Quoted By:
For those using v4 and struggling with actions, the source# and target# stuff seems to work relatively consistently. It still mixes up who's doing what sometimes, but it's way more reliable than just using prose to describe the action ("Character 1 is doing X to Character 2")
For example, if you want a character groping another from behind, the character prompts would be something like:
>girl, yukihana lamy, hololive, source#groping from behind, smug
>girl, kazama iroha, hololive, target#groping from behind, surprised
It's funny because the model kind of understands facial expressions out of the box, but it really struggles with any kind of actions when you are using more than 1 character
