6 years ago, I made this dude in my mind and I loved him since. I was changing his design, giving him new outfits, made 3D models of him, etc and now it's time to make him into a LoRA! I mostly made this LoRA to experiment with AI... but it's honestly a pretty good model if ya use it correctly and if ya do enough experimenting. anyways, here's some info and stuff
RECOMMENDATIONS (V1.1)!
Sampling Method: DPM++ 2M Karras OR DPM++ SDE Karras
Resolution: 512 x 512 (or lower)
RECOMMENDED!:
I HIGHLY recommend using EnvyBetterHands and EasyNegative
Lora weight = 0.5 - 1
I recommend using Anime-ish checkpoints (Anything V5/Ink is a good one!)
You can tick the "Restore Faces" box if you're having issues with the face
NOTES!:
putting "Masterpiece, Best Quality, best quality face, beautiful eyes" in the prompt should help with generating better images
This LoRA is an EXPERIMENT and is.... not the best with full body images. It's okay with Cowboy shots tho. At the moment it's the best with Upper body and Close-up images (especially with close-ups)
The character has multiple outfits and in the detaset there are multiple images with multiple outfits. to get his main outfit, put "black hoodie, black pants, black shoes" in the prompt.
IMPAINTING THE FACE IS OFTEN NEEDED
RECOMMENDED TAGS!:
Main tag/keyword: ArnasAtakiyo
NEGATIVE PROMPT: (low quality, worst quality:1.4), lowres, bad anatomy, bad hands, four hands, four legs, two people, cut image, censored, author name, water marks, web address, artist name, multicolored hair, blurred face, bad face, heart, blue heart, (easynegative:0.1), 1girl, gloves, extra fingers,
(Feel free to add more tags to the negative prompt... or remove some tags..)
also, sometimes ya might need to add "colored tongue, blue tongue" the postive prompt
WARNING!:
This model MIGHT ACCIDENTLY generate nsfw/graphic content. This has never happened to me, however I still wanted to make this small disclaimer since I don't want anyone complaining. to prevent this, put nsfw tags and the "injury" tag in the negative prompt
EXTRA INFO!:
Trained on 46 images of the character/3D model
CONTACT INFO:
Description
Trained on 67 images (of the original VRoid 3D model)
Works alot better with full body shots (set the image width or height to 1024 OR use hires. fix)
Requires description tags (brown hair, short hair, blue eyes, etc)
You might need to play with the loRA weight (I recommend 0.95 or lower (or higher sometimes.. but not too high))
WORKS BEST WITH HIRES. FIX ON!!