I couldn't find any LoRA models that had the specific makeup look I was looking for, in particular "Bimbo makeup", so I've decided to train my own.
Results seem to happen around 0.5-1.0 weight from what I've played around with.
I haven't done extensive testing on it, but from my preliminary testing it seems to have gotten the idea.
I'm brand new to training LoRA models, and this was the first to actually come out looking alright. My apologies if my inexperience shines through and makes this an actually terrible model. I had a specific use case for this going in, and I'm mainly sharing as a bit of a formality for people that are also looking for something similar.
By all means let me know what issues you have with it, but I probably won't be able to fix them unless you yourself have a suggestion on how to do it. I don't really know what I'm doing ¯\_(ツ)_/¯ and I'm kinda just throwing things at the wall and seeing what works.
Description
FAQ
Comments (16)
the bimbos are finally arriving!
One candle if by land, two if by sea. God forbid, they arrive from the air.
How easy is it to change it? Can you prompt colors easily with it? Is it just realistic or drawn too?
The lipstick is kinda hard to change, just due to the fact that a lot of the girls in the dataset have pretty much the same color lipstick, but I've been able to push the eye makeup colors around alright enough.
I've seen people use it on drawn models as well and it seems to work fine. Apparently better than if you had just prompted "makeup" without the LoRA. I'm still completely new at this, so I can't really pretend like this is anything more than a fluke. Basically, your mileage may vary. If it works, that's great, but if not, I've got no clue what to do to fix it.
@anon4324 (I think) if you wanted to make something very versatile, or very specific, you could inpaint images to make your own custom dataset with those things, I.e. specific coloured lipstick/eyeshadow etc. You'd then train using that faked dataset so you can teach it each individual one. You could teach it to make just black lipstick/eyeshadow for example, make a new dataset with that and teach it to be "gothic makeup".
I haven't worked up the courage to make a lora yet, but that's my understanding anyway!
@BDinc Yeah that's definitely an option. Especially if I decide to upgrade this at some point. Augmenting the dataset with different colors via inpainting on real photos, or from generated images if they're decent enough. I guess theoretically using the LoRA on existing images, sending them through img2img is also an option. Looks like this is more of a first step than what I had originally thought.
I'm also getting a bit better at creating LoRA as well. Something pretty special should be coming out soon that I hope people will like. Lord knows I've been putting way too much time and effort into it.
@anon4324 I definitely want to try it myself at some point, but at the moment getting a decent enough dataset and doing all the tagging is the hardest part for me! Well done for the results you got!
Which model was used with this great LORA, to make the renders ? From the renders prompts it has the hash cc6cb27103, but cannot be found here on CAI.
What a time to be alive.
Truly this is one of the Loras of all time.
curious what if any trigger words you have here? trying to use in a prompt and not getting great results, been playing with weights and such. from your example prompts it's hard to decide what actually gets this thing going. thanks, would really like to use this for makeup in prompts.
I don't think it has any trigger words.
To be fair, I made this when I was very inexperienced with making Lora. Honestly the results I got were probably a lucky fluke more than anything, and I only really needed it to work for 1-2 images since it was made for a project I was working on.
I mainly only uploaded it as a courtesy to people looking for something similar, but I had no idea it would actually have as much downloads as it does. I'm pretty sure it's badly made, but I don't really feel obligated to make a proper version any time soon. Kinda just bored with Image Generations right now. Maybe if there's demand for it, I'll do a proper version in the future now that I've got quite a bit of experience under my belt.
@anon4324 If you still have your training images and the text files with tags for them, you could look at what some of them got tagged with. Words like "mascara", "eyelashes", "lipstick", "eyeshadow", and "makeup" should all trigger it and use your lora to get samples of those tags. =)
@Jenika Yeah that would definitely be the plan for a v2.
I took a look at the training data, and it's not exactly "tagged" but more something like this
"a woman with long blonde hair and blue eyeshadow with pink and purple makeup on her face and her hand on her shoulder"
From what I understood at the time, that was how you were supposed to caption realistic images, but perhaps that's not the case anymore. I'm definitely more familiar with animated/anime style loras, and I've gotten quite skilled at that. I think a v2 is going to be a project I tackle soon (once I clear work off my plate), and I'll do what I can to leverage what I've learned to put out a realistic and anime model.