how to support different size in the latent space when training

[icelighting]

hello, thanks for the great work. there is a question when i train the control net. when inference the denoise steps, the ddim sample size is not just the64*64, how to make it