about paper

[oceanzhf]

This work is excellent and has provided me with enlightening thoughts. However, I still have a few questions I would like to ask the author. First, I noticed you mentioned that the change in 𝑃 represents the additional bias added to the AI-generated images by the image encoder. Why is it referred to as an additional bias? Could it perhaps be the inherent bias between AI-generated images and real images? Furthermore, although the 𝑃 values exhibit consistency after dimensionality reduction using t-SNE, the bias between AI-generated images and real images may itself fall within a very narrow range. This could cause the distribution of 𝑃 to reflect the appearance of the images you provided.

[xsc1234]

Thanks for your attention for our work, we are very glad to answer your questions: Q1: Why P it referred to as an additional bias? A1: It is discussed in Section 5.1 of our paper, the core of the bias is come from the relevance score, which is calculated by similarity between representations. P is the transformations that debiased image encoder performs on the representations of image, so it can be viewed as the additional bias.

Q2: the bias between AI-generated images and real images may itself fall within a very narrow range A2: It is exactly this paper wants to point out, the images with different semantics have the bias that fall within a very narrow range. It indicates that the bias may be like the watermark that is universal information for the image generation model and can be expressed by neural network visual models such as image encoders.