The eeveelution family LoRA
Also include "sky forme shaymin", "absol"
In v3 version add "kirara (from inuyasha)", "teto (from ghibli-nausicaa)"
If you want to try "feral" type, use (quadruped feral) prompt, and set higher weight.
v1/v2: base on Yiffy-e18
v3-YiffyMix : base on YiffyMix v30
v3-FluffyRock : base on FluffyRock-e27-o12
This is a LoRA model, you will need new version Automatic's 1111 SD-WebUI or additional-networks extension
Use e621 tags, no underscore, Stop at CLIP layers=1
Best setting = [network strength:0.5~0.8]
Description
Demo images generated on YiffyMix + lora (network strength:0.78)
Stop at CLIP layers=1
FAQ
Comments (8)
Thanks for this! Can you do lora of furry themed artists?
kinda new here where can I check what the format for yiffy-e18(and therefore this model) is?
https://huggingface.co/ckpt/yiffy-e18/tree/main , ckpt model, use 84000 VAE, Emarraming V1.
These seem very reliant on putting A LOT of triggers in to get anything near a decent image. Is this just the case with furry models? Or the Lora itself? Also seems very limited on poses.
In my experience less can be more. Lots of tags work fine on certain models, but there can be a lot of infighting with others' tags, meaning that you'll have to find the max amount of lora-specific tags you can use along with the right lora/character weights. I've seen it on mine and others so I'm pretty sure that's just lora's in general and how this dataset was tagged.
@PseudoSu Right, but most Lora's should to a degree be able to kick something remotely passable just from them being triggered. "high quality, hires, 1girl, solo, <lora triggers>" is usually the testing prefix for when I test out any lora (raytracing, realistic for realistic models) and then usually I see what it's going to produce with that being it's simplest form. If you want to hone in on details, that's when you should be able start spamming in more tags. When I tired this a month ago, the results were horrendous, until I started filling in more and more tags and even then they weren't great.
@BDinc I haven't touched the original in a while but I will say that this seems to be way more unpredictable, keeps giving me human hybrids on models it used to work on. I got really good results with LawlasYiffyMix in the past, but yeah the tags counts were through the roof. To be fair Yiffy18 sucks as a base in general, I've never gotten anything out of it beyond a low-quality blender render and mixes off of it are either pretty solid but mid or really good but hit-or-miss. A NAI or AOM based release would be better. Until then, Im'ma see what works since I think this version is really undercooked as well.
Edit: I've done more experiments and for some reason only my old prompts (with a LOT of what's in the images cut out) and as long as you have EasyNegative it works pretty well. Not stellar, but good enough. Don't use prose, just single words. Cut out all the fancy style stuff and use what your model works with. Maybe I'll take a crack at it someday myself, but this lora works well enough with some tweaking so it's unlikely.
@PseudoChmod There are models with baked-in Eeveelution support.