VRChat Furries

# VRChat Furries **Discord server** to see or share other generated images: [Discord Link](https://discord.gg/bC6U4GwbAt) ## 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