First things first, due to the datasets I trained on - this model produces nsfw content if you aren't specific about things like what the subject is wearing in your prompt. Depending on your CFG, you might have to fight with it a little bit. I don't test on any hardcore content, so I can't speak to that.
Next, I was going for a very specific style here. This model will evolve and I will be working to reduce the accidental nsfw stuff. There's some fine-tuning that has to be done.
That said, I'm really happy with this model and produces something different. It was trained on a specific set of images from the National Gallery of Art's archive.
https://www.nga.gov/open-access-images.html
I have found that the most reliable CFG for me has been between 5.5 - 7.5
Another thing about this model, I didn't test it on Euler samplers. It'll work just fine on them, but this was specifically trained around DPM++ 2M SDE (w/Huen and w/o Huen) & Karras as a scheduler. My sample images used about 50 steps, and 30 hires steps with the 4x_foolhardy_Remacri upscaler.
I like this model a lot, it's got a bit of personality that you have to fight with a bit sometimes... that'll get worked out in future versions I'm sure - or at least subdued.
Description
FAQ
Comments (2)
It makes cool stylized females, but for some reason they all have very distinctive facial features. Even if try use actress or characters in prompt - they will look like sisters. Probably author used 1-mage lora, and it produces image of same person.
If you look at demo images in header, and images of other users below - you clearly will see males and females characters have same face.
This was trained on a specific custom dataset, I'm working to reduce that impact on v2 which should be finished sometime next week. V1s are always a little surprising as people let me know what's working and what isn't. Mostly though, THANK YOU for the constructive comment. I take these right to my notebook for things to work on.