Part of my facial expressions series. All of these LoRAs are designed to mesh well with each other and have low impact on your image framing.
It's important to face your fears, or at least be able to put a face to them.
Jan 18, 2024 Update:
It was pretty clear that this model wasn't trained as well as the other core expressions, and didn't mix as well, so v2 is normalized much better but you may need to prompt it harder. It should be pretty much SFW out of the box, and degrade images less, but I imagine some people might still like v1 better for some of their images.
Using the LoRA
It has a hard time working well without hiresfix if the face isn't a prominent part of the framing, so try running with hiresfix if you aren't seeing the face reflect the expression. Check out the gallery for examples
Non-trained words like "wide eyed" work really well. As with my other expression LoRA's I recommend adding strength to the prompt words and not the LoRA weight.
Starter prompts:
<subject>, scared
<subject> <context> BREAK scared expression
Description
Reduced style and face transfer, and reduced occasional amplification of NSFW tendencies from the base model
FAQ
Comments (5)
Any tips on how to describe the images references to create this kind of Loras? Thanks.
For sure! Do you mean how they are captioned in particular, how the images are chosen? Or something broader?
@thegipper Exactly. How they are captioned, how the images are chosen. And number of epochs x repeats. Thanks!
@thegipper Yes, I would very much like to know how you trained your LoRAs. I'm experienced with training people and styles but I had no idea that it was possible to train facial expressions. I downloaded all of your LoRAs and they all worked very well with a (photo-realistic) character I trained. My intention is to render this character with a variety of facial expressions, add them to its dataset, and train again with those expressions specified. Then I tried to use your LoRAs to manifest a deep frown and was not able to do it. So I decided to train my own facial expression LoRA by collecting photos of that feature this type of frown. But my training failed. I think it's because I did not write proper captions. It would be good to know how you trained these. Otherwise, I will not know how to properly caption my own character for subsequent training.
What values do you have set in Ksampler?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.