面倒くさいプロンプトをアレコレ指定しなくても、ネガティブに１つ指定すればそれなりの結果になります。
顔のアップになってくるとリアルさが弱くなるので強度やCFGで調整してください。
背景には弱いです。あと、脱ぎやすいかもしれません。
DPM++ 2M Karras CFG:6～7
2024.04.25update
顔のアップも少しリアルになったと思います。
裸だと日焼け跡が出やすいかもしれません。
ポジティブにasian faceを指定するとよりアジア系の顔立ちに、
ネガティブにasian faceを指定すると西洋系の顔立ちになります。
--
Even if you don't specify all the troublesome prompts, you can get a certain result by specifying one negative prompt.
As the face becomes closer, the realism becomes weaker, so please adjust the intensity and CFG.
The background is weak. Also, it might be easier to take off.
2024.04.25update
I think the close-up of the face is also a little more realistic.
If a woman is naked, she may be more likely to get tan lines.
If you specify asian face in the positive prompt, the face will look more Asian,
If you specify asian face in the negative prompt, it will give you Western facial features.
Description
more realistic
FAQ
Comments (5)
I prefer without. The embedding makes the faces look all the same. Just like the face you always see with 1.5 models. No idea why we're not training these models to learn a million different faces.
Thank you for your comment.
If there is a model that allows you to easily create attractive and beautiful women with a wide variety of features, I would love to have one too!
However, the purpose of this EMBEDDING is to combine multiple tokens into a single token called pony_negative, so it seems a little different from what you are looking for.
It saves you time by eliminating the need to enter 30 to 40 tokens in the quality specification. (The first image shows the result with less than 20 tokens.)
Use this as a starting point and adjust to your liking.
If you are able to adjust the prompt to your liking, including the quality specification, you may not need this EMBEDDING.
By the way, I still plan to adjust both this EMBEDDING and PianoMix.
Well for me currently the problem is more toward the models than the embeddings, i mostly use everclearPNYByZovya, and when i try to make a black female i always get the same face over and over
@aubepinvial634 I see, it depends on the model and the prompt, but this EMBEDDING may have been influenced by the fact that I made it closer to my preferences based on PianoMix.
It may be good to use it with a weaker setting of around 0.5, but the rest is a matter of preference...
As you may know, if you randomly specify a female name at the beginning of the prompt, it will produce some variation.
This may depend on the checkpoint you use it with. For my the models I tend to use this is by far better to anythings else to get reliable good output. I wouldn't want to miss this. Kudos to the creator!