Wall of text about negpip prompting.

I did some experimentation since there are multiple ways you could prompt with it. For instance, if the goal is to have the subject wearing a white gothic dress, you could use the following prompts (and more I didn't test).

gothic dress, (black dress,:-1.0)
gothic dress, (black dress,:-1.0) white dress,
(black:-1.0) gothic dress,
(black:-1.0) gothic dress, white dress,
(black:-1.0) gothic dress, white gothic dress,
white gothic dress, (black dress,:-1.0)
white gothic dress, (black gothic dress,:-1.0)
(black:-1.0) white gothic dress,

And then you can also test with different colors as the theme to make sure it's stable. For instance, aqua theme. This is what those prompts give, with the first column being just gothic dress.

Generally and unsurprisingly the model seems to not understand what things mean the longer a comma separated segment is. So if you want to subtract the concept of blackness from the dress, you have can't just subtract black, you have to subtract black dress, and subtracting black gothic dress is not as effective. Though "white gothic dress, (black dress,:-1.0)" interestingly performed the best in terms of making everything white, while "gothic dress, (black dress,:-1.0) white dress," had more bits of the outfit in black. It makes sense why it might do that, since tag segments are sometimes interpreted as applying to different things on the image and may not not necessarily describe the same thing. So she might be wearing both a gothic dress and a white dress that's not black, but not necessarily a white gothic dress, which the model might think is entirely white.

Wall of text about negpip prompting.

I did some experimentation since there are multiple ways you could prompt with it. For instance, if the goal is to have the subject wearing a white gothic dress, you could use the following prompts (and more I didn't test).

gothic dress, (black dress,:-1.0)
gothic dress, (black dress,:-1.0) white dress,
(black:-1.0) gothic dress,
(black:-1.0) gothic dress, white dress,
(black:-1.0) gothic dress, white gothic dress,
white gothic dress, (black dress,:-1.0)
white gothic dress, (black gothic dress,:-1.0)
(black:-1.0) white gothic dress,

And then you can also test with different colors as the theme to make sure it's stable. For instance, aqua theme. This is what those prompts give, with the first column being just gothic dress.

Generally and unsurprisingly the model seems to not understand what things mean the longer a comma separated segment is. So if you want to subtract the concept of blackness from the dress, you have can't just subtract black, you have to subtract black dress, and subtracting black gothic dress is not as effective. Though "white gothic dress, (black dress,:-1.0)" interestingly performed the best in terms of making everything white, while "gothic dress, (black dress,:-1.0) white dress," had more bits of the outfit in black. It makes sense why it might do that, since tag segments are sometimes interpreted as applying to different things on the image and may not not necessarily describe the same thing. So she might be wearing both a gothic dress and a white dress that's not black, but not necessarily a white gothic dress, which the model might think is entirely white.