Approach
- A PyTorch-based GAN for handwritten digit generation on MNIST
- Core Vision Transformer components implemented from scratch
- Pretrained ViT-based image inference explored
Why it stays here
This is the lightest of the six projects, and its value is the foundation rather than the outcome: attention, patch embedding, and adversarial training are one thing to read about and another to write. When later work moved into VLA territory, that layer did not need relearning.