Week2-Case study: Synthesize your own AI postcard.

This week I case study an AI program called “Synthesize AI postcard”. Basically, this is an AI algorithm that utilizes three machine learning models to explore how machine learning algorithms look at the world and how they are able to synthesize images.

I used the front camera of Macbook pro to snap myself and hand the portrait to the algorithm. 

 

As we can see, it took a long period for algorithms to generate a postcard. After this long process, here is what I got:

The texts showed at the bottom perfectly explain how the project works:

Firstly, this project uses three machine learning models: 

1. Face-Api.js: A model used to recognize faces in an image

Link: https://github.com/justadudewhohacks/face-api.js

2. Pix2Pix: turns the recognized face into a new image (mentioned in peer’s case study last week)

3. StyleGan: trained on landscape image, transform the background.

General feelings:

I think this project is not complicated in terms of its theory. We can easily understand how it works, but I still cannot understand the code and model after checking the source on Github and the neural networks on runwayMl. I think this project is critical and meaningful. First of all, it reveals how AI algorithms interact with users(humankind). How do algorithms look at us? How do they process us? These questions are really philosophical. 

runwayML example

I experience two examples in runwayML: footwear design and oil painting swap.

This is the first model that attracts my eyes. I also want to create a similar project. But I don’t know how to set up. The input displayed on the screen is a vector but I don’t know what it is. I try to upload sneakers’ images to the programming, but nothing happens. So I move on to the next model.

The second model I try is a model that converts an image to the form of an oil painting. In the content image box, we can upload the image we want to transform, and in the style image, we can upload the oil painting which is in the style we want. The model will learn and analyze the style and feature of style image and generate a new oil painting based on the content image.

 

Unfortunately, it took half an hour to processing and I could not reach to a result.