Hiccupwzy

@akk-123 Interested in you lora idea. Have you try this idea and get some experiment results?

I think the image `w` can be not equal to `h` for that you can simply resize it into your preferred size. And I have a small question for you....

I change this line [decode](https://github.com/CompVis/zigma/blob/4f6ff22ea104927eea8e2a88e6149f987ac20805/sample_acc.py#L367) and divide by 0.18215. Then the raw picture can turn to raw colorful picture. But the generated picture still be corrupted. 

Still not. But I retrained the facehq1024 using zigman8 config file. Honestly, the performance of my model is not good as that mentioned in the paper.

Same question as you. In my experiment, the process of generating images (first stage) and generating video with the similar motion (second stage) are decoupled. And you can use any...