>>57661011Nai3 was based on waifuDiffusion model (an altered stable diffusion), which in turn was based on danbooru. Back then, since Danbooru was like 100% of its training data, characters with less art than 100 were almost impossible. Due to some legal issues with the company that arranged their payments (who wanted the model to become censored) Anlatan decided to work alone and remade their entire model. So Nai4 is not an upgrade of Nai3, it's an entirely new model. Therefore, much like other image generators such as Midjourney, it can now do natural language. Even though its tagging system and most of its base is still built from the danbooru database, only like 75% of the model is focussed on Danbooru. Because of this, and the incredible strength of the new model, it can accurately handle characters (with a little prompting) with like 20 images. It's really the most impressive anime image generation model out there. I have tried others, but I keep coming back to Novelai. Not sponsored, just enthusiastic haha. No idea if people are interested in this, but thought some people might like the info. Thanks for coming to my Ai Ted talk.
