[Other][LORA]HiRyS, it's AI-RyS! (LoRA) v0.1

0xSoul

WARNING: This model was briefly auto-hidden by CivitAI for preview images containing 'loli' content. There was no such thing in the removed images - the only difference between them and the images that were approved to stay up is that the removed ones had 'hololive' in the prompt. If you want to post your gens using this LORA, make sure not to use 'hololive' in the prompt, or scrub the EXIF data first!

The base Anything and NAI models aren't so great at generating images of a certain virtual YouTuber with horns, halos, and a gorgeous voice. Here's an attempt to fix that!

ALPHA VERSION: This is literally my second attempt at training a LoRA. I've gotten some good results with it but visual artifacts are common and unusual prompts may give some very low-quality outputs. I hope to upload better versions as I learn more about LoRA training.

Trained using AnythingV3 fp32 and its corresponding VAE; use those for best results. I've had some success using it with other models but results may vary.

For best results, use keywords that describe her in your prompt - e.g. long hair, multiple horns, [red|purple] hair, fangs, and perhaps her name as well. No trigger words are necessary to get a good approximation of her standard appearance, but here are some things to try:

• white dress is likely to get her official outfit; it's a bit overtrained so try putting it in your negative prompt if you want her in something else.

• official_outfit is, in theory, keyed to that official outfit as well, but it doesn't seem to affect output much. Again, I'm still learning; hopefully future versions will be better!

• kimono is likely to get her alternate outfit, although this one isn't well-trained and often comes out distorted or too short.

• double halo and star halo for better chances of getting her distinctive halos.

If artifacting is particularly bad on a prompt you're attempting, try turning the LoRA strength down and adding more keywords describing her to the prompt - I usually get best results around 0.75-0.85 strength.

(note: if Cover so much as gazes in my direction this is gone. Please do not sell images generated with this LoRA, sell or provide paid access to the LoRA itself or merges or retrains of it, or use this LoRA as part of a paid service.)