Classifier Free Guidance for StableVideoDiffusion

[tykim0507]

Hello, I have some questions about the implementation of classifier free guidance for the current I2V Stable Video Diffusion model.

1. In training, what is the probability of dropping out the first frame latent concatenated and the image latent going into the cross attention? I read your paper, but I couldn't find the drop out probability of the image condition.

2. Is linear classifier guidance considered in the implementation for training?

Thank you