How do you approach the trade-off between model latency and creative quality in real-time generative tasks at Mirage?
