Defending DreamBooth Without Overclaiming It

In an interview, you've just argued that DreamBooth (full fine-tuning) is the right personalization method for an AI-headshots product, citing its ~15-minute training time fitting a same-day SLA and the fact that per-user models are discarded after generating outputs. The interviewer follows up: "Okay — now design the same core feature, but as an always-on personalized-avatar product serving millions of users continuously, where a user's personalized model needs to be available on demand indefinitely, not generated once."

1. Does your DreamBooth recommendation still hold for this new product shape? Justify your answer using the same axes (training cost, storage/serving cost, identity fidelity) from the original comparison.
2. What technique would you switch to, and why does its trade-off profile fit this new product shape better?
3. What does this exchange demonstrate about how to structure a strong system-design answer in general?