Table 4.
Ethical-Lens achieves substantial improvement in bias alignment across various base text-to-image models, maintaining image quality
| Baseline | Methods | CLIPScore | Aesthetic | Blockout | Bias score |
|---|---|---|---|---|---|
| DD 1.0 | base model | 29.618 | 6.494 | 0.000 | 0.0968 |
| +Ethical-Lens | 28.686 | 6.443 | 0.045 | 1.0356 | |
| SD 1.5 | base model | 29.521 | 6.067 | 0.000 | 0.2902 |
| +Ethical-Lens | 28.601 | 6.209 | 0.040 | 1.0081 | |
| SD 2.0 | base model | 29.966 | 5.907 | 0.000 | 0.3012 |
| +Ethical-Lens | 28.851 | 6.140 | 0.042 | 0.8404 | |
| SDXL 1.0 | base model | 29.950 | 6.694 | 0.000 | 0.2654 |
| +Ethical-Lens | 28.506 | 6.780 | 0.037 | 1.0501 | |
| DALL·E 3 | base model | 28.584 | 7.057 | 0.007 | 0.6188 |
The table illustrates the comparison of the overall scores for different text-to-image models and our Ethical-Lens on the HumanBias dataset (italics). indicates that lower scores are better and indicates that higher scores are better.