Inspiration
Disney characters are always changing, sometimes as fast as this world is changing. Even sometimes stressful situations, deadlines and the huge time designers pass on the table, make them lose the creativity they would like to have 24 hours every day. But, what if someone gives them this incredible idea to come up with the best Disney character that will be remembered by every child and even their parents, for the next ten, or twenty years? What if it is SOMETHING that gives them that idea?
What it does
In our project, we worked with a small dataset of Disney characters for, using a Generative Adversarial Network, come up with some drafts of what could be a fantastic Disney character in the correct mind. Using the GAN algorithm, and the existing characters, we trained the model to create some really interesting shapes and really weird colors of possible Disney characters.
How we built it
In order to build our GAN we used a GitHub repository that has the general GAN functions implemented and then we adapted the code and the inputs so they fit the model. The images (training dataset) were obtained from several web pages but most were from the Disney gallery. The GitHub repository was obtained from the following link: https://github.com/llSourcell/Pokemon_GAN/blob/master/pokeGAN.py
Challenges we ran into
The first challenge was finding an idea of a project for three students, who never worked together before in a team. After picking an idea while brainstorming on Friday, we rejected it on Saturday because one member of the team retreated. So we came up with another simpler but more creative idea, which is generating new Disney characters. The next challenge was dealing with a very sensitive and complex training algorithm, and one of the newest deep neural networks, which are still in research. Despite our limited knowledge about deep learning, we could implement our model and run the code after a big struggle with TensorFlow and GPU.
Accomplishments that we're proud of
Even when we were negatively surprised by the absence of one member of our team at HackGT, we were highly motivated to work collectively and accomplish our goal, which is "using deep learning to help companies solve some problems". We never lost hope, even when we had little experience with deep learning and we could create a good communicating team, thanks to our different cultural backgrounds.
What we learned
We learned a lot about the neural networks algorithm we used but also we realized how difficult it is to come up with an idea to create a new character, which is even more difficult when we have to implement that in an algorithm.
What's next for New Disney Characters
We know that the essence of the Disney characters comes mostly from the huge artistic team that is behind them but we think that using Deep learning techniques we could help with the character design and power, even more, the creativity and efficiency of these designers and artists. We don't want this fantastic Disney adventure to finish now!
Log in or sign up for Devpost to join the conversation.