No single model wins at text-to-image.

We blind tested eight image models with human graders, then built a router that picks the right one for the job.