[Other] Unpainted version of VMS

[oshfdk]

Now that I think of it, it could be possible to train a more powerful model using weighted combinations of the training data points. E.g., if we combine 30% of a known ink kernel and 70% of a known paint kernel, using proper computation mimicking physical blending of paints, and then label this as ink, we can get a much larger set of training points for detecting faint ink and separating it from layers of paint.