Creating my own customized celebrities with AI
Gaming
Introduction
In a fascinating CS 229 final project, Kerry trained an autoencoder on 13,000 celebrity headshots from famousbirthdays.com. Using principle component analysis (PCA) on the encoding, Kerry developed a face editing tool that allows for facial expression changes, makeup application, and even the prediction of what a baby would look like if Barack Obama and Donald Trump had a child together. The project aimed to explore the mathematical aspects of PCA and its ability to generate various facial attributes. The findings were not only intriguing but also showcased the potential of AI in creating customized celebrities.
Overview of the Project
Using an autoencoder trained on celebrity headshots, Kerry employed PCA to create a slider-based face editing tool. With over 300 components, the tool could manipulate various facial features, such as brightness, skin tone, shading, smile, eyebrow height, and more. Each component represented a specific attribute, and adjusting the sliders allowed for customization of celebrity images.
Key Learnings and Findings
The project revealed several remarkable insights:
The average face: Combining thousands of celebrity images, Kerry created an average face that was symmetrical and well-proportioned. However, it appeared more feminine than expected, possibly due to the higher number of female celebrities in the dataset.
PCA components: The first component represented brightness, while the second component represented skin tone. Other components included shading, temperature, eye shadow, smile, eyebrow height, and more.
Mathematical correlations: The components discovered were a result of mathematical analysis and did not rely on human decision-making. This demonstrated the algorithm's ability to identify underlying patterns in the dataset.
Artificial face generation: Kerry's tool had a randomization feature that produced entirely new faces based on the distribution of the actual celebrity images. These artificial faces offered a glimpse into what celebrities might look like in a different universe or era.
Keywords
face editing tool, autoencoder, PCA, celebrities, celebrity headshots, average face, brightness, skin tone, shading, temperature, smile, eyebrow height, makeup, facial features, customization.
FAQ
Q: How did Kerry create the average face?
A: By combining thousands of celebrity headshots, Kerry used PCA to calculate the average face, representing symmetrical and well-proportioned features.
Q: Can the face editing tool replicate real-life celebrities?
A: Yes, the tool allows users to input a celebrity's name and replicate their face by adjusting the sliders accordingly.
Q: How accurate are the predictions of what a baby would look like from combining two celebrities?
A: The predictions for a baby's appearance are generated based on the features of the two chosen celebrities. However, it should be noted that the results are purely hypothetical and for entertainment purposes.
Q: Are the makeup components accurate in reflecting a celebrity's makeup preferences?
A: The makeup sliders offer a range of options, including the intensity of eye shadow and lipstick. While they might not precisely replicate a specific celebrity's makeup, they provide a customizable feature to experiment with different looks.
Q: How many components did Kerry identify through PCA?
A: There were over 300 components discovered through PCA, representing various facial attributes and features.
Q: Can the tool be used with non-celebrity faces?
A: The tool was trained on celebrity headshots, but with slight modifications, it could potentially be adapted to work with non-celebrity faces as well.