Table 3.

Ethical-Lens significantly enhances the toxicity alignment across various perspectives, mostly surpassing the performance of DALL·E 3

Methods	GPT4-V evaluation $\downarrow$					HEIM evaluation $\downarrow$
	Nudity	NSFW	Public	Politic	Culture	NSFW	Nudity
DD 1.0	0.044	0.078	0.158	0.163	0.041	0.037	0.051
+Ethical-Lens	0.023	0.009	0.048	0.042	0.023	0.023	0.041
SD 1.5	0.097	0.078	0.166	0.157	0.033	0.069	0.077
+Ethical-Lens	0.058	0.013	0.063	0.041	0.021	0.043	0.059
SD 2.0	0.068	0.056	0.184	0.155	0.030	0.049	0.054
+Ethical-Lens	0.052	0.015	0.060	0.033	0.007	0.036	0.037
SDXL 1.0	0.046	0.068	0.182	0.160	0.033	0.047	0.046
+Ethical-Lens	0.039	0.006	0.014	0.009	0.009	0.030	0.028
DALL·E 3	0.015	0.042	0.021	0.084	0.050	0.014	0.020

The table illustrates the comparison of scores across each alignment perspective within the toxicity dimension for different text-to-image models and our Ethical-Lens on the Tox1K dataset (italics). $\downarrow$ indicates that lower scores are better.