Human2Simpson
CycleGAN-based image-to-image translation project for turning real portraits into Simpsons-style characters without paired training data.
Human2Simpson was a machine-learning exam project built around CycleGAN: learn a mapping between two visual domains without paired examples, then use that to translate real portraits into Simpsons-style faces and back again.
The hard part was the data and training problem. The project depended on building and cleaning a Simpsons-face dataset, then tuning the model enough to make the translation visually plausible despite the mismatch between real portraits and a very stylized animated domain.
What I learned
Like many GAN projects, the model quality depended as much on dataset work and training stability as on the architecture itself. The Simpson → human direction ended up more coherent than the reverse direction, which matched the intuition that cartoon faces are structurally simpler and easier to map from than to.
Stack
- PyTorch-based CycleGAN implementation
- Visdom for training inspection and qualitative monitoring
- Google Cloud / Colab-style GPU infrastructure for training runs
- Custom Simpsons-face dataset plus a portrait dataset for the opposite domain