AI-Generated Sims Faces

The Sims is one of the most famous simulation games ever made. You create a character and manage their life. I spent hours in it as a kid, and one day wondered: what if I trained an AI to create my Sims for me?

Building a dataset that didn’t exist

There’s no public dataset of Sims faces, so I made one:

• Opened the game’s character-creation screen and wrote a script to repeatedly hit the “randomize” button, generating thousands of unique faces.
• Recorded about an hour of randomised characters, then wrote another script to split the footage into individual images and assemble a clean dataset.

Training the model

• Used StyleGAN, starting from a model pre-trained on celebrity faces (transfer learning), and fine-tuned it on my Sims dataset in Google Colab.
• Documented the learning curve: the network starts from random blobs and gradually resolves into highly detailed, game-accurate faces.

Why it mattered to me

It’s a small idea taken end-to-end: inventing the data, training the model, and generating something new. It’s also proof that a fun question is often enough to start a real machine-learning project.