Pushing Neural-Style to the limits...not really :|
- 0 Comment(s)
- May 18, 2017
I had to do it, I'm so goddamn picky on how everything gets old so fast, that after submitting artwork i'm ten steps away from that style that everybody is loving. Its not that I don't love it, its just that it become mundane, it loses its nostalgia, kind of like a roller coaster ride. It does get boring after 50x or so. So I like leaving a nice first ride impression.
So down to the nitty gritty, what are the secret ingredients to creating stunning neural network art ?, well I will tell you, its not the Style Loss or the Total Loss parameters, Its Visual Aesthetics, you see, our brain is extremely complex at looking at good things it finds nice, just like a rotten apple or a beautiful carved marble stone piece, you will find it when you create it. but you will be getting rocks and pebbles when you start..and that's ok.
So far i'm finding out the more relu's means better art, but the parameters I use so goddamn much has become soo second nature, I can technically visually see the effects before the get rendered.
But Victor, WHAT RELU's ???
Also, its gonna scare the crap out of your GPU's because it will eat more video ram than a Chinese buffet, But 9GB of Total GPU memory should render 512px just fine..
And also but not least, you will need to use VGG_ILSVRC_19_layers.caffemodel and VGG_ILSVRC_19_layers_deploy.prototxt
Plus the actual limits of Neural-Style are.... 5x NVIDIA TESLA K80's. Luckly I am on a mission to obtain these so called legendary nvidia artifacts for my personal artwork riches and put NASA Satellites to shame with my 12K Pixel Artwork.
artwork riches - Endless Creativity