I've seen horrible things.

But hopefully you don't have to. These embeddings mainly focus on composition, color schemes and anatomy. They are not perfect and should most likely be used along other negative prompt words, but they should do the bulk of the work.

If your model already does good compositions and anatomy, these embeddings might not help. The ones with fewer vectors might be better in this case.

Do note that the way embeddings work with prompt weighting seems somewhat broken, specially on high number of vectors. You can still fine tune, just avoid low multipliers for high vector embeddings, at least on Automatic1111.

These embeddings were also not trained, they were created based on the encoded vectors themselves.

Whatever you do, don't run this as a positive embedding. I'm not responsible for your loss of sanity.

Make sure you check out bad prompt as well.