BiasMap: Leveraging Cross-Attentions to Discover and Mitigate Hidden Social Biases in Text-to-Image Generation

Bias discovery is critical for black-box generative models, especially text-to-image (TTI) models. Existing works predominantly focus on output-level demographic distributions, which do not necessarily guarantee concept representations to be disentangled postmitigation. We propose BiasMap, a model-agnostic framework for uncovering latent concept-level representational biases in stable diffusion models. BiasMap leverages cross-attention attribution maps to reveal structural entanglements between demographics (e.g., gender, race) and semantics (e.g., professions), going deeper into representational bias during the image generation. Using attribution maps of these concepts, we quantify the spatial demographicssemantics concept entanglement via Intersection over Union (IoU), offering a lens into bias that remains hidden in existing fairness discovery approaches. In addition, we further utilize BiasMap for bias mitigation through energy-guided diffusion sampling that directly modifies latent noise space and minimizes the expected SoftIoU during the denoising process. Our findings show that existing fairness interventions may reduce the output distributional gap but often fail to disentangle concept-level coupling, whereas our mitigation method can mitigate concept entanglement in image generation while complementing distributional bias mitigation.