My first LoRA, and more of a proof-of-concept than anything. Mostly I wanted to see if I could actually train an SDXL LoRA with only 6GB of VRAM. The answer is yes, with concessions, and it will take 15 hours to do so.

It's a LoRA of an (in)famous moment that occurs very briefly in the anime "Yondemasu yo, Azazel-san" / "よんでますよ、アザゼルさん。", though some global viewers might know the scene better as "Cowing up!" or "Rinko sakuma vaca." A demon summoner's apprentice tries to summon a demon on her own, but messes up the ritual and summons the wrong demon. He's more powerful than the one she was expecting, and demands a greater sacrifice than the curry she cooked up. As a punishment for her failure, she is cursed to turn into a cow.

Because it's so brief, and because fan art is minimal (let alone fan art in acceptable quality), there weren't many unique images to select from for the dataset. I tweaked things here and there to better fill out a set, but it's still only 16 images, with probably too much reused framing. Since this is my first attempt and I don't know if/when there will be a second, I've elected to include my training data as a download if anyone wants to have a go at doing it better. Please enjoy the janky edits.

Still, I'm not too displeased with the outcome for what was essentially a test run. The tagging is a little wonky (I made them almost two years ago for an unrealized SD 1.0 version of this LoRA and didn't bother rewriting them for this test run). Anticipate a fair bit of tweaking and compromise, especially for visuals like the cartoonish hooves that many models can't make heads or tails of, or the horns which are only well-trained for a select few angles. Or the number of endowments, which requires counting and thus LLMs are out of luck. Careful application of other LoRAs is recommended, particularly ones for transformation concepts.

I find it works best as a spice, requiring a fairly rigorous prompt for the best results, but able to get you over the finish-line. Crank up the strength to get closer to the original scene (including poses), or tone it down for a subtler effect. Also, it leans pretty hard on the demi tf/furless side of things. You might have a tough time prompting actual fur back in.