I'm uncertain if this is an emergent behavior in the new safe pack training I did for Simulacrum v38, but check this out.
Running my flux1d normal 1d merged v32 simulacrum, which has a heavy bit of anime and booru tag training.
a girl waving, (black hair:2.0)
<lora:safe_fixers_v1-000028:1>
Pretty normal looking right?
a girl waving, black hair
<lora:safe_fixers_v1-000028:1>
Hmmm... she appears to have grown a finger.
Now here's the fun stuff.
a girl waving, (black hair:-1)
<lora:safe_fixers_v1-000028:1>
squints
Hmmmmm... what happens if we do like -4?
Okay not too much, but what happens if we add another tag in. Black hair is highly potent in this pack.
What happens if we try 1.0 with black hair, purple hair?
Yeah that looks pretty purply black, now what if we do -4 on black hair?
Oh hello there negative prompt in the positive prompt box.
As of right this minute, I'm uncertain what this means.
I could be weighting the lora, or I could be weighting some mechanisms in T5, or I could be weighting nothing and it's just using the ascii itself.
Additional testing is necessary to determine if this viable or not.
The model itself has a divergent CLIP_L and that could be heavily attributing to the negative potential, but it needs more testing.