Made from a dataset I am curating, "Unknown Wishes" based on some interesting looking art I pinned on pinterest.

It works surprisingly well on short prompts (a quirk of how Anima Standalone Trainer treats caption files that have multiple lines, ended up training with lines of my captions, instead of it entirely)

Trained for 110 something epochs, no repeats, good bunch of images, captioned entirely with "llama-joycaption-beta-one-hf-llava.Q4_K_M.gguf" by mradermacher if i'm not mistaken.

Left it "slowly cooking", constant learning rate and all that.

I adjusted the captions a bit, specially in the subset "Rusted Hunter" but besides giving each image a unique and "poetry-like" title, I didn't do much, so SOME of the images, by now, I have noticed their captions contain a bit of inaccuracy or are not 100% correct, to be corrected in a future version.

Trained on top of Anima Base, using Anima Standalone Trainer.

If you like the gritty and "rough" dark art sort of style, you will love this lora.