The inpaint part is currently missing and will be implemented later.

I think they messed up this model pretty bad. They added some
control_noise_refiner blocks but don't actually use them. There is a typo
in their code so instead of doing control_noise_refiner -> control_layers
it runs the whole control_layers twice.

Unfortunately they trained with this typo so the model works but is kind
of slow and would probably perform a lot better if they corrected their
code and trained it again.