Did I miss the part of the article where they create a Monet style from multiple Monet works and apply it, rather than using the style of an individual painting? I thought that was the point of this. They sort of do it in the last image, I guess. Seems to pick up so much actual color, though! (as opposed to perhaps the quality of the color or how it's used, which is just as characteristic of a particular style as the color itself)
There has been other recent work (Preserving Color in Neural Artistic Style Transfer) that allowed control, separate from the "style", over how much of the original painting's color palette to use:
It isn't too hard to extrapolate from this to using separate images as sources for the style and palette, and from there to transforming the "content" image's color palette to resemble the palette of the "style" image in terms of saturation, value, contrast, etc. (but not hue) by manipulating histograms, and then using that result instead of either the "content" image's or the "style" image's palettes.
Because if you have a photo of a bright red cardinal and a desaturated pastel drawing of a pale blue boat, the result you are likely to be most pleased with would be a pastel cardinal that is, well, pink, and not a pale blue cardinal.
Of course, the harder version of this would need to recognize the semantics of both the "content" and the "style" image in order to construct a more meaningful palette to be applied (which is what you're getting at), and that same semantic information will also help with transferring the style more meaningfully (eg. crude palette-knife houses in the background but small brushed strokes depicting the people in the foreground).
I agree. While this is super impressive stuff it seems to match palette far more than style. Van gogh has a distinct style even though not all of his paintings have the same color palette.
