An entirely imaginary South Indian woman.
Description
FAQ
Comments (7)
I wonder how SDXL version embeddings were trained? I faild in stable diffusion. I can only train V1.5 TIs.
Ah, I'd love to teach you.
This method automatically trains both 1.5 and SDXL simultaneously, and the entire training process takes less than 30 seconds.
First, you need the Embedding Merger extension.
Once installed, describe your subject in great detail. I do it in this order:
1: Hair colour and style
2: Skin tone
3: eye colour, size and shape
4: Mouth shape
5: Body type
6: Bust size
7: Age, name, nationality
8: Mixed with 2-4 celebrities that share some vague resemblance or has a feature you'd like to incorporate.
An example: Long voluminous platinum hair, tanned skin, pale blue eyes, pink lipstick, pouty full lips, very large expressive eyes, round cheeks, long eyelashes, prominent eyebrows, bombshell body, very large bust, 21 year old "Britt Lawrence" from Miami mixed with "Noemi Kovacs" mixed with "Helga Lovekaty"
If you have a 1.5 checkpoint loaded when doing this, the 1.5 embedding will be placed in the Embedding merge folder (in your regular embeddings folder) and the SDXL TI will be in a subfolder called sdxl. If you have an sdxl model loaded, the SDXL TI will be in the main embedding merger folder and the 1.5 will be in in a folder called sd1
@tvange365 Thank you very much! I got it, with my understanding, so actually we CAN NOT train Embedding directly (from photos inside our dadasheet) with-in SDXL (lots error will occur). What we can do is using Embedding Merge‘s function to convert these SD1.5 version embedding file to SDXL format. I'll try in my work, thanks again.
@AceNeo I'll write an article describing my process soon. Maybe later today.
@tvange365 Thanks a lot, 'cause I've been using the traditional way to generate embeddings through SD1.5's trainning procedure by processing with pre-selected images of individual persons (since they're not celebrities) inside my datasheet folders (with-in 200 steps normally). But this way it's not working in SDXL, maybe this method is out of date haha.
@AceNeo I'd say 200 steps is very, very low. I typically use 1500-3500 steps.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.