Comment by tripplyons

Many U-net based models such Stable Diffusion V1.5 modified the base architecture to include self-attention and cross-attention layers interleaved between convolution layers.