latents = torch.randn(
    (batch_size, unet.in_channels, height // 8, width // 8),
    generator=generator,
)
latents = latents.to(torch_device)