Every year the pokemon company ends up hiring talented artists to work on the new designs and features of a new generation of pokemons. This is a very time and energy consuming process as one generation of pokemon inludes about 100 different designs. With that in mind, we decided to try to use machine learning to generate new pokemons base on the old design and features in a hope that furture artist would be able to use these as a refrence to make their life easier.
Our machine learning model uses unsupervised representation learning with deep convolutional Generative Adversarial Networks to extract key features of the images and generate new images base on it. We ran the Kaggle complete Pokemon data set (about 800 pokemons in total) though it and the link to this data set is found here. https://www.kaggle.com/datasets/kvpratama/pokemon-images-dataset.
As our result, we were able to generate pokemon images with features resembling legs, wings or other features.
The following pictures are an overview of machine generated pokemon image from our model. We can see the generated result getting better overtime.
Generated Image 1 of 100
Generated Image 20 of 100
Generated Image 40 of 100
Generated Image 60 of 100
Generated Image 80 of 100
Generated Image 100 of 100
One idea that we think would improve our result is to train pokemons with physically similar and expand the data set mirroring images and twerking colors of the image for a better learning curve.