Merge two embedding sequences regardless of modality, e.g., image with text in Stable Diffusion U-Net with encoder-decoder attention.