This LoRA is designed to represent one particular type of afterimage: a stretched, distorted version of the subject itself. In Japan, this type of effect is often referred to as an "obake".

The training was done using a meager dataset of only 33 training images. As a result, the LoRA tends to generate an excessive number of arc-shaped obake effects.

For effects such as attack_trail when swinging a weapon, the effect may actually work better without this LoRA.