Finetune after&with TI?

[Nialpo]

Hello,
Is it possible to train textual embedding, get a new specific text token (previously unknown by network),
and then fine-tune the network with this token included?
I mean, fine-tuned or dreambooth or lora-tuned network has more versatility and not frozen, but they train using unknown token/vector like SKS. Probably it is better to start training with TI token/vector which is already pointing to the right custom instance?