Team: Shirley, Justin, Apurva, Ryan
The purpose of our project is to generate anime-like characters from a portrait image of a person. Anime, also known as Japanese animation, has gained traction specifically in Western communities. Nowadays, through the influence of social media creators, many individuals are interested in transforming their profile photos into anime-style images that resemble characters from their favourite animated shows. Our motivation for this project stems from the fact that commercial image editing software falls short in achieving this style transfer that we and many others are hoping for. Additionally, creating anime images manually in specific styles requires professional artistic skills. In order to accomplish our goal, we can develop a deep learning model, specifically a Generative Adversarial Network (GAN), which would be capable of accurately translating features of a human face into an anime-style illustration. We will work towards a transformation which would preserve the unique characteristics of the individual while applying the stylistic elements and shapes of features of the given style of anime. For example, anime characters tend to have colourful hair, small noses, and large and expressive eyes whereas human faces do not always possess these distinct facial features. GAN’s are specifically good at generating high-quality images along with transferring styles between different domains because of their architecture which involves two unique networks, a generator and a discriminator. However, it might be less competent at shape transformation, which will be a major challenge for this project. GAN variants such as CycleGAN and StyleGAN are excellent at style transfer tasks, and CycleGAN in particular can learn the mapping between human faces and anime styles without needing exact pairs of corresponding images. For our specific project, we will use a modified version of CycleGAN that incorporates data from annotated facial landmarks to generate a more detailed anime character.