Question about Autoencoder loss function, and training pipeline confirmation about stablesr

[IceClear]

Hi.
For the loss config, we do not train the first stage model, i.e., VQGAN, so it is fixed and the loss is set to identity. You may refer to the training process of Latent diffusion for a better understanding. In the code, you can find that this loss config does nothing during the training of the diffusion part.

For the training pipeline, you can refer to our paper for details. Your understanding is mostly correct. Only one thing: after training the diffusion part, we just generate the training data using the diffusion part. After that, the diffusion part is not involved in the training process anymore and we just use the generated data to train CFW, the same as training VQGAN.

[DZY-irene]

Thank you for your reply!