Hacker News Reader

Comment by jobigoud | Hacker News Reader

GAN creating faces purely at the pixel level still seems a strange approach to me. In some years it will feel very restrictive. I guess it's the only tractable method at the moment.

Is anyone working on a GAN to generate bone structure then flesh and skin/mouth/eyes textures and pipe the result in a ray tracer?

It's incredible what can be done in 2D solving directly for the result, but imagine where this goes when this works in volume and multiple levels more driven by physics.

View on HN

I guess finding training data for bone structure, flesh, and skin textures is harder than finding pictures of faces...

Oddly, one's bone structure likely possesses more intrinsic privacy protections despite being, arguably, less personally identifiable than one's face.

Of course, this is also why training data is more difficult to acquire, as you mention.

Wouldn't being less personally identifiable mean it also offers more privacy protection?

He means that there are rules preventing the data from being shared, making it harder to find a training set.

Upon further reflection, my observation's mostly a straw man. HIPAA covers a patient's face when a dermatologist takes an image just as a x-ray covers a bone structure.

Privacy is hard to talk and reason about without defining everything specifically.

This projects images were sourced from Flickr. You can find medical imagery on Flickr reasonably easy as well it turns out.

That defeats the purpose. The purpose of deep learning is to let the machine solve the problem using only goal data. If you are going to program it with bones and flesh models then you don't need a neural net.

I disagree. If the purpose is simply to get a single-use photograph, then yes. But for this type of application I think we want to create a virtual human that can, at the very least, be photographed from two different angles and ideally put into motion. To create an virtual entity.

You wouldn't "program" it with flesh and bones, you would generate a life-like but original new skeleton in the same way we generate these images, except the space of solution is the space of possible skeletons instead of possible pixel configurations. And then generate soft tissues that are also original, conforming to biology constraints and also constrained by the underlying skeleton. Same for skin, created from scratch but believable and driven by the underlying tissue.

With enough training data, I would think you could create alternate views from the first view via ML methods, rather than doing skeletal structure by ML and then physics modeling to get views.

But what if you can train other feature lines that will generate backgrounds and other humans? Then add in another feature line that can morph an existing human line into different views, poses and/or animations.

Yes, the next step is 3D:

https://www.youtube.com/watch?v=HO1LYJb818Q

Reply on news.ycombinator.com