Is your feature request related to a problem? Please describe.
Classifier-free guidance is useful to make conditioning stronger in diffusion models. For this, a portion of the conditioning passed as input is set to a value that means "no condition". Currently, this is set to -1. However, this value might not be suitable for certain applications (e.g. text representations are usually normally-distributed, so -1 would not make sense here).

Describe the solution you'd like
Give users the possibility of specifying the "unconditional" value.