What are the sources for the images?
#3
by
Kulusic
- opened
They produce great results, but the embedding tends to conform to the same light/shadow shapes. I'd like to gather more images in the same style to create a more robust embedding.
Yeah, they are trained on images with that type of lighting so it will try very hard to conform to that. These were trained on a few other sources from a game, but I mostly (99%) use other generated AI images with a mix of midjourney and my own 1.5 model I trained that can do a variety of styles to train my SD V2 embeddings to prevent any copyright infringement or annoy artists in general.