Question on LDM Conditioning Mechanisms (Concatenation vs. Cross-Attention)

[HONGcogml]

Hello,

I’d like to confirm my understanding of the conditioning mechanisms in Latent Diffusion Models (LDMs):

A. Spatially aligned image conditions (images/masks/semantic maps)

1. Downsample(interpolation) to the latent resolution → concatenate with the input Zt along the channel dimension and feed into the UNet once at the input stage.
2. It is channel-concatenation–based injection, not cross-attention–based injection.
3. τθ=ID

B. Text / token-based conditions (y)

1. Create an embedding τθ(y) via tokenizer/Transformer.
2. Inject via SpatialTransformer (cross-attention) at multiple UNet stages, using the embedding as context (K,V).
3. Not channel concatenation; it is cross-attention–based injection.

Could you please let me know if this is accurate?
Thank you.