Our capstone project, FaceCraft, has been a journey of innovation, learning, and overcoming challenges. Our goal was to develop an AI model capable of generating entirely new human faces using Generative Adversarial Networks (GANs). The journey has been filled with ups and downs, technical hurdles, and moments of triumph.
We encountered numerous technical issues along the way. For instance, PoPOS, our initial attempt at image preprocessing, didn’t yield the expected results. To tackle these challenges, we leveraged WSL (Windows Subsystem for Linux) and Docker to create a consistent development environment, ensuring that all team members were on the same page regardless of their operating system.
Central to our project was the StyleGAN2 architecture, renowned for its ability to produce high-quality, photorealistic images. We experimented with different GAN variants, including StyleGAN-T and StyleGAN-ADA, each offering unique benefits and challenges. These experiments were crucial in refining our approach and understanding the strengths and limitations of each model.
Our model, FaceCraft, generates images based on user prompts and random seeds. This dual capability allows for a broad range of outputs, from specific requests to entirely random creations. Additionally, we incorporated filters and the ability to manipulate images in latent space, providing users with extensive control over the generated faces.
This project has been a remarkable research endeavor. We’ve not only created a powerful AI tool but also significantly expanded our knowledge of AI and machine learning. FaceCraft has been a testament to our hard work, collaboration, and the endless possibilities of AI-driven creativity.
Comments