Updated: Jul 1, 2025
concept
Applies a dynamic, action-shot, cinematic visual style to an image. The model will generally add a lot of motion lines, perspective shots, dramatic and stylized poses, and other stuff. Characters will naturally start spinning and jumping etc. when you apply this lora, but if you really want to make them stay still, I'm generally able to do it by explicitly specifying the position they should be in, e.g. "standing, arm up, reaching" or something similar.
The keyword is "epicaction". Just combine that with any general description of the action, e.g. running, jumping, spinning, reaching, backflip, surfing, whatever.
The model was trained on a dataset of 114 manually selected images scaled to 1024x1024 for training. The dataset was repeated twice and run with a batch size of 2, for 114 steps per epoch. The model was trained for 50 epochs for a total of 5,700 training steps with a learning rate of 10e-4. The Adafactor scheduler and optimizer were used.
I'm actually having a lot of fun with this so hopefully you do as well!