Each step can be inverted, so the model can trace a generated sample back to the simpler distribution.
The model accounts for how each transformation stretches or compresses regions. This lets it calculate probabilities under suitable designs.
ELI5
For example, imagine stretching a grid drawn on a flexible sheet without folding it over itself.
The grid becomes uneven, but you can still trace a stretched point back to its original position.
Why use reversible steps?
They let the model move between the simple starting distribution and the learned distribution in either direction.
What limits which steps can be used?
The model needs transformations whose inverse and change in volume can be computed efficiently.

