[bduffany]

The "Problem #2" that they mention also has an easy solution.

If the "intermediate" variable / internal representation is just "the input image but with edges only," then of course you can see internal representations.

When you compute that intermediate variable for the line drawing, it will just happen to behave like the identity function for that particular case. So if you have already filtered out non-edges then the transformation is basically a no-op.

The "types" mismatching as they mention is not a concern because the type is just "image" i.e. a big vector of HSL values or something. Edge detection is just a convolution filter so it's going to have approximately the same type as the input.