[Textual Inversion Embedding]
Trainer is not me.
IF IT IS YOURS COMMENT WE CAN TALK.
The OG Korean doll-aesthetic mod for Stable Diffusion.
» I like to use "Ulzzang-6500" embeddings (I used that for all sample images).This is my type....I couldn't avoid generating with this.... «
-Tasuku, who made ChilloutMix
Works best for semirealism and photorealism, but most likely only on SD 1.x (tested on 1.4 and 1.5). Is often used in combination with the ChilloutMix and aom2h-basilmix. No bias in terms of eye and hair color and also not mimicking an individual person! Also handles angles well, you can throw pretty much anything at it and use it for inpainting of the face as well. There is a significant bias towards pale skin, as a heads up.
How to use
Simply download and put into your /embeddings folder in the Automatic1111 folder. Then use 'ulzzang-6500' in your prompt. Reduce or increase the weight of the embedding by using this syntax:
(ulzzang-6500:1.2) increases the weight, or values below 1 to decrease (such as 0.8)
I recommend putting ulzzang-6500 towards the end of the prompt, it's strong enough.
Inpainting
Another useful approach with this model is to inpaint on top of already existing images that were generated without the embedding if you want to see how they look with the ulzzang face.
In Automatic1111, on the img2img tab -> Inpaint tab paint over the face (you can include the hair as well). Then describe that specific masked area in the prompt, which means that you should remove "bad hands" or "perfect body" from the prompt.
Select "Inpaint only masked area" and set the steps to around 70 (this is important, inpainting and img2img require higher step counts). Set the denoising strength to 0.4 to 0.75. The more closely the existing face resembles the ulzzang aesthetic the lower you can set this value. Set the resolution to 512x512, this is plenty for the face region.
Note on the concept
The embedding is neither intended to provoke childish features (the 'doll-like' refers to an uncanny level of supposed beauty standards, like a slim nose, aegyo sal and tear-drop chin) nor does it have a bias towards them from my testing. If the face in a generation should appear too young that's a bias from the latent model and can be countered by adding 'loli' and 'child' to the negative prompt.
For the LORA fans there is a similar model as well named Korean doll.
V1 demo image by yatagarasu, second one by me, third one by Bloodsuga
V1.1 images by author.
I upload it here Because I want to use it.