Style Transfer - Quick notes

4 April 2019

Description of Style Transfer

Version of the paper I worked from is here

Updated version of the paper here - see Fig 2

Some simple successes

Almost the view from the office

There are lots of variables. So that I’m keeping almost everything constant, I’ll be using this view as the content for style transfers. Look at the shape of the moon.

Voyager Animation

Vger

I haven’t seen this done elsewhere, but it’s simple. Each frame is the result of a separate - but consistent - training from random image towards a sequence of frames I created from nasa’s free voyager model.

It’s pretty enough, but I think that training from the previous frame’s output image will improve the outcome.

How does it cope with big bold colours?

Because everyone uses the style of Starry Night.

Cost vs iteration. Only the first 2800 iterations (optimisation cycles) are in the animation, but the jumps in cost are noticeable

… another quick bold example …

Hulk Smash. Or, The Test that Failed.

Breakout after 23500 iterations. Grey goo.

Breakout after 99900 iterations. Not better.

needs less white

And what about this as a style?

Question: How well do you think this will come out?

Swap them around

Question: Can I try to transfer the style of a photograph onto a painting?

style

content

Answer: No.

A couple more just for fun

The dappling is art, not compression artifact

… just to prove that you’re not restricted to 800x600. Generation time seems to scale with number of pixels - well, doubling them to 1000x1000 below roughly doubled time and not much worse. Not A Benchmark :-)

Church at Auvers

Notice dark 'strokes' of separation and retention of (some) straight edges