Textual Inversion of Cara Delevingne.
I wasnt happy how SD massacred my girl, so I tried to fix that.
This is my first attempt at doing textual inversion, most of the time I have no clue what Im doing but it turned out well I guess...
Trigger word is caradelev1ngne_ti for V3!
it is now happily cooked but lowering the weight to 0.98 sometimes does help.
v1 was trained with 7 images for 800 steps
v2 was trained with 8 images for 3000 steps
v3 was trained with 21 images for 4600 steps
I tested the TI with some of the most famous prompts from this site, mostly got usable results. Will be improving over time, as I learn more.
Oh and the training data had necklaces and chokers in them so you might want to put that into negative if it generates too much stuff around the neck lol
Description
retrained with better images and more steps. i checked the pt file with inspect_embedding_training.py and it sits comfortably at 0.1915
hope to see some nice images from you :)
FAQ
Comments (11)
I checked the vector strength and it's 0.08, so it's under-cooked by a large margin. if this is 800 steps, run it to at least 2000 and check that one and maybe an 1800 step version as well.
Great insight thanks, I'm already cooking a second version. Test higher step counts aswell.
Here is the script I use to check vector strength - https://github.com/Zyin055/Inspect-Embedding-Training
@Bozack3000 thanks man appreciated. I learned that 0.2+ is overbaked but couldnt figure out when that started. do you know a good value to stop? e.g. 0.15 or something?
You can save an embedding as often as you like, so set that to every 100 steps and you should come super close.
I just ran a training that saved every 500 steps (using a 0.005 learning rate) and one of them hit .2061 so you should have no problem.
Wanted to add - at such a low vector strength, your TI is most likely being influenced significantly by the info for Carla already in the model you're using.
I noticed in some models it would not invoke her at all or sometimes even a male lol. I was scared of overbaking but will try some more later
I love her so hard from Valerian. Thanks for making this. :)
I must be missing something obvious but I can't get it to work. I am using the SD v1-5-pruned.ckpt model with 1111 (which I git pulled a couple of days ago).
This TI is in my TI folder and shows when I click the "Show/Hide extra networks" in TextToImage.
The console even logs
"Textual inversion embeddings loaded(2): caradelevingne_ti, MyOwnFace"
I do use "a portrait of caradelevingne_ti" or something like that in my prompt, but it's just some random person.
So, I must be doing something wrong. Any idea?
I'm new to this site and can't figure out how to tag more than one resource, but thanks for the CD model! Here are some shots of her combined with more anime looks: https://civitai.com/posts/166595
great pics!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.