So, in the DL specialization course 4 I learned that a network can be trained to generate an image based on a content image and a style image. One example was a content image Stanford University and van Gogh’s Starry Night for style image. And the generated image was bluish in color.
What if I want the network to generate in van Gogh’s style, not necessarily in Starry Night’s style? Like, the network learned from the given style image bluish colors and wavy skies, but not all van Gogh paintings are like that. Night Cafe, for example, uses reds and greens and takes place indoors.
That’s a really interesting question! Well, “all the artist’s work” might be pretty challenging. If you look at many artists, their style evolves quite a bit over their career. E.g. take a look at Picasso’s work: when he was young, he could paint in the style of the old Masters and then he went through quite an evolution and even did ceramics in addition to his incredibly varied paintings.
It’s been 4 or 5 years since I listened to the lectures from Prof Ng on this section of DLS C4, so I don’t remember if he discusses anything relevant to this question. But consider how the style part of the cost function is defined and think of ways you could extend that idea to include multiple style images, if not an artist’s complete oeuvre. Maybe you could run some experiments using two or three Van Gogh images like Starry Night, Night Cafe and Sunflowers and try different ways to select input from multiple images. If you do try anything like that, it would be interesting to hear/see what you find! You could also take a look at some of the papers on which this work is based and see if they say anything about this. Prof Ng includes references in the assignment.
2 Likes