In Anithing and Abyss series of animation style models where the existing LoRA would create an image without a separate appearance prompt, the face style was fixed.
To improve this, the female face normalization data was refined again so that various types of facial expressions could be generated.
Along with this, the phenomenon that the hairstyle was fixed with a side parted hair and the phenomenon that the hair accessories were overexpressed has been improved.
In order to correct this phenomenon, we manually modified all the hair parts of the existing dataset photos and images.
Now you can freely express your face style, expression, and hairstyle.
In addition, the newly shared LoRA doubled the learning rate compared to the previous LoRA, and deliberately overfitted it.
This is a part that has been adjusted to be advantageous for combination by lowering the weight when used in combination with other models or LoRA.
The advantage of low weight is that it can suppress image breakage caused by model characteristics or LoRA combinations.
Even if various prompts are applied, it has the effect of responding to stable image generation.
Recommended trigger: hanbok
Recommended weight: 0.3 ~ 0.8 (The default value of 1 can be used, but chromatic aberration, such as image breakage, becomes more pronounced.)
In Anithing and Abyss series of animation style models where the existing LoRA would create an image without a separate appearance prompt, the face style was fixed.
To improve this, the female face normalization data was refined again so that various types of facial expressions could be generated.
Along with this, the phenomenon that the hairstyle was fixed with a side parted hair and the phenomenon that the hair accessories were overexpressed has been improved.
In order to correct this phenomenon, we manually modified all the hair parts of the existing dataset photos and images.
Now you can freely express your face style, expression, and hairstyle.
In addition, the newly shared LoRA doubled the learning rate compared to the previous LoRA, and deliberately overfitted it.
This is a part that has been adjusted to be advantageous for combination by lowering the weight when used in combination with other models or LoRA.
The advantage of low weight is that it can suppress image breakage caused by model characteristics or LoRA combinations.
Even if various prompts are applied, it has the effect of responding to stable image generation.
Recommended trigger: hanbok
Recommended weight: 0.3 ~ 0.8 (The default value of 1 can be used, but chromatic aberration, such as image breakage, becomes more pronounced.)