VRChat Furries

Discord server to see or share other generated images: Discord Link

New version : V4.2

Trained for the Yiffymix v3.3 model (https://civitai.com/models/3671?modelVersionId=143769)

YOU NEED TO REDUCE THE STRENGTH OF THE MODEL TO <0.8

Work better with CFG Scale between 3 and 6
Works with clip skip 1 and 2

Contains :

Rexouium (keyword: rexouium)
Avali with kitavali and davali (keywords: avali, kitavali or avali, davali)
Nardoragon (keyword: nardo)
Deuzear (keyword: deuzear)
Wickerbeast (keyword: wickerbeast)
Protogen (keyword: protogen)
Taidum (keyword: taidum)
Awtter (keyword: awtter)

Also works well with anime boy and anime girl

LoRAs working well with VRCFurry

Gacha splash LORA v3.1 and v4 : painting style + epic background and arrangements + improves faces (biased towards shorts + thinner body + centered subject) https://civitai.com/models/13090?modelVersionId=26536
HD Helper : avoid duplicate faces on high resolution pictures (https://civitai.com/models/110071?modelVersionId=118644)
Fluffyrock Quality Tags : slightly improve characters + more likely to generate furry instead of humans (https://civitai.com/models/127533?modelVersionId=151790)
Micro Waist : self descriptive (https://civitai.com/models/4602/waist-slider-microwaist)

Negative Embedding working well with VRCFurry

I tried a bunch of different negative embedding and those are the ones that seems to give the best results

Boring_e621 (https://civitai.com/models/87781/boringe621-negative-embedding-improve-image-quality)
veryBadImageNegative (https://civitai.com/models/11772/verybadimagenegative)
BadDream (https://civitai.com/models/72437?modelVersionId=77169)
EasyNegative (https://civitai.com/models/7808?modelVersionId=9208)

I recommend to use all of them à the same time.

Keywords

Improvers :

good lights : makes volumes pops more
good colors : better color choices
good composition : more assertive composition choices
good face : faces I prefer
bestart : used this keyword on the best art of my dataset (few example so it is pretty weak/random)
aestheticcartoon : more assertive colors and texture choices
aestheticmuscles : targets no fat, no steroid or bara look (sometimes the position gets a bit stiff)
musculegs : to counter act the strong biasis of most models toward thick upper body and thin legs

Worsener : (to use in the negative prompt)

bad face, ugly, dumb face : used in dataset, pretty self descriptive
bad hands : used in dataset, pretty self descriptive
bad muscles : unaesthetic muscle insertions, poor definition or disproportionned
bad abs : same but specific to abs
ghbelly : used in the dataset for muscular bodies with gh/steroid bubble gut, and distended abs
thin legs : self descriptive
small eyes : self descriptive (but pretty weak as there are very few examples in the dataset)

Non vrcfurry keywords but that I like to use with this model

gorgeous eyes : improve eyes
extremely detailed furs, 4k, 8k : add details but often biased towards small eyes

Styles (keywords) :

comic
thick lines
photography
render
painting
swaggy
draft
hardstyle
detailed style

Composition (keywords):

rim light
dynamic scene
complex scene
bicolor
monocolor
morning

Body / Face (keywords):

cute
teenager
( - / slightly/ very / extremely) muscular
ripped muscles
(spiky / wavy / feather) hairs
relaxed
vore

Clothes (keywords):

( - / tight) shirt
short
legging
hoodie
bandana
croppedhoodie
muscles through clothes

Position (keywords):

side view
3/4 view, three-quarter view
back view
looking back
looking down
looking away
defying look
dynamic position
seated / sitting
crouching
lying on
leaning on
facing down

Suggestive / NSFW (keywords):

dick through clothes
dick sticking out
autosucking
jacko pose
grabbing pec
first person view

Others (keywords):

hmoma : furry with human (at least 2 persons)

Tips & tricks:

Improve avali look by adding those keywords: cute, big eyes, feathers, black eyes
Get better faces by adding these keywords: close up portrait
Sometimes you can get better faces by changing hairs / furs colors

Good positive prompt keywords :

Kind of a meta I found across all my models

good lights, good colors, good composition, monocolor, bicolor, good face, (cute:0.7), teenager,

good lights, good colors, good composition, are keywords I added on the best artworks in the dataset
monocolor, bicolor, are often associated to higher quality / artistic artworks in the dataset with interesting color management
(cute:0.7), teenager, cute put emphasis on eyes and make them much better and teenager couteract the "chibi"/"kid" look it could create while still keeping a cute look

and some bonus keywords that can make it look even better, even if it depends a lot on your tastes ^^

masterpiece, extremely detailed furs, 8k, soft fur, perfect composition, detailed shading, dynamic pattern with accents, detailed style, soft,

Good negative prompt keywords :

(neck fluff, fluffy neck,:1.2), lores, bad anatomy, warped, bad hands, BadDream, easynegative, boring_e621_v4, (ghbelly, bad face, ugly, dumb face, thin legs, flat hairs, bad abs, fluffy neck, neck fluff, fluff on the neck:1.2), (anime boy, human, human ear:1.4), verybadimagenegative_v1.3

There seems to be a bias towards females, so if you want to generate males, use:

female, woman, girl, 1girl,

If the shirts are cropped / always lifted, use:

crop top, cropped shirt, bra, shirt lift,

Planned for next version (V5):

Better tail control
Reference sheet addition ?
better dataset captionning

VRChat Furries - v3.1