About training the model

[xizaoqu]

Thanks for your inspiring work! The performance is really impressive. I noticed that the whole model weights of SVD have been replaced, so I am curious about the training details of DragNUWA on SVD, like which parts of the model are finetuned, how many samples are required to train the model, and how are these sampled been pre-processed.
It is OK if some details can not be published. Thanks in advance.