Table 5.

Ethical-Lens substantially enhances the bias alignment across various dimensions, mostly surpassing the performance of DALL·E 3

Methods	GPT4-V evaluation $\downarrow$			HEIM evaluation $\downarrow$
	Gender	Race	Age	Gender	Race	Age
DD 1.0	0.958	0.983	0.990	0.662	0.911	0.745
+Ethical-Lens	0.256	0.504	0.584	0.215	0.590	0.504
SD 1.5	0.777	0.817	0.906	0.559	0.784	0.800
+Ethical-Lens	0.163	0.392	0.607	0.283	0.639	0.597
SD 2.0	0.674	0.752	0.877	0.622	0.865	0.854
+Ethical-Lens	0.209	0.396	0.627	0.418	0.678	0.752
SDXL 1.0	0.840	0.858	0.940	0.633	0.788	0.757
+Ethical-Lens	0.196	0.406	0.571	0.255	0.594	0.590
DALL·E 3	0.332	0.497	0.838	0.353	0.574	0.760

The table illustrates the comparison of scores across each alignment perspective within the bias dimension for different text-to-image models and our Ethical-Lens on HumanBias dataset (italics). $\downarrow$ indicates that lower scores are better.