New textual inversion won't allow details.

[Toddzilla3000]

I created a textual inversion and the likeness is uncannily good. Unfortunately, whenever i use it, the subject is always in the same room with the same styled hair. No matter the prompt or the weights.

The pics i used for the input are well varied - different hair lengths, different environments.

Did I mess up while creating?

[claychinasky]

Probably that's the concept of over training an embed. Try less steps but maybe higher lr. This is the most difficult part of training TI compared to DB or HN, to know where to stop training in case of sustaining a flexible embedding.

[Toddzilla3000]

Thank you sir! Since it saves every 500 steps, I'll try with each embedding, using the same prompt and seed. and check it out.

I appreciate your time :)