A AI model allows you to translate Nike sneakers into Adidas style and vice versa.
This project utilizes cycleGAN to deal with sneaker style translation. By training the cycleGAN with Nike and Adidas sneakers' side image, the model can receive a rough sneaker design sketch, and transform them into black&white sneaker design
Personal music streaming system focused on listening to whole album with my own album cover.
Cover Covers is a personal music streaming environment that appreciates the concept of album. It offers an experience of listening to albums with self-created album cover.
With the emergence of streaming services like Spotify, people gained easy access to music in a unit of songs. It brought music consumers a lot of convenience, but being able to curate a playlist with favorite songs, people stopped listening to albums as a whole. Listening to whole albums require time and sometimes endurance, but it provides us with a different experience that is certainly worthwhile. Even knowing this, I always find my playlist with songs from different albums.
I wanted to force myself to listen to albums. It's not that listening to individual songs is bad. I genuinely wanted to listen to albums, but my laziness and various excuses (eg. I'm too busy to sit down and listen to music for 30 minutes) stopped me from doing it.
So, I decided to design a process that would make listening to albums a fun activity. Human beings tend to value their own creations than those of others' regardless of its quality. The more beginner one is, the stronger this tendency. That's why my first pasta tasted better than restaurants'. If I were able to recreate an element of a music album, I would be eager to listen to it.
Not being able to draw and paint, I used a set of digital tools including machine learning and Photoshop to cover an album cover (as in covers in music). Then, I created a system where I can simply type a name of an album and have it played on my streaming website. It automatically swaps original album cover images with my own covers hosted on my blog.
A recreation of traditional Korean folk painting in the eyes of an AI
The project is a recreation of a Minhwa painting called “The Tiger and the Crow” using a styleGAN2 model – a creation of a neural-net styled version of Minhwa painting.
Minhwa (민화) is a traditional Korean folk painting that has a unique feature that the painter is, most of the time, anonymous. Minhwa can be drawn by anyone. However, as many traditions get lost over time, the production of Minhwa paintings have been diminishing since the early industrialization era. Despite the current efforts to revive the tradition, it is slowly being lost as new generation steps away from folk paintings.
If everyone can be a Minhwa painter, I thought of a notion that an AI can also be an anonymous painter – an AI as an anonymous creator in order to pay respect to the long tradition of Minhwa creations. Through this project, I wanted to emphasize that anyone can be a Minhwa painter regardless of who you are. When even an AI can be a Minhwa painter, why can't you be one?
Although the product of this particular machine learning model seem odd and “imperfect” compared to the many Minhwas drawn by famous artists, the imperfection of the created painting portrays the essence of what folk paintings are – I am not striving to create a perfect folk painting, but a statement to the community that there is no “perfect” Minhwa.
The project involves using StyleGAN2, a machine learning model that creates similar styled images using a given training dataset images, as the main component. I was able to collect approximately 500 different Minhwa images with various styles, and trained on top of an existing pre-trained model to expedite the process.
This project let you control my body, which means it takes your image as input and generate a images of me with the same pose
In this project, I use CycleGan, which is a powerful and useful machine learning algorithm to train my own model. It can generate images of myself based on the pose images (my name is Edmund so the project name is ‘This is Quite Edmund’). Those pose images are the result of a pre-trained machine learning model, which is called ‘DensePose’, it can return you an image showing poses of all the people in the input image.
During the training process, I took some videos of mine, and get frame images of those videos as the dataset of my own images. And then I upload those images to DensePose to get the results and store them as the dataset of pose images. After training for a long time, I eventually get a satisfactory model.
Finally, I wrap DensePose and my model up in a workspace, so I get another brand new model by combining them together. By using this model, you can send in your own images, and will get images of me with the same pose as you. You can image you are actually controlling my body.
