# scale and decode the image latents with vae
latents = 1 / 0.18215 * latents
with torch.no_grad():
    image = vae.decode(latents).sample