FairCoT: Enhancing Fairness in Diffusion Models via Chain of Thought Reasoning of Multimodal Language Models

Thank you for your valuable feedback and the positive remarks about our paper. We have carefully considered your comments and made corresponding revisions to address the concerns raised.

W1: Use of Bullet Points

We appreciate this observation and have revised the manuscript to eliminate bullet lists, opting instead for a more narrative style that enhances the flow and professional appearance of the text. All updates are available in the modified manuscript (updated methodology on pages 4-6; and updated experimental results on pages 9 and 10 that we highlighted in red).

W2: Concern About Reliance on CLIP for Attribute Prediction

CLIP was used to accurately predict attributes in several related works [2,3], after showing impressive gender and racial attribute prediction performance in the original CLIP paper [1], providing an automated method to label attributes and being more reliable than classical classifiers due to its zero-shot capabilities. To ensure results are more accurate, we integrated wider demographic attributes for races that are hard to distinguish visually, such as combining White, Middle-Eastern, and Hispanic Latino, into one class and Southeast Asian and East Asian into one class, and considering two age groups young and old inspired by bias literature [4]. Furthermore, we have extended the application of CLIP to include attire attributes beyond just gender and race, enabling powerful automated labeling that applies to a broader range of tasks beyond demographic classification.

W3: Length of CoT Path

As detailed in the appendix, the average CoT path length in our experiments is approximately 3.6 steps, with a maximum of 6 and a minimum of 3 (section 6.2.2). Regarding its applicability for other applications, while our current expertise is primarily in fairness for vision language models, our methodology can potentially be adapted for other vision-language tasks such as zero-shot image classification, Visual Question Answering (VQA), and image captioning with minor modification related to the desired tasks.

W4: Qualitative Examples and Image Control

The images including several people were indeed intentional, aiming to demonstrate the method’s capability to generate diverse scenarios, and were limited to professions that interact with humans (e.g. a doctor in various environments that has to interact with patients). We clarified this in the updated appendix and provided additional comparative results with other baselines while including our face-shot results (we generated both face-shots and full body images), and other tasks like Judge image generation. Regarding the green noise observed in some images, this is attributed to the specific version of Stable Diffusion used SDXL for the doctor generation task. This artifact varies between versions and is a known issue with the model rather than a direct consequence of our FairCoT approach that does not alter the model weights. Therefore, we included further qualitative results for other tasks, baselines, and models in the appendix (Figures 6,7,8,9 on pages 14-17).

[1] Shen, X., Du, C., Pang, T., Lin, M., Wong, Y., & Kankanhalli, M. (2024). Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations.

[2]Shrestha, R., Zou, Y., Chen, Q., Li, Z., Xie, Y., & Deng, S. (2024). FairRAG: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 11996-12005).

[3]Chuang, C. Y., Jampani, V., Li, Y., Torralba, A., & Jegelka, S. (2023). Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070.

[4]Shen, X., Du, C., Pang, T., Lin, M., Wong, Y., & Kankanhalli, M. (2024). Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations.

We also utilized Llama-3.2-11B-Vision-Instruct for inference to validate the efficiency of our method. A sample of the prompts is provided in the appendix on page 25 sec 6.3.5 of the updated manuscript.

The performance of our method using llama-3.2-11B-Vision-Instruct was comparable to that using GPT-3.5 Turbo. We observed similar improvements in fairness and diversity metrics across the evaluated attributes compared to other baselines, and further improvements in age attribute compared to GPT-3.5 FairCoT. Moreover, LLaMA achieved state-of-the-art performance in multi-concept generation, revealing its generalization power compared to GPT-3.5. Interestingly, this demonstrates the power of the vision modality in improving the reasoning capabilities of LLMs, as earlier versions (e.g., LlaMA 3.1) fail and hallucinate when prompted to perform this task.

References

[1] Serengil, Sefik, and Alper Ozpinar. "A Benchmark of Facial Recognition Pipelines and Co-Usability Performances of Modules." Journal of Information Technologies, vol. 17, no. 2, 2024, pp. 95–107, doi:10.17671/gazibtd.1399077. Accessed at https://dergipark.org.tr/en/pub/gazibtd/issue/84331/1399077.

[2] Serengil, Sefik Ilkin, and Alper Ozpinar. "LightFace: A Hybrid Deep Face Recognition Framework." 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), IEEE, 2020, pp. 23–27, doi:10.1109/ASYU50717.2020.9259802. Accessed at https://ieeexplore.ieee.org/document/9259802.

[3] Serengil, Sefik Ilkin, and Alper Ozpinar. "HyperExtended LightFace: A Facial Attribute Analysis Framework." 2021 International Conference on Engineering and Emerging Technologies (ICEET), IEEE, 2021, pp. 1–4, doi:10.1109/ICEET53442.2021.9659697. Accessed at https://ieeexplore.ieee.org/document/9659697.

Thank you for your prompt response and for recognizing the practical insights our manuscript provides. We appreciate your acknowledgment of our efforts to tackle significant questions in AI and your constructive feedback, which has been invaluable in refining our work.

Contribution and Technical Innovation

We understand your concern regarding the perceived limitations in our contribution and technical innovation. Our contributions are multifaceted and lie primarily in the novel adaptation of the Chain-of-Thought (CoT) technique to enhance fairness in text-to-image (T2I) generative models. While CoT has been employed in language models for reasoning tasks, to the best of our knowledge, its application to bias mitigation in image generation is unprecedented. Our method involves generating fair chains of thought using a Large Language Model (LLM) capable of generating images. We allow the model to iteratively reassess its reasoning until it produces fair images. These chains of thought are designed to generalize to any LLM that generates images or any text-to-image model connected to an LLM capable of reasoning. To enable adoption with any text-to-image model, we propose using an LLM to produce refined text prompts that can guide any Stable Diffusion (SD) model toward fairer outputs. Moreover, we extend our assessment of bias and fairness beyond classical types such as gender or race, aiming to make fairness research more inclusive of underrepresented minorities (e.g., women wearing hijab) who are often overlooked in the literature. To improve the zero-shot performance of CLIP in detecting people wearing religious attire, we proposed an attire-based method. Our method successfully debiased three versions of Stable Diffusion (SD), in addition to the DALL-E model, demonstrating its generality. It is model-agnostic and leverages LLMs to generate refined prompts without the need for fine-tuning or additional training data. Furthermore, our method generalized to multi-face generation and multi-concept scenarios beyond professions, such as images involving children, laptops, and animals.

Bias in the Experiment

We acknowledge your concern that the experiment may exhibit some bias.Our decision to use CLIP for attribute prediction was based on its robust zero-shot capabilities, ensuring accurate results even on images unseen in training data, such as low-quality images generated by text-to-image models or those involving underrepresented attributes. Classical classifiers often fail in such scenarios; for example, the DeepFace classifier incorrectly predicts a white female wearing a hijab as a Black male, regardless of the actual religion or attire.

To cross-validate our results, we compared CLIP's attribute predictions with those of DeepFace [1-3] on 500 images generated by our method . We also compared these predictions against annotations from expert labelers on Amazon Mechanical Turk for age, gender, and race attributes. The results confirmed that CLIP provides more reliable attribute predictions in this context.

For attributes like religion, we introduced an attire-based enhancement to improve CLIP's attribute classification accuracy. This method reduces reliance on potentially biased features by focusing on neutral, observable attributes associated with specific religions.

We recognize that biases in CLIP might have led to lower CLIP-T scores for alignment due to its inherent biases (e.g., predicting a female doctor as a nurse, as discussed in the literature). To mitigate this, we added a penalty during the training of our chain of thought for the CLIP-T score, interrupting training if the CLIP-T score falls below a threshold (alpha × CLIP-T0) where alpha<1. Notably, the alignment score at inference achieved comparable performance regardless of CLIP's biases. Our CoT refinement process is designed to counteract biases by encouraging the LLM to reason toward diverse and representative image generation. By iteratively refining chains of thought, we aim to balance the representation of different demographic groups in the generated images.

Our qualitative and quantitative results demonstrate that our approach significantly reduces bias compared to baseline models, even when considering the inherent biases of CLIP and MLLMs.

Thank you for your thoughtful analysis and feedback. We have addressed each of your concerns as outlined below, aiming to clarify our contributions and methodologies.

W1: Technically Limited Contribution

Our key contribution is adapting CoT effectively to enhance fairness in image generation, which significantly advances over existing methods where traditional automatic CoT adjustments fail [1]. This contribution is both state-of-the-art and practical, requiring no additional training while being generalizable across various tasks. Furthermore, our attire-based CLIP enhancement method improves zero-shot classification performance in CLIP for religion, which can be extended beyond addressing bias attributes prediction.

W2: Biases in MLLMs and CLIP

CLIP is specifically utilized for attribute prediction due to its efficiency and accuracy in predicting gender, race, and age [2], as recognized in the relevant bias literature [3,4]. We also proposed our attire-based method to enhance CLIP's attribute classification for complex attributes like religion(sec 4.3). On the other hand, our chain of thought refinement method aims to debias MLLMs for image generation task by stimulating the reasoning of MLLMs for fairness resulting in more equitable image generation. We measure bias using normalized entropy, with both qualitative and quantitative results demonstrating the fairness of our images across various demographic attributes (We included additional qualitative results in the appendix of the updated manuscript).

W3: Computational Efficiency

In practice, our pipeline is optimized for efficiency; after initially training seven CoTs, no further training is required. During inference, for any given task, only a single call per prompt to Stable Diffusion is necessary, as detailed in Figure 1. This method avoids the need for finetuning, RAG, external datasets, or additional GPUs, making it both effective and practical.

Q1:Choice of "Attire" as a Representation of Religion

We did not intend to imply that attire directly signifies religion; rather, our goal was to enhance fairness by representing individuals in diverse attires who are often underrepresented. This point is clarified further in our report and noted as a limitation (please refer to the updated manuscript on page 10 lines 514-519).

Q2: Compatibility with the SD Set of Models

As depicted in Figure 1, an LLM generates a set of SD-compatible prompts from the CoT, which are then used to generate images. This method not only adds reasoning capabilities to SD models but also enhances their performance in addressing bias and fairness (please refer to sec.6.3.3-6.3.5 in the appendix of the updated manuscript on pages 23-26).

Q3: Improvement of Figure and Table Captions

Thank you for suggesting improvements to the captions of our figures and tables. We revised all captions to enhance clarity and relevance (please refer to the updated manuscript for improved captions).

[1]Shaikh, O., Zhang, H., Held, W., Bernstein, M., & Yang, D. (2023). On second thought, let’s not think step by step! Bias and toxicity in zero-shot reasoning. In A. Rogers, J. Boyd-Graber, & N. Okazaki (Eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 4454–4470). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.acl-long.244

[2]Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., ... & Sutskever, I. (2021, July). Learning transferable visual models from natural language supervision. In International conference on machine learning (pp. 8748-8763). PMLR.

[3]Shrestha, R., Zou, Y., Chen, Q., Li, Z., Xie, Y., & Deng, S. (2024). FairRAG: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 11996-12005).

[4]Chuang, C. Y., Jampani, V., Li, Y., Torralba, A., & Jegelka, S. (2023). Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070.

Impact of Biases in CLIP and MLLMs on Our Method

We appreciate your concern regarding the potential biases present in CLIP and Multimodal Large Language Models (MLLMs) and their effect on our method's results. We acknowledge that these models are not entirely free from biases, as highlighted in prior studies [1,2]. Our decision to use CLIP for attribute prediction was based on its strong zero-shot capabilities, which help ensure accurate results even on images that are unseen during training, such as low-quality images generated by text-to-image models or those involving underrepresented attributes. Classical classifiers often fail to generalize well in such scenarios. For instance, the DeepFace classifier [3-5] incorrectly predicts a white female wearing a hijab as a Black male, regardless of the actual religion or attire. To validate our results, we compared CLIP's attribute predictions with those of DeepFace on 500 images generated by our method. We also compared these predictions against annotations from expert labelers on Amazon Mechanical Turk for age, gender, and race attributes. The findings confirmed that CLIP provides more reliable attribute predictions in this context.

For attributes like religion, we introduced an attire-based enhancement to improve CLIP's attribute classification accuracy. This approach reduces reliance on potentially biased features by focusing on neutral, observable attributes associated with specific religions.

We recognize that biases in CLIP might have led to lower CLIP-T scores for alignment due to its inherent biases (e.g., predicting a female doctor as a nurse, as discussed in the literature). To mitigate this, we added a penalty during the training of our chain of thought for the CLIP-T score, interrupting training if the CLIP-T score falls below a threshold (alpha × CLIP-T0) where alpha<1. Notably, the alignment score at inference achieved comparable performance regardless of CLIP's biases.

Our CoT refinement process is designed to counteract biases by encouraging the LLM to reason toward diverse and representative image generation. By iteratively refining chains of thought, we aim to balance the representation of different demographic groups in the generated images.

Our qualitative and quantitative results demonstrate that our approach significantly reduces bias compared to baseline models, even when considering the inherent biases of CLIP and MLLMs.

Computational Cost of Multiple Inference Calls

We understand your concern regarding the computational cost associated with multiple inference calls in our method. While our pipeline does involve additional computations compared to a single-pass approach, we have taken steps to optimize efficiency and manage the computational overhead.

Firstly, the initial generation of the chain-of-thought (CoT) pool is performed offline. Once generated, the CoT pool can be reused across different prompts and tasks, eliminating the need for repeated CoT computations during inference. Secondly, during inference, each input prompt requires only one additional LLM call to generate the refined prompt, followed by a single text-to-image model inference. This approach minimizes the number of additional inference calls while achieving significant performance gains in fairness.

We conducted a cost-benefit analysis, including runtime and resource utilization, which we have included in sec. 6.2.2 on page 22 of the updated manuscript. The analysis demonstrates that the computational overhead is manageable and justified by the substantial improvements in fairness and diversity metrics achieved by our method.

Moreover, we believe that the additional computational cost is acceptable for fair image generation that generalizes to any bias type or concept, especially since it can be applied upon request—when fair generation is specifically needed by the user.

Clarification on How SD Fits into the Figure 3 Pipeline

We apologize for any confusion regarding the integration of Stable Diffusion (SD) into our pipeline. Allow us to provide a clearer explanation:

Figure 3 Pipeline: In our pipeline depicted in Figure 3, the MLLM generates a refined Chain-of-Thought (CoT) process for fair image generation.

Enhancing SD with MLLM Reasoning: At inference (Figure1) the MLLM generate prompts the are specifically formatted to be compatible with SD models.By incorporating MLLM-generated reasoning into the prompts, we provide SD models with enriched context, leading to fairer image generation without modifying the SD models themselves.

Instead of using handcrafted prompts written by humans to generate fairer images with SD, we utilize prompts generated by an MLLM. These MLLM-generated prompts are designed to be fair and specifically targeted for image generation, as they are produced by an MLLM capable of generating images. To adapt our approach to SD, we generate prompts that effectively guide the SD model's image generation process.

Similar to research in reasoning where prompting LLMs to "think step by step" improves their reasoning compared to hand-crafted chains of thought, we demonstrate that MLLM-generated chains of thought enhance fairness in image generation more effectively than human prompting and fine-tuning methods.

Our approach equips SD models with improved context derived from the MLLM's reasoning process, enabling them to generate images that are fairer and more representative without altering the SD models themselves. This integration ensures that the benefits of the MLLM's reasoning are seamlessly transferred to the SD models through the refined prompts.

References

[1]Agarwal, Sandhini, et al. "Evaluating clip: towards characterization of broader capabilities and downstream implications." arXiv preprint arXiv:2108.02818 (2021).

[2]Berg, Hugo, et al. "A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning." arXiv preprint arXiv:2203.11933 (2022).

[3] Serengil, Sefik, and Alper Ozpinar. "A Benchmark of Facial Recognition Pipelines and Co-Usability Performances of Modules." Journal of Information Technologies, vol. 17, no. 2, 2024, pp. 95–107, doi:10.17671/gazibtd.1399077. Accessed at https://dergipark.org.tr/en/pub/gazibtd/issue/84331/1399077.

[4] Serengil, Sefik Ilkin, and Alper Ozpinar. "LightFace: A Hybrid Deep Face Recognition Framework." 2020 Innovations in Intelligent Systems and Applications Conference (ASYU), IEEE, 2020, pp. 23–27, doi:10.1109/ASYU50717.2020.9259802. Accessed at https://ieeexplore.ieee.org/document/9259802.

[5] Serengil, Sefik Ilkin, and Alper Ozpinar. "HyperExtended LightFace: A Facial Attribute Analysis Framework." 2021 International Conference on Engineering and Emerging Technologies (ICEET), IEEE, 2021, pp. 1–4, doi:10.1109/ICEET53442.2021.9659697. Accessed at https://ieeexplore.ieee.org/document/9659697.

Thank you for your continued engagement and for providing further feedback on our work. We value your insights and would like to address your remaining concerns to clarify our contributions and methodologies.

Limited Contributions and Generalization of the CoT Technique

We understand your concern regarding the novelty and generalization of our approach. Our contributions are multifaceted and lie primarily in the novel adaptation of the Chain-of-Thought (CoT) technique to enhance fairness in text-to-image (T2I) generative models. While CoT has been employed in language models for reasoning tasks, to the best of our knowledge, its application to bias mitigation in image generation is unprecedented.

Our method involves generating fair chains of thought using a Large Language Model (LLM) capable of generating images. We allow the model to iteratively reassess its reasoning until it produces fair images. These chains of thought are designed to generalize to any LLM that generates images or any text-to-image model connected to an LLM capable of reasoning. To enable adoption with any text-to-image model, we propose using an LLM to produce refined text prompts that can guide any Stable Diffusion (SD) model toward fairer outputs.

Moreover, we extend our assessment of bias and fairness beyond classical types such as gender or race, aiming to make fairness research more inclusive of underrepresented minorities (e.g., women wearing hijab) who are often overlooked in the literature. To improve the zero-shot performance of CLIP in detecting people wearing religious attire, we proposed an attire-based method. Our method successfully debiased three versions of Stable Diffusion (SD), in addition to the DALL-E model, demonstrating its generality. It is model-agnostic and leverages LLMs to generate refined prompts without the need for fine-tuning or additional training data. Furthermore, our method generalized to multi-face generation and multi-concept scenarios beyond professions, such as images involving children, laptops, and animals.

We conducted experiments using Llama-3.2-11B-Vision-Instruct for inference and reported the results in the updated manuscript. A sample of the prompts generated by LLaMA is provided in the appendix sec. 6.3.5 on page 25. We also included the multi-face and multi-concept results on page 8. Below are the complete results:

Additionally, our method offers several advantages over existing prompt-based and fine-tuning methods. Prompt-based methods require human labor, may introduce bias, perform minor changes, and can lead to inadvertent consequences. Fine-tuning methods are computationally expensive, can lead to memorization, and do not generalize to other bias types not included in the fine-tuning process.

Thank you so much for all your feedback and help! We will take these into consideration later in our manuscript.

2. The enriched context is still within the tokens limit of SD. In fact, it was one of the inspirations for using an LLM to generate prompts that are SD-compatible. We previously tried using a short summary of the original CoT and feeding it to the SD model. Below are the SD-XL results for using summary CoT: | Model | Prompt | Gender | Race | Age | Religion | CLIP-T↑ | |----------------------|--------------|--------|-------|-------|----------|---------| | SDXL | summary CoT | 0.80 | 0.51 | 0.85 | 0.66 | 0.254 |

Thank you for your thorough review and constructive feedback. Your suggestions have been instrumental in refining our methodology and presentation. Below, we detail how we have addressed each of the points raised.

W1: Evaluation Using Author-Selected Prompts

Our choice to use profession-based prompts aligns with methods utilized in baseline comparisons such as finetuning, debiasing VL, and fair Difussion methods, ensuring consistency across evaluations and highlighting our approach's state-of-the-art capabilities [1-4]. Our method also demonstrates generalizability beyond typical profession-based scenarios, accommodating diverse settings and subjects, as illustrated in sec.4.2.2, 4.2.3, 6.2.2. Moreover, the professions list was generated by an LLM to guarantee fair evaluation.

W2: Concern About CLIP-T Usage in Methodology and Evaluation

We employed CLIP-T to ensure that the generated images align closely with the initial prompts, addressing a common limitation where images deviate significantly from specified attributes (e.g., generating an image of a 'normal female' when 'female doctor' is intended), and this metric was widely used in the literature [1-2]. While CLIP-T primarily verifies prompt alignment, we evaluated fairness using normalized entropy.

W3: More Detailed Prompt Information in the Main Text

Acknowledging the feedback on prompt clarity, we have referenced to the list of prompts used directly in the main text (please refer to page 7 lines 345-347 in the updated manuscript). The omission was originally due to space constraints, but we understand the importance of transparency in prompt categorization to enhance reader comprehension.

W4: Details on Hand-Labeled Images

The labeling was performed by a PhD student using Amazon Turk Sandbox. We will include updated labels using Amazon Mechanical Turk, with quality control ensured by employing random master labelers. Random master labelers achieved 96.7% agreement with the initial labels provided by the PhD student. We are committed to sharing this hand-labeled dataset, along with our codes, upon acceptance of the paper to facilitate reproducibility and further research.

W4: Sparse Figure and Table Captions

We thank you for pointing out the sparsity of our captions. We have revised all figure and table captions to be more descriptive, thereby improving the clarity and informativeness of our visual and tabular presentations (please refer to the updated manuscript).

Q1: Misleading Term "Training Stage"

We intended this to refer to the training of the Chain of Thoughts (CoT), which does not involve further training post-generation. We revised this to the "CoT generation stage" to better reflect the process while aligning with the terminology used in CoT reasoning research (please refer to page 4 in the updated manuscript).

Q2: Specifics on Language Model Version and Alternatives

We specified that we used the 'GPT-3.5 Turbo' version during our initial studies on page 22. Our updated version in camera-ready will incorporate 'Llama-3.2-11B Vision Instruct' in inference testing to diversify our language model usage, enhancing the paper's reproducibility and reducing reliance on proprietary models. We plan to share the Llama-3.2-11B Vision Instruct results before the rebuttals deadline.

Q3: Editorial Corrections

The noted typographical errors, such as missing spaces, have been corrected. We appreciate your attention to detail.

[1] Shen, X., Du, C., Pang, T., Lin, M., Wong, Y., & Kankanhalli, M. (2024). Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations.

[2]Shrestha, R., Zou, Y., Chen, Q., Li, Z., Xie, Y., & Deng, S. (2024). FairRAG: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 11996-12005).

[3]Chuang, C. Y., Jampani, V., Li, Y., Torralba, A., & Jegelka, S. (2023). Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070.

[4]Friedrich, F., Brack, M., Struppek, L., Hintersdorf, D., Schramowski, P., Luccioni, S., & Kersting, K. (2023). Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893.

Thank you for your thoughtful follow-up and for giving us the opportunity to address your remaining concerns. We greatly appreciate your detailed feedback, which has helped us to further improve our work.

W2: Impact of Biases in CLIP and MLLMs on Evaluation and Results

We understand your concern regarding the potential propagation of biases present in CLIP and Multimodal Large Language Models (MLLMs) into our method and evaluation. We acknowledge that these models are not entirely free from biases, as highlighted in prior studies [1,2]. Our decision to use CLIP for attribute prediction was based on its robust zero-shot capabilities, ensuring accurate results even on images unseen in training data, such as low-quality images generated by text-to-image models or those involving underrepresented attributes. Classical classifiers often fail in such scenarios; for example, the DeepFace classifier incorrectly predicts a white female wearing a hijab as a Black male, regardless of the actual religion or attire. To cross-validate our results, we compared CLIP's attribute predictions with those of DeepFace [3-5] on 500 images generated by our method . We also compared these predictions against annotations from expert labelers on Amazon Mechanical Turk for age, gender, and race attributes. The results confirmed that CLIP provides more reliable attribute predictions in this context.

For attributes like religion, we introduced an attire-based enhancement to improve CLIP's attribute classification accuracy. This method reduces reliance on potentially biased features by focusing on neutral, observable attributes associated with specific religions.

We recognize that biases in CLIP might have led to lower CLIP-T scores for alignment due to its inherent biases (e.g., predicting a female doctor as a nurse, as discussed in the literature). To mitigate this, we added a penalty during the training of our chain of thought for the CLIP-T score, interrupting training if the CLIP-T score falls below a threshold (alpha × CLIP-T0) where alpha<1. Notably, the alignment score at inference achieved comparable performance regardless of CLIP's biases. Our CoT refinement process is designed to counteract biases by encouraging the LLM to reason toward diverse and representative image generation. By iteratively refining chains of thought, we aim to balance the representation of different demographic groups in the generated images. Our qualitative and quantitative results demonstrate that our approach significantly reduces bias compared to baseline models, even when considering the inherent biases of CLIP and MLLMs.

W4: Transparency of Evaluation Criteria and Process

We appreciate your request for more transparency regarding our evaluation criteria and process. Below, we provide the details of the questionnaire used in our study:

Questionnaire:

• Instruction: "Select the religion of the person based on the visible attire."

• Options:

o Neutral

o Muslim

o Christian

o Hindu

o Sikh

o Jewish

The initial labeling was performed by a PhD student with expertise in fairness. To ensure the reliability of the labels, we conducted a validation study using Amazon Mechanical Turk (AMT) with Master Workers who had an approval rate of over 98% and more than 1,000 approved HITs. The agreement between the initial labels and those from AMT workers was 96.7%, indicating a high level of consistency.

We considered the task manageable for a single labeler because the determination of religion was based primarily on visible attire, such as specific religious garments or symbols (e.g., hijab, cross), which are generally easy to identify regardless of the image quality. In cases where religious attire was not visible, the image was labeled as "Neutral." We recognize that the varying quality of generated images could potentially affect label accuracy. However, the high agreement rate between the initial labels and the AMT labels suggests that the labeling task was indeed straightforward and reliable.

Thank you for your insightful comments and suggestions. We value your feedback and have carefully considered each point to enhance the clarity and impact of our work.

W1: Technical Contribution is limited and [1] used iterations for fairness

We appreciate the comparison to prior work by He et al. [1], which we were not aware of as it was published on October 28, 2024 after the ICLR submission deadline. We have referred to it in the updated version as concurrent work. Here, we highlight the important differences. First, the learning phase of our approach constructs several templates for Chain of Thought (CoT), that is, improving the MLLM reasoning for fair image generation that will be used to extract appropriate prompts that promote fairness. In contrast to [1], which fine-tunes the model by updating its weights, our approach does not require access to the model parameters and therefore is applicable to closed-source models. Second, at inference stage, our model selects a chain of thought that is closer to CoTs that are constructed at the learning stage and is shown to generalize to multiple biases on which it hasn’t seen at the learning stage -- in sec. 4.2.2, 4.2.3, 6.2.2 we have shown that for multi-face generation, multi-concept generation, including laptops, kids, and dog breeds. This is in contrast to [1] which requires retraining for new types of biases and has been evaluated only in the case of gender and ethnicity bias with primary experiments on 6 professions. We expanded our literature review to clarify these distinctions and highlight the novel aspects of our approach, as highlighted in the updated version lines 115-119.

W2: Comparison with IDA as Relevant Baseline

IDA was published after our submission and has yet to make its model weights or code public. For this reason, we are unable to obtain a direct comparison on all the biases that we address and on the metrics that we used including normalized entropy and CLIP-T score (and used in the literature [2,3]). However, since the paper reports results in terms of gender KL divergence over 6 professions, we evaluated our gender KL divergence on 2 common professions we share: Housekeeper and Teacher where we achieved superior performance of 0 divergence between genders. We will include those results in the updated manuscript. We commit to including IDA in our baseline comparisons and conducting a thorough evaluation once their resources become available.

W3: Clarification of Iterative CoT Updates

To better illustrate our method, we have included samples of how the chain of thought improves from zero-shot CoT, to the first iteration, and later to convergence (sec. 6.3.1-6.3.3 in the updated version). We also added for inference, an example of doctor chain of thought generation and text prompts generation to show how the trained Cots help in debiasing other tasks. This process exemplifies how our method can be generalized to other contexts, such as generating fair images for different medical professions (sec. 6.3.4 and 6.3.5 in the updated version). We will include a detailed toy example in the appendix of the final version. This example will walk through the iterative refinement process using nurse images, showing how feedback on bias and alignment is used to adjust CoTs until optimal fairness metrics are achieved.

W4: Presentation Improvements

We have revised Figure 3 and related visual materials to enhance clarity (please refer to updated Figure 3 on Page 6). Moreover, the overall presentation is improved, including limited usage of bullet points.

Q1: Image Quality in Figure 6

The CLIP-T metric focuses on prompt alignment rather than visual quality. The quality of images in Figure 6 is attributed SDXL model rather than a deficiency in FairCoT as our method doesn’t alter model weights. For example, SD v1.5 and v2.1 achieved higher-quality images we now include in the appendix. We also included in our report additional comparative images from baseline methods to further demonstrate the capabilities of our approach with detailed image examples (Figures 6,7,8,9 on pages 14-17) .

[1] Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, and Xiaojuan Qi. 2024. Debiasing Text-to-Image Diffusion Models. In Proceedings of the 1st ACM Multimedia Workshop on Multi-modal Misinformation Governance in the Era of Foundation Models (MIS '24). Association for Computing Machinery, New York, NY, USA, 29–36. https://doi.org/10.1145/3689090.3689387

[2]Shrestha, R., Zou, Y., Chen, Q., Li, Z., Xie, Y., & Deng, S. (2024). FairRAG: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 11996-12005).

[3] Shen, X., Du, C., Pang, T., Lin, M., Wong, Y., & Kankanhalli, M. (2024). Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations.

Dear Reviewer wptm, as we are approaching the end of the author-reviewer discussion period, could you please let us know if our responses have adequately addressed your concerns? If there are any remaining comments, we are committed to providing prompt responses before the discussion period concludes. Thank you for your time and consideration.