Bias in distort selection procedure?

[Ofnuts]

Looking at the code, the "granularity" is implemented by scaling down the image by the granularity factor, applying spread noise and blur, and scaling the image back. I think that what happens is that the spread noise on something close to a layer edge biases the shape away from the edge because it will produce pixels mostly towards the center of the layer (the "balancing pixels" fall outside the limits and are not taken in account).