V2 is a merge of v1 and a Lora extracted from a dreambooth model. Might work better in some cases, worse in others. Not absolutely necessary if V1 is working for you.
This is a Lora of Canadian pop-punk singer Avril Lavigne. it was trained on the same 28 images as the embedding in my profile, though these were 768x768, whereas the embeddings training data was 512x512
Description
Took the dataset used on v1.o and trained a dreambooth. After testing various steps* I selected What I felt was the best checkpoint, extracted the lora, and then merged it with v1.0 to make this one. Works better in some models than the first, works worse in others. Personal preference, I suppose. If you enjoyed v1.0, this is not necessarily a must-have, but I offer it up for your consideration nonetheless. This is probably the final version for this individual.
*I ran it for several hours, generating models every so often at both epochs and every 200 steps. Took a ridiculous amount of drive space because I am a dummy
FAQ
Comments (2)
That looks superb! Why's the file so large though?
Because I wasn't paying attention. I was expecting it to be the same size as V1, bet when I extracted the Lora from the Dreambooth model, it created a larger file, and then when I used Supermerger it automatically adjusted the size of V1 to match. Didn't even notice it was that large until after I updated.