Additional Examples for the GanGogh Collection
We enhance the strongest (left side) and second strongest (right side) found in the image relative to the rest of the decomposition.
Free Style Manipulation
Interpolations between two archetypes on a grid. The archetypes are freely chosen, they are not necessarily present in the image decompositions before stylization. Indeed, this is perfectly possible to apply to natural photographs, as shown below.
The figures below show some of the 256 archetypes computed on the GanGogh collection. The full set of 256 archetype visualizations can be found here
Left: the second archetype seems to code only for "circle on rough canvas". While this is definitely the defining characteristic of the contributing images, it is not helpful for stylization. The other rows are examples of degenerate archetypes, i.e. archetypes with a single contribution.
Right: Due to the presentation of the data (we only crop the outer 10% of the images) archetype 101 seems to code for the shape of the frames.
Archetypal analysis allows to decompose an image into its contributing styles. Below are some examples of these decompositions. The contributions of the archetypal styles are visualized in three ways:
- as a texture synthesized using the archetypal style (left column)
- as stylization of the image in questions, using a unit vector for each contributing style (top row)
- as a sum of the (strongest three) contributing images
Low Quality Decompositions
The left image shows a non-sparse image decomposition, which is thus difficult to interpret. The strongest three components seem to represent the absence of texture, but it is not clear what their contribution to the image style is. This kind of artwork likely pushes the pre-trained VGG19 network to its limits since it bears no resemblence to natural images.
The right image shows a case in which all three strongest components are very similar. Again, all of them are arguably relevant, yet the decomposition is not helpful for analysis or manipulation.