Association of Objects May Engender Stereotypes: Mitigating Association-Engendered Stereotypes in Text-to-Image Generation

Weaknesses

• W1 (the complexity issue): As shown in Section 4.2 of the paper, we conducted experiments to evaluate the impact of MAS on the computational load of the T2I diffusion model. Table 6 demonstrates that MAS effectively mitigates stereotypes while maintaining image generation efficiency and quality.

• W2 (the computational overhead issue): The time complexity of the Sensitive Transformer is consistent with that of the Transformer. For an input batch size of $b$, sequence length $N$, and an $l$ -layer Transformer model, the computational complexity is $O(l(bNd^2 + N^2d))$, where $d$ represents the dimension of the word embeddings.

  We conducted MAS tests on our local server. As reported in Table 6 of the paper, the performance is similar to the original T2I model, with a difference of about $20$ seconds over a generation task of $100$ batches with a batch size of $10$. This results in an average efficiency decrease of $0.02$ seconds per image, which is negligible for practical generation tasks. Regarding the scalability of MAS, we applied it to mainstream T2I models for stereotype mitigation. Table 2 demonstrates the effectiveness of our MAS across different stable diffusion pipelines. Additionally, our MAS implements modular integration across various pipelines, allowing stereotype mitigation by simply embedding our MAS module into the original T2I workflow.

• W3 (the adaptability issue): Currently, our approach has only been tested for effectiveness in Text-to-Image generation. Though the concept of our approach looks feasible for Text-to-Video and Text-to-Audio generation, honestly speaking, we can not make a concrete answer about its adaptability without further deep exploration.

Overall, we will carefully revise our paper based on these valuable comments.

Weaknesses (the first part)

• W1 (TIT CLIP mapping issue): In fact, the association-engendered stereotype in T2I can only manifest in the generated images. Therefore, we need to use images as a medium to connect the relationship between prompts and stereotypes. In our task, we consider each image's stereotype to focus on the description of the object and its sensitive attributes. Thus, each image includes two descriptions: the original 2D CLIP description of the image's meaning and the stereotype-focused description that concentrates on the object and its sensitive attributes. As shown in Figure 1 of supplementary_images.pdf, although the stereotypes in the paper are derived from probabilistic statistics, we annotate the stereotypes of the images through probabilistic descriptions rather than using the probability model for prompt-image-stereotype association mapping. Finally, in Figure 1 of supplementary_images.pdf, we provide examples of stereotype descriptions.

  Training data type: <prompt, image, stereotype description>

• W2 (technical details): Our Sensitive Transformer is based on Transformer[1]. As shown in Equation (5) of the paper, $\text{Sensitive Matrix}(V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$, we changed the Transformer's task from translation to generating sensitive constraints.

  The Sensitive Transformer is trained together with the diffusion model under distribution alignment guidance. The training objective of the Sensitive Transformer is to generate sensitive constraints that align with the given prompts. Additionally, to generate the optimal sensitive constraints for the prompt, we optimized the generated sensitive constraints. As shown in Equation (6), $\sigma^* = \mathop{\arg\min}\limits_{\sigma \subseteq S_{\mathcal{Y}}} \sup |\sigma(p_x^{v(s)}) - p_x^{u(s)}|$, the optimization objective is to minimize the total variation distance between $\sigma(p_x^{v(s)})$ and $p_x^{u(s)}$.

  [1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017, 30.

• W3 (the issue of handling multiple attributes): As shown in Figure 2 of supplementary_images.pdf, the Sensitive Transformer generates sensitive constraints for each object in the original prompt using the prompt-image-stereotype embedding obtained from TIT CLIP. These sensitive attribute values are generated in the form of a dictionary. If the original prompt already specifies certain sensitive attributes, they will not appear in the sensitive constraints dictionary. By simultaneously constraining multiple sensitive attributes, we suppress the association-engendered stereotypes, achieving this in an implicit manner.

• W4 (the evaluation issue): As mentioned in lines 219-221 of the paper, we provide a brief description of the evaluation of SDTV values presented in Table 2 and Table 3. The detailed calculation method for SDTV values, including single and multiple objects and sensitive attributes, is provided in Appendix B. Given the potential inaccuracy of classifiers, we evaluate the SDTV values of these images using mathematical statistical methods rather than relying on classifiers.

• W5 (the MLP suggestion): Initially, we had the same idea as the reviewer: to use MLP to map prompts to their stereotypes. However, as shown in Figure 1 of supplementary_images.pdf, given our training data, if we skip the images and only use prompts and stereotype descriptions, we encounter the problem where the same prompt corresponds to multiple different stereotype descriptions, as illustrated in the table below.

  prompts	stereotype description
  a photo of a doctor and a nurse.	doctor, nurse, male, female,...

$doctor always male, ...$ [M-O & M-SA] | | a photo of a doctor and a nurse. | doctor, nurse, female, female,... $doctor always male, ...$ [M-O & M-SA] | | a photo of a doctor and a nurse. | doctor, nurse, male, male,... $doctor always male, ...$ [M-O & M-SA] | | ... | ... |

This problem causes the MLP's loss to remain constant, preventing it from learning effective features. By encoding the image into the embedding, we can address this issue while preserving the semantic relationship between the prompt and the image, which is the main advantage of TIT CLIP over MLP.

• W6 (about Figure 3): In Figure 3(b), the left image of the terrorist predominantly displays Middle Eastern facial features in terms of the sensitive dimension of race. After mitigation, the Middle Eastern features are covered by a mask, and the generated image then focuses on the terrorist's inherent features rather than displaying obvious racial characteristics.

Note: Due to the word limit imposed on each rebuttal by NeurIPS, we have included our responses to W7-W10 and the replies to the "Questions" in the Author Rebuttal cell. We kindly request the Reviewer L6vF to switch to the Author Rebuttal cell to view the second part of our detailed responses. We sincerely appreciate your review.

Weaknesses

• W1 (the algorithm issue): The output of Algorithm 1 is the embedding of prompt, image, and stereotype.
• W2 (about the suggestions of output examples): We sincerely appreciate your constructive suggestions. We will add examples of outputs from MAS and other baselines to the paper to demonstrate the mitigation effects alongside the results reported in Table 3.
• W3 (the experimental settings): We describe the main pipeline and key evaluation methods in Section 3. Due to the paper's page limit, we have to place the detailed experimental settings in the appendix.

Questions

• Q1: Please see W1.

• Q2: Please see W2.

• Q3 (the definition of extreme): Extreme refers to a scenario in a sensitive attribute dimension where the occurrence probability of a specific sensitive attribute value significantly exceeds the combined occurrence probability of other attribute values. We refer to this specific sensitive attribute value as an extreme attribute value. For instance, when using "an image of a beautiful woman" as a prompt and generating some images, considering the woman's race, 90% of the images may depict the white race, while other races only account for the remaining 10%.

Clarifications

• As mentioned in lines 132-136 of the paper, stereotype-free means that a T2I model should generate images with equal probability across different sensitive attribute values, avoiding significant disparities in probability distribution caused by the dependence of sensitive attributes on the object being generated. Stereotype-free distribution means that the T2I model generates images with equal probability for each sensitive attribute value, resulting in a probability distribution of sensitive attributes close to a uniform distribution.

• Please see W3.

Overall, we will carefully revise our paper based on these valuable comments.

Weaknesses

• W1(the acronym issue): We sincerely appreciate your constructive suggestions. The original acronym, Text-Image-Text CLIP, was chosen from an encoding perspective, focusing on the encoding of Text (prompt, stereotype description) and Image. We agree that this name may not be precise enough and does not fully reflect the project's theme. We will rename it to Prompt-Image-Stereotype CLIP. This new name better describes the objects that CLIP encodes and highlights the key aspects of the prompt and the stereotypes manifested in the generated images, making it more aligned with the project's focus than TIT-CLIP.
• W2 (about the default representation stereotype): We did not deny that default representations of identity are stereotypes; therefore, we use the words "may not" in lines 8-10. Our focus in lines 8-10 and Figure 1 is on the emergence of new and subtle stereotypes when the sensitive attribute of race is associated with the object of a house, rather than on the stereotypes inherent in the representations themselves.

Questions

• Q1 (the default stereotype issue): As stated in W2, the primary purpose of our paper is to demonstrate that the stereotype can be engendered by the association of two objects when these two objects individually are not stereotypical. This association-engendered stereotype is not a simple inheritance or addition of individual default representational stereotypes but is based on the interaction between the two, engendering a new stereotype.
• Q2 (the dataset issue): We have provided a detailed description of the experimental datasets in Table 8 of Appendix C.1 of the paper. Most of the data comes from publicly available datasets, which can be accessed via the references provided in the paper.

Missing Limitations

• In lines 328-334 of the paper, we already discussed the limitations of our method and the potential social impacts. We would extend that in the revised paper if that were not enough.

Overall, we will carefully revise our paper based on these valuable comments.