It is deeply silly that the input and the output go in the same bucket. These things have how many context heads, in how many layers, and you can't count to two? It's not like it's novel - diffusion's had separate positive / negative prompts the whole time.