This LoRa will allow you to use your favorite character with it without altering its face.
(Minimal alteration can technically happen on eyebrows, mouth and eyes)
I've made huge progress in training Loras very recently so I'm pleased to say this one is excellent by my standards.
You can even combine it with my other flux Loras, no problem !
Trigger word : Ahegao - Please use also "tongue out", seems decisive
Other useful words (reinforcements) : crossed eyes, drooling, dr00ling, s1de (for side view)
Weights : I think 1 is fine, it definitely doesn't alter the face, but then if you are using more than 2 loras (your character and the ahegao), you might want to play with weights. Not sure it's good to go lower than 0.6 though. Haven't tried more than 1.2 either. Interested in your feedback on this.
Known issues (from my little time testing) :
Tongue can sometimes come out a bit weird
Drooling tongue, unfortunately, despite my best effort, very rarely comes out. Instead, it sometimes messes the tongue. Lesson learned, I will really stick as much as possible to a single concept in my future Loras, and just make more versions if doing variations. If you got a drooling tongue, good for you, the ahegao gods are with you :-D
Hairstyle can be altered (quite rarely) by the lora as the dataset pics were a lot of e-girls and cosplayers. Actually for me it's almost an improvement rather than an issue, lol. But just saying it's best to include the hairstyle you want in your prompt just to be sure. 90% of the time it's ok though.
For the same reason, your character can get a bit of mascara and eyeliner. No big deal.
Crossed eyes don't come always, as well as rolled backwards eyes which are rare. But they can, and do come. Add some prompts of your own, be patient, make many generations, etc
It's possible the Loras drives you more to close-up since it was trained on faces. YMMV.
This one I don't really need to give you prompt examples, it's easy ;-)
Have fun liberating the power of sweet ahegaos in all your characters !
Description
FAQ
Comments (18)
Oh man, Thanks a lot for this! your work on Non Face Altering LoRAs are amazing! Keep tweaking and fine-tuning them! Looking forward for more updates and releases!
Thanks :-)
will it just cross the eyes or is the open mouth hard coded in just like every other one?
EDIT: yeah it's just like the others, gives them all heavy eyeshadow and makes them look like e-girls.
I still appreciate the free Lora though.
I'm doing these loras for me first, and this doesn't bother me. I think there are several things that don't cross your mind :
- This is an Ahegao lora not a "crossed eyes" Lora specifically
- I found out it is way more challenging to train multiples concepts rather than a single one. I've tried that with my POV blowjob lora with sucking and licking and it wasn't good.
- I could make a second version to exclude eyeliners but this is a lot more work than you think
- Try to find at least 30 - 40 quality pictures of girls doing ahegao that are NOT e-girls or cosplayers and don't have eyeliners. Good luck with that.
- If I didn't write out in the description the dataset was using egirls I'm sure you wouldn't even have known
Thanks for trying it out anyway. Btw, your account is more than one year old and you haven't posted a single picture or contributed anything on civitai, just saying...
@valentinkognito365 No problem! Sorry if I came across as rude. I will still play around with it, because it's a new LORA!!!! :)
@cdhasshman no worries. I kinda get your point though, it would be nice to have more control over the lora by having crossed eyes or not. The thing is i'm realizing its a distilled flux limitation. Distilled flux is just plain bad with prompts when using loras. I suggest you try dedistilled and negative prompts for eyeliners it might work. Meanwhile, i was more trying to have a non face altering ahegao ASAP rather than a perfect attempt, because I hate LoRas altering my beloved characters. At some point I might come back on it to improve it, but it's always that or making new ones like positions etc. Dilemma :-D Anyway, I like people making comments rather than everyone being silent so thx for your previous comment anyway, it shows me interest on the crossed eyes
maybe it being unable to drool is because of flux? I saw a drool lora where they put "translucent drool" in the front when prompting. I actually made an ahegao lora a few days ago and captioned images with cum on face as "white translucent liquid" even though only 3-4 out of like 60 or so images had it and obviously it didn't work the best but it still popped up. On the other hand images I tagged with drool didn't work as well despite having more images than that.. I made an image on civit ai generator where I was able to get drool with this lora on the first try. So I just assume that putting translucent while prompting and/or captioning is what works best. If I reuse my ahegao lora and instead put translucent drool instead of just drool while generating an image maybe I'll actually get drool.
Thanks for this advice, i might do that. But I'm waiting a bit to see the progress of people doing training tests with dedistilled. I believe it could increase dramatically the prompt adherence of the lora and allow concepts such as the drooling to come out. Dedistilled is pretty much my god-mode model since I've tried it.
I saw you posted one, well done !
Well done! This is shockingly good. I've tried many ahegao (and other facial expression) LoRAs and nothing has come close to the consistency of this one; ahegao or otherwise. And it really does not affect characters' features. Thanks so much for sharing.
Thanks, I appreciate your comment :-) Good news is I have ideas to improve it in the future
@valentinkognito365 Awesome - I look forward to the update. I've tested it a fair amount now and the only issues I experience are the ones you've described. The only thing I have to add is that the resulting image is consistently soft/slightly blurred.
As someone who definitely hasn't been able to figure out how to train a successful non-character LoRA, I'd be really interested to learn about your workflow and config. I'm 100% with you in your frustration with bad LoRAs wasting Flux's potential so I'd be happy to train and share some useful ones.
@272727 interesting, i did not notice the blurring/softness. Will look closer if related to the lora or not. About how I do that I plan to do articles about it so motivated folks like you can do it
@valentinkognito365 One can easily overcome it by adjusting one's workflow to bring back sharpness and detail after the initial generation.
My first suspicion is that it's a result of the main area of the camera's focus (the face) being masked out in the dataset. So perhaps the model is inferring that the ahegao concept means slightly out of focus, which presumably the majority of the unmasked area of the photos in the dataset would be.
If I'm correct, training with some non-closeup shots in which some of the unmasked area is in perfect focus might help? Such a minor issue though - I'm only really writing this because I find it interesting; I don't think doing anything about it is warranted. The main thing is that it does its job really, really well. I'll test your other face-focused LoRAs over the weekend and let you know if I experience the same thing with them.
@272727 I really doubt it's due to this focus theory. I try not to take blurry samples but then sometimes I take relatively bad quality ones and fix them using a dedicted software for that. It could be also overtraining of the model, the quality can lessen if it's trained too much. Anyway, next step for this model is clearly a retrain using the dedistilled model. I'm sure this way we could get much more easily crossed eyes or drooling tongues through prompting
I tested this Ahegao model and the results are fantastic! The facial expressions it generates are spot-on, with clear details, exactly what I was looking for. Thanks for sharing/creating it!
Hello! Amazingly good lora! What's your workflow to non face altering? If you want to make it public, of course. I tried myself to make one with not so good results. I blurred faces and captioned the blur (trying to show it wasn't part of the lora), but now the 90% of my generations with that lora have blurred faces.
"Weights : I think 1 is fine, it definitely doesn't alter the face, but then if you are using more than 2 loras (your character and the ahegao), you might want to play with weights."
so.... if you aren't using a character Lora, and you're only using this Lora by itself..... how would you know if its altering the face or not? how would you know what face its gonna generate..?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.