One GPU, a shared queue, and measured response time. Experimental output.
Choose a garment to begin.
Queue time and model startup are separate from streaming latency.