After spending the last 3 weeks and about ~100 attempts on trying to create this LoRA, I finally got a version I'm kinda happy with. Trained on 919 images. There is still a lot to improve but it can give good results.
The sample images are raw outputs (txt2img + Hires. fix). I used the embeddings UnrealisticDream, BadDream (by Lykon) for some examples. All examples made with RealisticVisionV51 (by SG_161222). CFG Scale of 7 works fine.
Main Keywords are:
tongue out (749 trainers)
ahegao face (407 trainers)
drooling (99 trainers)
wet tongue out (77 trainers)
wide open mouth (44 trainers)
Limitations:
Subjects further away from the camera don't really work that well
Bias towards nose piercing (In a lot of training images)
Bias towards animal ears when using "ahegao face"
Keyword "wide open mouth" sometimes produces terrifying outputs
Keywords "tongue out to the side" and "licking lips" don't work (Not enough training images)
Might not work for males. Didn't test
There would still be a lot to improve. I used a lot of medium quality (and some kinda low quality) images. But its hard to find more high quality images. Also I wasn't able to train above 512x512 since my RTX 3060 Ti couldn't handle more. That might could have improved quality also. I would love to make a SDXL Version but i'm too poor for the required hardware, haha.
Note: The example images have the wrong LoRA name in the prompt. Tongues_V62-000012 was my original name before changing it to TonguesAndAhegaoV1.
I appreciate your feedback!
Description
FAQ
Comments (20)
Looking good. If you need help with training hit me up on discord :D
Thanks for the offer. I might get back to you after my vacation!
Could I generate images of other ethnicities besides Caucasian if I wanted? Were there training images of that?
All ethnicites were included in the training images. But most of them were caucasian. Best would be to try it out yourself.
I was using this while inpainting faces, and i got some interesting results from enabling it but putting ahegao face in the negative prompt
I didn't try it on inpainting. What results did you get?
@Tepcaly usually models have certain looks for faces that show up over and over, it just gave me different faces than i normally get
I see. Maybe because it was trained on ~1k handpicked images
What did you put in the prompt to get them to have smiles/playful expressions while sticking out their tongue?
You can see all the prompts i used for the example images if you click on them
Your work is very outstanding, achieving the customized functional vision I hoped for. Unfortunately, I spent the vast majority of my time customizing the description of my training subjects, but due to limitations in the EMBD format functionality, the customized ideas in my work were not implemented. I will share my customized ideas and templates with you, hoping to provide inspiration for your future optimization iterations.
TEMPLATE【close up of/a woman sticking out her tongue,there is a thick/natural/thin white coating on her tongue and the color of her tongue body is pink/light red /white/red/light purple/ brown,with/without obvious saliva adhesion on her tongue,her tongue fully/partially/slightly/ extended forwards/downwards and flattened/narrowed with her lips pressed tightly together/with her mouth widely/moderately/slightly open, her upper teeth/lower teeth fully/partially/slightly/ bared and her teeth almost gnawed on her tongue,while/her teeth not bared and /her jaw remained relaxed/her jaw jutting outwards】
I spent 3 months using this template to describe 300 pictures of my training set (including design and translation, I am not a native English speaker). Your training set is up to 900 pieces, so it will be a considerable workload.
Thank you! I also spent a huge amount of time in making custom descriptions only to find out later, that just a short trigger word works best. I also created a template similar to yours before but it didn't work out. In the end I just created folders for each concept or multiple concepts. For example I have those folders: "tongue out", "ahegao face, tongue out", "ahegao face, drooling, wet tongue out" ...
Some concepts, I tried to implement, didn't work at all. I realized the more images I had, the better the concept worked. Maybe it would be completely different when finetuning. But my graphics card is not good enough for it. Also I'm currently working on a game, therefore my time is limited.
@Tepcaly Creating a detailed custom description for each photo may be a clumsy approach. I am not very familiar with Lora's training process. If we could separate several sub concepts and bind specific trigger words, it would be very convenient and effective, but unfortunately EMB cannot achieve it.
What prompts do you use to have the camera tilted a bit to the side instead of full front view? I can´t seem to do that no matter what prompts I use.
Try keywords like "profile" or "from side" in the prompt. Maybe even with weight. Or "from front" in the negatives. It's trial and error, though. You can see my prompts used on the example images when you click through them.
Hey, do you mind sharing the training data? I would like to give it a try and train an sdxl lora for ahegaos
Sorry but I don't share the training data. I would like to train a SDXL Lora of this myself at some point, though. Unfortunately, I don't have the financial resources for the hardware right now
@Kasmir Sad, but ok. I have quite a large set of training data myself, but i thought i could maybe get some more diversity with your data aswell. Thanks for the reply anyway :)
For some reason whatever I do I cant get it to ahegao face. Is there any other extension that you have to make it better?
I found that sometimes when i use other Loras at the same time, it can interfere with the Lora i intend to use.
Try exporting your image to inpaint and modify it there using your Lora for Ahegao, works fine with me...
(Oh and also your prompt can be problematic to...)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.