We present a method for Geometry-aware style transfer and show results of various artistic models. Our geometry-aware style transfer is performed using a portrait image bank for texture style transfer, and the artist-specific geometric-style model for geometric style transfer.
For geometric stylization, we use the artists' portrait image bank, and build the geometric style model Ƥartist (similar to artist distributions). Ƥartist is then sampled to produce a set of stylized landmarks. The input image is than warped to the stylized landmarks using TPS interpolation, resulting in Іg - a geometrically stylized image.
For texture style transfer we use the algorithm proposed by Gatys et al. [1], where Іg serves as the content image, and a random sample from the portrait collection serves as the style image, resulting in a portrait that is stylized in both geometry and texture Іg+t.
In this figure, the 1st row contains the results of our geometric stylization ( І g ). The 2nd row contains the results of using the texture stylization algorithm of Gatys et al. [1], on the input image, without performing geometric stylization ( І t ). The 3rd row contains the stylization results of our g + t style transfer application ( І g+t ) . By combining both textural and geometrical stylization to model artistic style, we achieve image stylization that is more visually appealing, maintains higher artistic credibility, and present higher variation between images stylized using different artistic models.
In this figure, we show results of automatically stylizing photographs using our method on different faces. We can identify the signature geometric style of each artist: a small face with large eyes for Keane, a wide chin and large distance between the eyes for Leger, an elongated face and nose for Modigliani, distorted spatial locations of the different facial features for Picasso. Notice the second to last image, we can identify the signature Picasso style of integrating a profile into a frontal face drawing. In the last row, notice the signature style of Foujita; round, child-like face, with pointy chin, small mouth and large distance between the eyes.
Since our stylization model is comprised of both texture and geometry, we can modify each
of these style components separately, and use different combinations of artistic models to
transfer texture and geometric style independently.
In this figure, each of the portraits corresponds to a different combination of style weights (𝓌g ; 𝓌t), which are used to balance the geometric and texture style of the output image, using two artistic models - Keane and Leger (marked with red rectangles). By separating each style component, we can choose favorite style aspects of different artists, and create stylized images with inventive new styles.
Style interpolation may be extended to an arbitrary number of artistic models. This figure shows style interpolation examples between 4 different artistic models (marked with red rectangles); Foujita, Meckenem, Modigliani and Leger (clockwise, from upper-left). Each of the images corresponds to a different combination of style weights for each artistic model. By adding a geometric component to the style interpolation, we can explore numerous new portrait styles.
[1] L.-A. Gatys, A.-S. Ecker, and M. Bethge. A Neural Algorithm of Artistic Style, 2015