This is kind of experimental.
It works on any realistic model with a normal negative prompt, or on AnyLoRA with my usual negative prompt with negative embeddings.
See the examples for the various comparisons.
A weight around 5.1 works fine from my tests. Triggers with edward candra
.