The combined versions lean most heavily toward realism.
I am generating images in more cartoon/amine styles and using those to balance the datasets one concept at a time.
Version 2.0 and Refined both do fisting really well. It was the easiest concept to train, so I will likely wait to do self-fisting until last.
Recommended settings:
Cfg: 5
Steps: 25-28
Weight of 1 or 0.9.
Hi-res your image with a lower weight 0.6 or 0.7.
This Lora does a lot of different things. Look at the prompts to see how to group triggers, and please share your posts to my model.
Description
This is a completely re-worked version.
Added and removed source pics.
Relabled dataset manually
Balanced the number of images for each concept (not perfectly, I must admit).
It's mostly predictable, but there are some happy accidents as well.
Some concepts are still overbaked while others are a little under. I'm including weights in the triggers to account for this.
It has a lot of triggers. Do not use them all at the same time if you want it to behave.
This cost me about 5000 buzz to train, including an earlier version that I did not release.
I'll be testing it more over the next week.
Give me a buzz donation to get it live, or just wait a week.
FAQ
Comments (3)
it might need a new version or it might be me who needs to improve my prompting
Start with simple prompts, and put the triggers early in the prompt. Look to the examples to see how best to mix the triggers. It was trained on all photos so if the style bleed is too much try to hires fix it with my lora weight turned down.
New version out. Give it a try.