UPDATE: Currently working on an SDXL version of this Lora. Will make it available soon so I can get some feedback and work out any issues.
Puckered, pouting, kissy face, whatever you may call it. This LoRA should help push the face in the right direction. Check out the examples to see how different strengths behave with different models.
Testing was done with ComfyUI which apparently handles weights a little differently, so you'll need to play with the strength a little. I've found a weight around 0.7-0.8 works well. If you happen to use ComfyUI, remember you can copy the workflow from the preview images.
I'm still new to creating LoRA's so I don't know if the trigger word is needed. Seemed to work fine regardless of what word I used to describe the lips, but let me know if you have issues with it. If in doubt, "pouting" should be the trigger.
Description
FAQ
Comments (14)
I've been waiting for this a long time
Glad to hear it'll be of use to someone! If you happen to have any other ideas for Loras I'd be glad to try them. The whole process is rather cathartic.
Works well with both real and animated checkpoints. Though it does have the tendency to create stern or angry eyes / eyebrows depending on the setup, nothing an angry negative prompt can't fix
I found something similar during my testing. Some checkpoints will lean towards angry/stern if you use the word pouting. Swapping it out for puckering should help.
@OcularPatdown I'll give that a shot, thank you! Regardless, 5/5
If, in the training dataset, the facial expression is described properly portion (frown eyebrows, etc.) then the only pixels the AI will train on is what hasn't been described : the lips.
When a LoRA generates something that it shouldn't, in this case something else than the mouth, then it's a database issue. Really tricky exercise. But damn rewarding when you spent days on it and it works well ;)
This LoRA is brilliant. Fackin' love the idea. Despite that little issue that can easily be fixed later with Inpaint anyway.
Just adding a few words to describe the facial expression and the mood to the dataset (but NOT the mouth obviously), then a kohya_ss re-train with those changes, should/could solve the problem.
@hansolocambo Thanks for the tip! I'll absolutely keep that in mind next time I train something similar. Glad you like the LoRA.
Will this work with SDXL models too?
While the images I used to train the Lora exceed what's needed for 1.5 checkpoints, I don't believe this Lora will be compatible with SDXL models due to how it was trained. You could potentially setup an img2img workflow but I haven't worked with SDXL models before so I can't guarantee the quality you'll get.
The coolest thing is that it works. What's not so cool is that I couldn't avoid lip enlargement (a side effect I guess) at any weight.
I appreciate the feedback. From what I recall I never tagged the training data with different lip sizes, which might have helped in this particular situation. Did you want to upload some of your work so I can check out your workflow?
I Love this lora so much!!, an SDXL version would be amazing 😁❤
I'm glad you like it! I'll look into Lora training for SDXL and see how things work out.
please, create a puckered lips Illustrious version model too.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.