Including VAE

[ghub29]

I've noticed model output now requires a VAE supplement. It looks like this may also cause issues when using it as a base for Loras. With Kohya, I'm unable to use it. It throws and exception, e.g.

venv\lib\site-packages\torch\nn\modules\module.py", line 2152, in load_state_dict raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for AutoencoderKL: Missing key(s) in state_dict: "encoder.mid_block.attentions.0.to_q.weight", "encoder.mid_block.attentions.0.to_q.bias", "encoder.mid_block.attentions.0.to_k.weight", "encoder.mid_block.attentions.0.to_k.bias", "encoder.mid_block.attentions.0.to_v.weight", "encoder.mid_block.attentions.0.to_v.bias", "decoder.mid_block.attentions.0.to_q.weight", "decoder.mid_block.attentions.0.to_q.bias", "decoder.mid_block.attentions.0.to_k.weight", "decoder.mid_block.attentions.0.to_k.bias", "decoder.mid_block.attentions.0.to_v.weight", "decoder.mid_block.attentions.0.to_v.bias". Unexpected key(s) in state_dict: "encoder.mid_block.attentions.0.to_to_k.bias", "encoder.mid_block.attentions.0.to_to_k.weight", "encoder.mid_block.attentions.0.to_to_q.bias", "encoder.mid_block.attentions.0.to_to_q.weight", "encoder.mid_block.attentions.0.to_to_v.bias", "encoder.mid_block.attentions.0.to_to_v.weight", "decoder.mid_block.attentions.0.to_to_k.bias", "decoder.mid_block.attentions.0.to_to_k.weight", "decoder.mid_block.attentions.0.to_to_q.bias", "decoder.mid_block.attentions.0.to_to_q.weight", "decoder.mid_block.attentions.0.to_to_v.bias", "decoder.mid_block.attentions.0.to_to_v.weight".

Wondering I need to embed the VAE, if there is a post training process to do this or some other tool.