After much tweaking and tuning, I now have a version of my model that I am content with releasing publicaly. This, is PRIMEVAL NOVA V2.
While not good at realism or photo like images, it instead is able to capture abstract ideas and concepts from prompts and make them into imaginitive scenes. I use it mostly to help inspire new ideas, concepts, and so on as a basis for larger projects.
I can't give a exact epoch or steps for training, since it went through so MANY diferent finetunes layerd on top of one another. Just know it's a mangled mess of images and fun.
This model works best without over tuned prompts, giving it freedom to fly. For negatives, avoid generic negative lists and instead only use what really is opposing of your idea. Making a hellish scene? write "angellic, peaceful" as the negative. It's with simple nudges I find this model shines.
This model needs no VAE, as it is built in. I also suggest playing around with the sampler methods when generating images, as each one has it's own strengths and weaknesses.
Description
FAQ
Comments (5)
Awesome composition! Could you make a full checkpoint based solely on training data generated images from this checkpoint?
I could, yeah. for the best results though, I would have to generate a lot of diferent images and tag them. Maybe another time I will, just to "cement" the style even more.
If anyone else wishes to do it as well, feel free. Be creative, go wild lol
The advantage as you produce the content is that if you check the setting to produce a txt file with every rendered images, you can use it for captioning!
Amazing Generations,
My niece hand drawings are better than this wtf