Inspired by @bakariso's <a target="_blank" rel="ugc" href="https://civarchive.com/models/133181/caco-chan-cacodemon-from-doom">Lora</a> for SD1.5. Most sources for the dataset are from arts by <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/18617250/artworks">Substance20</a>. I made it myself at first, because I like the original design Doom's and Doom II's Cacodemon and that humanized version.Z-image turbo: Horn-like plates on the knees and elbows don't really work, but other parts fairly well. Just the horns sometimes are not in the right quantity. Also, the blue tongue or the inside of the mouth needs to be examined promptly if necessary, and it still works not perfectly, but quite well. Strength is 0.95 — the best in most cases. Starting from 0.9, it begins losing features; at 1.0 it is sometimes difficult to achieve photorealism. Pony: So, it has some problems that are fine for me, but be aware that Lora will require some additional editing like inpainting most of the time, if you want to get perfect images at the start, you will probably be disappointed.The first of all that quantity of horns and the biggest problem. They should be four, but that usually happens in full-body images, then closer, then worse. From another side, the eye totally messes up in long distance, so it requires Addetailer with a high Inpaint denoising strength (0.4 by default, 0.3 and less is not enough for sure).Additional epochs made better the horns but worse the eye. I tried to find the balance between those two main problems. Also, Caco-chan should have elbows and knees plates like on her neck and cheeks, but results are bad, so even though it's possible to use, I don't recommend trying to do it. Her mouth should be blue inside like with Demi from Subverse, but combining with common lips it's not so easy and I don't have images for it.

Files with SHA256: c72afbd3bd313afec8799cf4b934ab7b62c6ae5778a8e137bfce31870bbe12f6