One Step Generation

This two-stage training process, whilst pragmatic, raises some deep questions: why can't we train a one-step generator directly, without needing to first train a diffusion model?