Yes, other people have made ahegao models, but I wasn't satisfied with their flexibility. I decided to try to make an Ahegao face LoRA that wouldn't overpower the rest of the subject. It was VERY DIFFICULT. Over 50 failed attempts later, it's still not perfect. But I'm calling it good enough for now.
Read below for usage tips. If folks ask, I may add a writeup on the process of making it.
Usage:
Strength will vary a lot per model and subject. It the model needs more encouragement for men than women. I advise experimenting in this order:
-Start at full weight with no triggerwords.
-add "ahegao" to the prompt if you aren't getting an effect
-change to "ahegao face" if you still aren't getting an effect
-Adjust weight and CFG scale to taste if you're getting an effect but it's generating distortions/artifacts
Tips n' Tricks
-adding "pleasure" should push it more toward a smile and away from idiotic. Sometimes.
-"drool" is something it's trained to know, but it can't replicate it well. Try it out, or add it to the negative if you're getting lines coming off tongues.
-"piercings" may need to be in the negative prompt if it's putting stuff all over people's faces
Compatibility:
I trained this on F222, since it's an ingredient in a lot of popular photorealistic nsfw merges. It'll likely have issues with models that don't don't have that in their DNA.
Description
-base model: F222
-100 training images, 1800 class images
-18 training repeats, 1 reg repeat
-64 dim, 16 alpha,
-Unet: 6e-5, TEnc: 75e-6
-Polynomial scheduler, 1.75 power,
-noise offset=0.03
-batch: 6 * 3 gradient steps
-12 epochs
FAQ
Comments (12)
This Lora distorts too much in every case. Other models made by you work well. What makes this difference?
I think the difference is that human facial expressions are way more complex than poses. Like, our brains are very well-trained to recognize differences in faces between people, and discern meanings from super subtle changes in expression. Teaching a program to learn a new -expression- without learning the -face- was my goal here. I was able to make models that replicated the -same people- from the training images without much distortion. But the aim was to be able to make -any person- have that expression. (Which this model /can/ do, it just needs a lot more work in prompting for it to do it well. Also hires upscale helps.).
Or I just messed up something simple and hopefully someone more experienced will point it out to me.
I think this is the best ahegao model I've tried, maybe your settings or the model you're using isn't compatible.
@brandnewfakeuser No offense. Making differences in facial expressions without making any facial changes is as hard as making perfectly shaped hands in StableDiffusion. I hope that can happen in the future.
@Vissehwy i've wasted over 150 000 steps and around 10 restarts from scratch on unreleased lora where hands, fingers and toes is very important. Aand still no good result. Making small details neat is incredible difficult
"Over 50 failed attempts later, it's still not perfect"
Just leaves me wondering how many great loras you decided not to share because they weren't up to your standards.
looking good in inpaiting.Thank you.
@chesterskej How do you mask the subject for inpainting? Masking the Full face turns it into a different person. Selectively masking mouth area doesn't work well either. I'm using URPM inpaint model btw.
The best ahegao lora!Especially work well in realistic works.Thx!
looking forward to SDXL version if it's in the works!
One of the best loras on here
Do you mind sharing you training data?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.