A bit too overbaked on a very miniscule dataset.
Use this at strength 0.6 for a bit of leeway.
I might have to re-train this later... Dammit.
