My art vocabulary probably isn't good enough to give proper feedback, but a few differences stick outThis is not img2img.
I used the left one to generate a openpose rig - then used that rig to generate the right side images..
For those who don't use SD, the left image itself is not used at all to make the right one(s), I only used the left image to generate a skeleton rig which is basically a JSON file of lines that represent bones / fingers that can be used to pose characters, the pixels / original image is then thrown away, it's just a way to copy the pose, not the actual aesthetic.
Noxian Nights sequel?
There are small fixes / clean ups that could be done on the below images..
I was rushing a little because I'm trying to test how well the LORA works..
The goal should be:
- very consistent poses / facial expressions without relying on a crutch like img2img, because ideally you want to create multiple clothing types
- have very deterministic outcomes when you prompt or pose something, you don't want "randomness" coming through your LORA / Model otherwise it makes it very difficult to create assets.
View attachment 3069262 .
- Nomo's default shading for his female characters tends to be paler with heavy contrast from pink compared to what the ai is spitting out.
- The expressions need to be more pronounced.
- Nomo tends to draw female characters with distinctly pointed and upturned noses which is only sort of present here.
- A lot of nomo's work has 'sketchy' elements or very textured brushes that I think an AI would struggle to imitate.
- The bull skull tattoo would probably need to be manually added to any picture with bare shoulders.
At the moment it seems like this would be better for adding minor elements like new backgrounds rather than new CG. Which could still be useful, you can do a lot just by swapping out a few small elements.