U-Net vs. DiT for a Scaling Roadmap

Your team ships a text-to-image product on a U-Net-based diffusion model. Leadership wants a roadmap for the next two years, including whether to invest in scaling the current architecture further or migrate to DiT. They ask you to brief them on the actual trade-off, not just "DiT is newer."

1. Describe the mechanical difference between how U-Net and DiT process a noisy image and incorporate conditioning.
2. What is the strongest architectural argument for eventually moving to DiT, stated precisely rather than as "Transformers are better"?
3. Is "we should migrate immediately" a well-supported conclusion from what this subject covers? Justify your answer.