Text-to-image latency

Setup

I run text-to-image on the same RTX 5090 rig that handles video work, with the GPU power cap locked at 403 W so thermals stay predictable run to run.

What I measured

Wall-clock time from job accept to finished image in the app logs — one image at a time, no batch queue tricks.

Numbers

2–2.5 s median per image across recent jobs. The spread is real: different resolutions and step counts show up in production, and I am not flattening that into a single fake precision number.

What’s next

Tighter harness with fixed resolution and steps so the card and the write-up can be compared apples to apples.

Method

Median over logged production jobs in the shared image workflow; single GPU; no batching.

Caveats

Resolution and step count vary by workflow; range comes from app logs, not a fixed harness.

← All benchmarks