They built a system that can be trained on real-world data so it can output more real-world-like data. It's much more sophisticated than what you could hand-tune with Perlin noise and can have coherent terrain features across multiple scales which are harder to do in classical Perlin noise.