The field of generative AI, particularly in the realm of diffusion models, is witnessing significant advancements aimed at addressing inherent biases and improving fairness, especially in sensitive applications like healthcare and geoscience. A notable trend is the development of methods that debias text-to-image diffusion models without the need for extensive re-training or reliance on external datasets. These methods leverage self-discovering processes to identify and mitigate biases related to gender, race, and ethnicity, thereby enhancing the equity of generated images. Additionally, there is a growing focus on adapting diffusion models for specialized domains such as medical imaging and geoscience, where data scarcity and domain-specific challenges prevail. Innovations in this area include techniques for counterfactual medical image synthesis and the generation of multi-modal pairwise data, which are crucial for scientific computing and research. These developments not only improve the quality and applicability of generative models in specialized fields but also address critical ethical concerns by promoting fairness and reducing stereotypes in generated content.

Noteworthy Papers

• DebiasDiff: Introduces a plug-and-play method for debiasing text-to-image diffusion models, eliminating the need for re-training or external datasets, and significantly outperforms previous methods in mitigating gender and racial biases.
• FairDiffusion: Presents an equity-aware latent diffusion model that enhances fairness in medical image generation, supported by the creation of FairGenMed, a dataset for studying fairness in medical generative models.
• Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis: Proposes a method to adapt diffusion models for medical imaging, enabling counterfactual image generation without the need for large datasets, demonstrating significant performance gains.
• A Novel Diffusion Model for Pairwise Geoscience Data Generation with Unbalanced Training Dataset: Introduces UB-Diff, a diffusion model for generating multi-modal paired scientific data, showing superior performance in generating reliable and useful data for geoscience applications.
• OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes: Develops a quantitative measure of stereotypes in text-to-image models, revealing that newer models still generate images with widespread stereotypical attributes, especially for nationalities with lower Internet footprints.