Normalization is currently applied directly to latent['samples']. This may cause various side effects (ie, bypassing already applied normalization still returns normalized latents).
We can fix this by cloning input samples and incapsulating them in a new latent.
Normalization is currently applied directly to
latent['samples']. This may cause various side effects (ie, bypassing already applied normalization still returns normalized latents). We can fix this by cloning input samples and incapsulating them in a new latent.