Implementation of quickjkee/modulation-guidance for Anima models in ComfyUI.
This pack implements modulation in Anima's diffusion forward path using CLIP pooled embeddings:
CLIP(p) + w * (CLIP(p+) - CLIP(p-))
You can find example workflow in examples
It adds CLIP to your llm text encoder model, because CLIP is eternal.
In general, my findings are:
-
Reduces unpromted horniness levels of Anima. (This was my first time using it, god it's horny for no good reason...)
-
Often improves prompt following, particularly with tags, and characters.
-
Can reduce color leaks.
-
Can improve visual quality (subjective).
-
In some cases can be worse, but generally outputs with CLIP were preferred.
-
Text remained on same level. In some outputs it's better, in some worse, +-same overall.
-
Seems to weaken styles, but needs more testing.
- Core modulation idea and adapter format from
quickjkee/modulation-guidance(MIT).