#cross-attention

Question 1: Select all that apply

Which statements about latent diffusion models (LDMs) as introduced by Rombach et al. are true? Select all that apply.

The autoencoder is trained first, and once trained it can be reused for many diffusion models1
The latent is a single flat vector, so the denoiser is an MLP rather than a convolutional U-Net2
A larger downsampling factor $f$ always gives better samples, because the diffusion model then gets cheaper to run3
In cross-attention, queries come from U-Net features and keys/values from the prompt encoding $\tau_\theta(y)$ 4
The decoder $\mathcal{D}$ runs at every denoising step to check the image in pixel space5