Thursday, November 8, 2018

Appearance Transfer

Computer networks are able to generate photographically realistic images from a simple written description. They can recognize and distinguish the various elements in a photo. And they can alter the weather or the time of day in a video clip.

In this post, we'll see how they can start with a simple outline drawing and complete it to match the appearance of another object. So, for example, a line drawing of a purse can be rendered to match the color scheme of a shoe.


The input is a simple outline drawing of a handbag (top row). The network fills them in using the colors from a photo of a shoe (left-most image).

Even more remarkably, the system can start with the basic input of a mannikin pose (top row in array below), and then transfer to that pose the appearance of a photo of a model in a completely different pose (left column).



The system will generate a plausibly photographic image of that model in the desired pose.



This video by Two Minute Papers summarizes the research (Link to YouTube)


For those of you who are interested in how these computer innovations will affect creative jobs, check out this talk by Andrew Price of Blender, who sounds an optimistic note. (Link to YouTube)
----
Scientific paper: A VARIATIONAL U-NET FOR CONDITIONAL APPEARANCE AND SHAPE GENERATION
Previously: Text to Image Synthesis
Image Translation
Semantic Image Mapping

5 comments:

Salty Pumpkin Studio said...

Thank you. That is amazing! A manikin on the desk has a new dimension.

Juan Álvaro said...

This is scary.

Roberto Quintana said...

As usual, This is a very interesting post!
Unfortunately, by the time I got to 25:25 on the Blender guy’s video I was crying so hard I was unable to see the ‘Optimistic‘ part. -RQ

James Gurney said...

Roberto, I know what you mean. The more I thought about the Blender video the less ideal it seemed. Do we really want to have computer intelligence doing the art for us? It might make the 'workflow' more efficient and inexpensive, but I worry that such would inevitably erode our capacities and turn us into art directors rather than artists.

Roberto Quintana said...

James-
The future is at our doorstep, and just like ‘Climate Change’ it does no good to deny it.
Soon pretty much all, or nearly all, of the tedious, boring, and even semi-skilled work will be automated. I don’t know about you, but all the early ‘hack-work’ I did allowed me to improve my skills and learn from all the talent around and above me.
Just like the wild fires and the record-breaking ‘century’ storms that are devastating our communities, AI will have devastating consequences on our society. First the lower and entry-level jobs will become scarce and then even middle management and production jobs will be gone. When computers can program themselves, where will all the Techies and Nerds find work? Even the Art-Directors aren’t safe w Cybernetic AI at the helm. The only thing people will be useful for will be as consumers, but consumers without incomes?
Where is that at?
I’m glad I’m still an ‘Optimist’ though, I’d hate to be a ‘Pessimist’ with such a bright future ahead. -RQ
(p.s. -I'm no robot)