It generates a few variations of input image. You can choose how much you would like generations to vary between Low, Medium and High.
It works really well on generations from Structured or Artistic model. Not so much on generations from Creative model. It's better to use Similar model in that case.
Let's look at some examples.
Input Image is generated using Structured Model with Prompt Fractal tree of life
Its variations :
Note : Images are shown in 512x512 resolution in the app, kindly download them to get 2048x2048 resolution
For Info on all models, check getting started guide.