Mixed captioning trained: tags+nl (natural language)+nl_tags+tags_nl

This one includes adaln so it is a bit bigger in terms of file size. I have another version that also included 1 or 2 more images in the dataset (the newest of the character) and I trained it in the same way, but without the adaln. When I compared both, the first one produced better results in terms of the character likeness and body proportions more consistently. On the other hand, the newer (no adaln) was better at generalizing only sacrificing a bit of thiccness (for some reason she always appears thinner).