AutoencoderKLOutput

>>> import lucid
>>> from lucid.models.generative.stable_diffusion._autoencoder import (
...     AutoencoderKLOutput,
... )
>>> out = AutoencoderKLOutput(
...     reconstruction=lucid.zeros(1, 3, 256, 256),
...     latent=lucid.zeros(1, 4, 32, 32),
...     kl=lucid.zeros(()),
... )
>>> out.reconstruction.shape, out.latent.shape
((1, 3, 256, 256), (1, 4, 32, 32))
Eight times smaller in each direction and four channels wide: this is
the compression that lets diffusion run at all, and the reason it is
a latent model rather than a pixel one.