Pure Eros Face - PureErosFace_v1
Mots-clés et tags associés
Prompts recommandés
[:(detailed face:1.2):0.2],shiny eyes,looking at viewer
(half_body:1.1), (short hair:1.1), PureErosFace_v1, blurry background
Prompts négatifs recommandés
long hair,full body,disabled body,DeepNegative
nsfw, nude, nipples, plant, full_body, bad_prompt_version2, day, sunlight, long hair, extra limbs, extra arms, extra hands, extra fingers, extra legs, extra digit, deformed limbs, deformed arms, deformed hands, deformed fingers, deformed legs, deformed digit, malformed limbs, malformed arms, malformed hands, malformed fingers, malformed legs, malformed digit, fused limbs, fused arms, fused hands, fused fingers, fused legs, fused digit, mutated limbs, mutated arms, mutated hands, mutated fingers, mutated legs, mutated digit, mutilated limbs, mutilated arms, mutilated hands, mutilated fingers, mutilated legs, mutilated digit, fewer limbs, fewer arms, fewer hands, fewer fingers, fewer legs, fewer digit, missing fingers, disconnected limbs, disconnected arms, missing arms, disconnected hands, disconnected fingers, disconnected legs, disconnected digit, missing limbs, child, missing hands, missing legs, missing digit, poorly drawn limbs, poorly drawn arms, poorly drawn hands, poorly drawn fingers, poorly drawn legs, poorly drawn digit
Paramètres recommandés
samplers
steps
cfg
resolution
Paramètres haute résolution recommandés
upscaler
upscale
steps
denoising strength
Conseils
When drawing high-quality faces, do not use the "detail face" tag at 0 steps to avoid deviation from the original embedding semantics.
Add "shiny eyes, looking at viewer" prompts to open eyes if generated faces have closed eyes.
Use negative prompts like "long hair" to fix outfit quality issues due to hair length biases in training data.
Avoid full body prompts as the dataset is mainly close-up images; this might affect quality.
Add "disabled body" if facial or body deformation occurs.
Try using the DeepNegative embedding alongside PureErosFace for better results.
Points forts de la version
Training Config:
initialization text: ulzzang face,kpop idol
number of vectors per token: 1
lr: 0.05:10, 0.02:20, 0.01:60, 0.005:200, 0.002:500, 0.001:3000, 0.0005
(Here is not the complete lr formula, because I manually selected several iterations as branches during training, but I did not record it completely...here is just a part)
batch size: 10
images: 100+ images (close-up portrait girl face)
size: 512x512
steps: 1000
This embedding will generate good-looking girl faces, close concept of kpop idols or Instagram girls. For SD1.5 based models.
what means "Pure Eros"?
> Pure Eros is a simple translation of the Chinese word "纯欲", which is a popular memes in the Chinese internet, the english words close to the semantics of this word are "ulzzang face", "korean idol face", so obviously this embedding will be close to the Asian aesthetic.🤗
Recommended config:
1. prompt
[:(detailed face:1.2):0.2]: When drawing high-quality faces, do not use the "detail face" tag at 0 steps, otherwise it may lead to deviation from the original semantics of embedding
shiny eyes, looking at viewer: If your generated faces often close their eyes, adding "shiny eyes, looking at viewer" will open them instantly
2. negative prompt:
long hair: add it to fix some errors caused by training data bias...(although it does not always affect the output, my experiments found that it seems that the length of the hair in a certain scene will affect the quality of the clothes. In short, the relationship is very subtle, you can try to control the length of the hair to change the outfit quality)
full body: Most of the data set is close-up, so it doesn't perform well on the "full body" scene, I'm trying to fix it in the next version
disabled body: If a deformed body structure is produced, please add it
DeepNegative: A negative embedding I released can be used with PureErosFace very well, it is recommended to try
Recommended model:
It is best to use a model with a realistic style. If it is an anime-style model, it may output a mask-like face.
1. LOFI
merged model I released, works fine with this embedding
https://civitai.com/models/9052/lofi
examples https://imgur.com/a/nFRE1Z4.jpg
2. izumi
Most of my example images are generated by izumi:
https://civitai.com/models/1364/izumi
examples and test: https://imgur.com/ZirqVWK.jpg
3. F222
Also produced a good-looking face, although I only experimented with a few, but they all look good for work
https://civitai.com/models/1188/f222
examples and test: https://imgur.com/ixyVXTP.jpg
4. HassanBlend
https://civitai.com/models/1173/hassanblend-1512-and-previous-versions
examples and test: https://imgur.com/no066sF.jpg
Backup
Contributeur
Détails du modèle
Type de modèle
Modèle de base
Version du modèle
Hash du modèle
Mots entraînés
Créateur
Discussion
Veuillez vous log in pour laisser un commentaire.
















