A common frustration I have with diffusion models is the averaging out of emotions and expressions. Characters' expressions usually feel a bit bland and lifeless.
This lora tries to offer a more vivid and intense portrayal of the terrified/intense fear/panicking emotion. It can do either the sort of angry, **********-overload, contorted-eyebrows, twisted-pained type of fear, or the anticipatory, tearful, panicking, sheer-terror kind of fear. Both styles are portrayed in the examples.
The keyword is "xtrmterror". It's the only keyword required, and it's also the name of the file, because I hate it when people aren't clear about their keywords. Although you can often get decent results with the keyword alone, it's more effective when you combine it with a few other commonly used words for emotions. An example would be "xtrmterror, scared, tears" or "xtrmterror, angry, constricted pupils".
Seems to work best with 1024x1024 images, although the difference is small, and it might just be my imagination. Upscaling and using facedetailer will generally improve it quite a bit, since it likes to have a lot of pixels to work with for the face area. Different character loras can have fairly widely varying results. Some kinda neuter the effect, while others work great.
The lora was trained on 36 images with a max resolution training size of 1024x1024. Training was conducted over 70 epochs for 1,610 training steps with a batch size of 2 and a learning rate of 10^-4. The Adafactor scheduler and optimizer were used. Training was performed on my own 4060ti 16GB. Please feel free to contact me with any other questions about the model or the training.
My example images were generated with WAI-NSFW-Illustrious.