Ashenthorn

I really liked the old model, even though it took me a LONG time to fine tune and get the results I liked.

I’m really having fun with the new model and will continue to play with it.

The differences that I currently find lacking with the new model are:

1. The lack of artist styles recognition.
2. a certain generic default e-girl sameness look… which is made worse by #1.
3. multiple images with the same prompt can look extremely and randomly different (paradoxically, considering #2… but also made worse by #1).

I think (hope) these can/will be improved with time, model training, and prompt tuning.

It’s funny… I’ve spent the past year and a half working on my prompts and fine-tuning my generator(s), for the style/results that I like. I got it to where I was very happy with the results - it was generally nothing resembling a lot of the “generic” images I’ve been seeing elsewhere on the web, and I used a lot of combinations of styles to get there - especially using the artist styles plugin.

Using the same prompts with the new model has generated some cool images, but it definitely seems to fall back to a “samey-ness”, and a definite “default style”.

Hopefully, prompt-tuning and model learning will help. I am having fun playing with it so far.