train dreambooth from textual inversion checkpoint

How can I use diffusers to train dreambooth from a textual inversion checkpoint ?
from the examples/train_dreambooth.py it seems it accepts only a hugging face model repo as pre-trained model or maybe I am misunderstanding something?
thanks

Start with examples/train_dreambooth.py and inspect how it accepts the pretrained model argument. Compare that entry point with the expected textual inversion checkpoint format, then document whether the workflow is supported and what input or conversion steps are required. Done means the issue has a clear, reproducible answer.

Written by the indexing model from the issue text.