Bias Begets Bias: the Impact of Biased Embeddings on Diffusion Models

With the growing adoption of Text-to-Image (TTI) systems, the social biases of these models have come under increased scrutiny. Previous approaches only identify such biases and fail to diagnose their sources. In this paper, we conduct a systematic investigation of one such source: embedding spaces. First, we demonstrate theoretically and empirically that an unbiased text embedding for the input prompt is a necessary condition for representationally balanced diffusion models. Next, we investigate the impact of biased embeddings on the alignment of images and prompts, a common technique for evaluating diffusion models. We find that biased multimodal embeddings like CLIP result in lower alignment scores for representationally balanced TTI models, thus rewarding unfair behavior. Finally, we develop a theoretical framework through which biases in alignment evaluation can be studied and propose bias mitigation methods. By specifically adapting the perspective of embedding spaces, we establish new fairness conditions for diffusion model generation and evaluation.