I remember doing some random experiments with these two researchers to find the best way to condition the stage B on the latent, my very fancy cross-attn with r

I remember doing some random experiments with these two researchers to find the best way to condition the stage B on the latent, my very fancy cross-attn with relative 2D positional embeddings didn't work as well as just concatenating the channels of the input with the nearest upsample of the latent, so I just gave up ahah. This model used to be known as Würstchen v3.