Hacker News Reader

Comment by ma2rten | Hacker News Reader

I feel like it's misleading of them to say that a computer painted the picture. That makes it sound like all works of Rembrandt have been fed into a Machine Learning algorithm and finally the computer spit out this piece.

After having watched the YouTube video, I fell like actually the picture was more or less photoshopped together from existing Rembrandt paintings. The role of the algorithms was "only" to generate the most average / representative of his works.

View on HN

I used to paint old masters portraits as a hobby (including Rembrandt). I think it's a bit disingenuous to describe this as 'painting'. On top of ma2rten's criticism that it basically photoshopped previous examples to create a most representative work, I'd add that 3D printing with texture is not painting. Rembrandt, in particular, was known for painting with countless layers of semi transparent glazes to give his paintings a sense of depth and an ethereal glow. He'd also mix his pigments with bindings aside from pure linseed oil, sometimes adding chalk or glass. 3D printing the final result with none of that process makes for a very limited reproduction.

Also, I'll add that I think it's very impressive, but I think they're overselling it here.

The 3D printed element is the least interesting part of the process because it's literally superficial. There's no reason you couldn't set up a robotic brush painter next to a robotic pigment factory to recreate all of the above.

However - what's interesting to me is that the only way to make this work was by defining the features that were going to be included, analysing and averaging them separately, and gluing them together to create a whole.

This wasn't a "Throw everything into an RNN and see if something useful falls out" project. It deliberately separated specific target features.

That works well for an artist with very consistent and narrow clothing, lighting, and background choices, and limited model poses.

It might not work so well for artists who included a much wider range of imagery - which is probably most of them.

I'm guessing recent progress in texture synthesis [1] and content generation [2] using deep convolutional nets made this possible. It would be interesting to see what was actually fed into the generating system. That would be a piece of art truly speaking magnitudes.

[1] http://arxiv.org/abs/1505.07376

[2] http://arxiv.org/abs/1511.06434

Those techniques aren't capable of generating images at high resolution that even remotely look as crisp as what was done here.

You can read on their website [1] what they did, and seems that it involved a lot of manual labour still.

[1] https://www.nextrembrandt.com/

I'm also getting the feeling that they are being dishonest about the techniques they've used.

It's also very misleading to say that it's in the style of Rembrandt when there are so many factors that make his works unique missing from this painting.

Reply on news.ycombinator.com