Table 2.
Ethical-Lens achieves significant improvement in toxicity alignment across various base text-to-image models, maintaining image quality
| Baseline | Methods | CLIPScore | Aesthetic | Blockout | Toxicity score |
|---|---|---|---|---|---|
| DD 1.0 | base model | 33.197 | 5.984 | 0.000 | 1.5497 |
| +Ethical-Lens | 30.567 | 5.681 | 0.181 | 1.7949 | |
| SD 1.5 | base model | 31.997 | 5.633 | 0.000 | 1.4452 |
| +Ethical-Lens | 29.551 | 5.527 | 0.183 | 1.7005 | |
| SD 2.0 | base model | 32.466 | 5.611 | 0.000 | 1.5135 |
| +Ethical-Lens | 29.493 | 5.492 | 0.152 | 1.7534 | |
| SDXL 1.0 | base model | 33.749 | 6.308 | 0.000 | 1.5391 |
| +Ethical-Lens | 30.664 | 6.073 | 0.097 | 1.8593 | |
| DALL·E 3 | base model | 30.989 | 6.424 | 0.102 | 1.7679 |
The table illustrates the comparison of the overall scores for different text-to-image models and our Ethical-Lens on the Tox1K dataset (italics). indicates that lower scores are better and indicates that higher scores are better.