ALI LIN
02Ali Lin Lab
2026

Teardown No 1

An API's documentation tells you what it returns. Only measurement tells you what it withholds.

The Teardown harness cropped to its middle 60 percent, showing the prompt, seed and guidance controls down the left, a timing readout for fal-ai/flux/dev in the centre, and two generated garment images on the right.

Problem

These APIs publish what they can do and withhold what they cost you. The one timing the response returns is the one that never moves.

Solution

Turning "it feels quick" into segments of numbers.

Methods

I made 47 structured calls across five experiments and logged every measurement to disk.

Next Step

The method ports to any inference API, but the findings do not.

02/05
← IndexAli Lin

TEARDOWN № 1

an instrumented teardown of a generative image API

measurement harness + interactive reportTypeScript / React / fal API2026shipped

the instrument running, 25.2s · click to play
open live ↗

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

  1. 1Step sweepwhat does another step cost, and what does it buy
  2. 2Guidance sweepis the second knob the same kind of knob
  3. 3Seed determinismdo identical parameters return identical bytes
  4. 4Latency serieswhere the waiting actually goes
  5. 5Friction loghow far the documentation sits from the wire

FINDINGS

  1. 1The hidden segmentQueue sd 277 ms against inference sd 9.9. The segment carrying nearly all the variance is the one the response body never returns.
  2. 2Identical bytesThree runs, three different inference times, one sha256, zero of 262,144 pixels different. Determinism holds and the documentation never says so.
  3. 3Past 28S20 to S28 moves 7.655 percent of pixels. The next eight steps move 1.488.
  4. 4No resting pointNo 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.

Ali Lin