Topic: Create a Galaxy
Inspiration: Inspired by One of Memo Akten’s work as below: 
With the input video of an eye, it’s output is a Stardust. He mentioned that this is not done by style transfer.
What’s more, I’ve tried to make galaxy like images in Processing before, and it is very troublesome, especially to mimic the foggy and cloudy like texture. I used noise() in Processing for mimicking, but the result is not very good. Here’s an example:
So this time, I decide to use style transfer to make a galaxy image from the images of eyes, with input style image like this:
Process: I’ve encountered difficulties in different parts. First, when downloading the dataset, Because of some reason, it retries more than 3 times to download it. And when I tried to run the train.sh file, it always has the error message of readByte Error. After searching, I knew that this is due to the dataset is corrupted. So I redownload a lot of times and follow the instruction afterward to train the model. I trained 3 models, and it takes a lot of time. Currently, two of them are still under training. But the process is fluent.
Result:
This is the input image of the eye:
Here’s the style transfer result:
Which the result is very bed with eyes. I think it’s because the galaxy input image doesn’t have an obvious style.
Then, I think of what if I don’t use the eye picture, but some other images. I chose wall images instead. And here are the results.
Input:
output：
I think the result with wall images are much better and kinda realistic. Although the images of galaxy hardcoded by Processing are more colorful and rich. But I think the texture made by machine learning is more natural, especially the last one.