TEARDOWN № 1
an instrumented teardown of a generative image API
WHAT
Teardown is a measurement study of a hosted inference API (fal.ai's flux endpoint) instrumented from a browser the way an integrator would hit it.
WHO
Anyone who has waited for a generation and wondered what the wait was made of.
Anyone building on one of these APIs, who has to pick a default and cannot pick it from the documentation.
WHY
A benchmark answers which model is better. Nobody was measuring what happens inside the one you already chose.
METHODS
5 experiments
- 1what does another step cost, and what does it buy
- 2is the second knob the same kind of knob
- 3do identical parameters return identical bytes
- 4where the waiting actually goes
- 5how far the documentation sits from the wire
FINDINGS
- 1Queue sd 277 ms against inference sd 9.9. The segment carrying nearly all the variance is the one the response body never returns.
- 2Three runs, three different inference times, one sha256, zero of 262,144 pixels different. Determinism holds and the documentation never says so.
- 3S20 to S28 moves 7.655 percent of pixels. The next eight steps move 1.488.
- 4No adjacent guidance pair falls below 17.167 percent, and the last rung still moves 57.591. It never converges.
TEARDOWN № 1 is the first in a series. The method, sweep and diff and time and log and publish with sources, ports to any inference API. The findings do not: they are one model, measured.
