Support Video to Video generation

[alfredplpl]

I would like to v2v by your model. I think we need to add two points on opensora/sample/pipeline_videogen.py.

1. Create the encode_videos function like the follow:

   def encode_videos(self, videos):
       latents = self.vae.encode(videos)
       ...
       return latents

2. Add code on the prepare_latents function so that the function copes with noised latents.

Any idea?

[LinB203]

Very creative idea. We can make further attempts. But I believe that before that, our model needs to be generalised enough, which means a lot of data to train.