should also mention this work's limitations and future works

References

[1] Liu, Ji, et al. "Fair representation learning: An alternative to mutual information." Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2022.

[2] Teo, Christopher TH, Milad Abdollahzadeh, and Ngai-Man Cheung. "Fair generative models via transfer learning." Proceedings of the AAAI conference on artificial intelligence. Vol. 37. No. 2. 2023.

The overall quality of the paper is quite low due to several critical issues:

• Presentation and Clarity: The paper's presentation requires significant improvement to enhance clarity. Specifically:

1. There is no clear motivation for the use of PC and why it is particularly relevant for the task. What advantages does it offer over existing methods that also aim to learn the distributions of fair latent codes, such as Gaussian Mixture Models (GMM)? Furthermore, it is unclear if integrating PC into this problem poses any unique challenges, or if it is merely a straightforward "plug-and-play" approach.

2. The method lacks a concise, high-level summary, making it difficult to follow. Nowhere in the methodology section is it explicitly explained how PC is integrated with different generative models. Based on my understanding, the authors use PC to learn the distribution of latent codes to match a reference distribution. However, the description of this approach is vague, with statements like "The training algorithm is the same as the previous case" or "An overview of the proposed method can be seen in Algorithm 1," which fail to provide sufficient detail.

• Limited Contribution: The paper’s contribution appears to be minimal, as it primarily focuses on applying PC to learn a set of latent codes without introducing substantial innovation or advancement beyond this.

• Inadequate Evaluation: The experimental evaluation is limited in scope, considering only a single dataset, a single attribute, and a single model. This restricted setup weakens the generalizability and robustness of the findings. Some other models should be considered such as Stable Diffusion or StyleGAN. Some attributes for face generation should be considered such as "BlackHair Young Smiling Moustache".

• Lack of Analysis and Ablation Study: The paper does not include any analysis or ablation study that would help elucidate the specific effects and impact of the proposed method.