Fair Text-to-Image Diffusion via Fair Mapping
Authors: Jia Li, Lijie Hu, Jingfeng Zhang, Tianhang Zheng, Hua Zhang, Di Wang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | With comprehensive experiments on face image generation, we show that our method significantly improves image generation fairness with almost the same image quality compared to conventional diffusion models when prompted with descriptions related to humans. Comprehensive Experimental Evaluations: Finally, we conduct comprehensive experiments of our methods to ensure our generated images fairness and quality. |
| Researcher Affiliation | Academia | 1Provable Responsible AI and Data Analytics (PRADA) Lab 2King Abdullah University of Science and Technology 3SDAIA-KAUST 4Institute of Information Engineering, Chinese Academy of Sciences 5University of Auckland 6The State Key Laboratory of Blockchain and Data Security, Zhejiang University 7Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security EMAIL |
| Pseudocode | No | The paper describes methods using text and mathematical equations, such as equations (1) through (7), and a diagram (Figure 3) illustrating the training and inference stages. It does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about providing source code, nor does it include a link to a code repository in the main text or references. |
| Open Datasets | Yes | We select a total of 150 occupations and 20 emotions for the fair human face image generation following [Ning, Li, and Jianlin Su 2023]. For sensitive groups, we choose gender groups (male and female), racial groups (Black, Asian, White, Indian) and Age (young, middle age, old) provided by [Kärkkäinen and Joo 2021]. |
| Dataset Splits | No | The paper states: 'Standardized prompts are used for occupations and emotions. More details are in Appendix.' However, it does not provide specific details on dataset splits (e.g., percentages or counts for training, validation, and testing) within the main body of the paper. |
| Hardware Specification | Yes | On an Nvidia V100, our debiasing method trains on 150 occupations in just 50 minutes, demonstrating impressive efficiency. |
| Software Dependencies | Yes | In our experiments, we use stable Diffusion v1.5 [Rombach, Blattmann, and Dominik Lorenz 2022] as the base model and implement 50 DDIM denoising steps for generation. |
| Experiment Setup | Yes | In our experiments, we use stable Diffusion v1.5 [Rombach, Blattmann, and Dominik Lorenz 2022] as the base model and implement 50 DDIM denoising steps for generation. Standardized prompts are used for occupations and emotions. |