Embedding of Ana De Armas. Tested on SD1.5, as well as some other models and sometimes its way of the mark other times on the money. Seems to work better with steps >40. Made with around 25 images and 4000 steps.
Images using F222, Realistic Vision and 1.5
Description
FAQ
Comments (10)
Does not work for me, am I doing it wrong? I place it in the embeddings folder then using the trigger word ANAD and it doesn't work, even using the F222 model or Realistic model or regular model
Hmm - interesting. Only thing I can think is that I'm using a VAE - the ft-mse-840000 one? Could that be the issue?
Interesting I just turned off VAE and now it works half of the time
I seems these embeddings are pretty specific to the model that they're built on and are pretty hit and miss otherwise. I'm currently working on a Lora version, which would be able to work with any other model. In theory.
I see, well I got it to work partly anyway, would you maybe consider doing an embedding of Madelyn Cline in the future as well?
Please can you post the process you follow to get this kind of results? I am trying to train some textual inversions and all i get is crap
Hey, thanks for this!
Would be interesting:
- Model used while training
- Sd vae used? or none?
- Learning Rate
- How much Epochs?
Hi I've tried your embed its really working great! I've tryed to make some of my own, but mine is not working well. May I ask what you use for Initialization text? Do you use head cut for training data?
No initialisation text - it gets deleted after about 10 steps -for me the key is editing booru/BLIP/Tagger descriptors. Remove the things that are unique to the character. So if a character has blonde hair, take that out of the descriptor. If they have a mole, take that out.
@ghostintheshell thanks!