FLUX Schnell Local Setup: Fast Images, No Data Leaves

FLUX Schnell is a distilled, few-step text-to-image model that runs entirely on your own machine, which means your prompts, reference images, and generated outputs never leave the device unless you deliberately export them — and that local-only path is exactly what makes it a defensible privacy choice over cloud image generators. For a beginner, the practical question is not "is it good art?" but "can I prove nothing left the box?" You judge that by two things: where files are written (data residency) and what the software records (log policy).

What does the measured latency tell you?#

Our Hax Local-AI Latency Index[/glossary#hax-latency-index] held at 119.2 ms (measured 2026-07-03) and 120.8 ms (measured 2026-07-04) — a spread under 2 ms across two consecutive days, with an estimated 8.3 to 8.4 tokens/s on the control channel (estimated). Think of first-response latency like the moment a kettle starts hissing: it does not tell you the whole boil time, but a stable, low value means the pipeline reacted without a network detour. Because generation stays local, this number is not fighting a variable cloud round-trip (not measured / 측정대기).

How do you set it up in five minutes?#

1. Install a local runner (a ComfyUI-style graph app) while offline. 2. Place the FLUX Schnell checkpoint into the models folder — our bench counted 32 checkpoints installed (measured 2026-07-04) alongside 63 LoRA (measured 2026-07-04). 3. Load a basic text-to-image workflow and set a low step count; Schnell is timestep-distilled for roughly 1 to 4 steps (estimated by model design). 4. Pick a sampler from the 44 installed (measured 2026-07-04) and generate. 5. Cut the network and confirm generation still works — if it does, nothing was calling out.

How do you judge data residency and log policy?#

Open the output directory and confirm images and prompts write only to local storage. Watch outbound connections during a run; a truly local setup makes none. Check the app's log settings and set retention to the minimum you can tolerate, because a private model with verbose local logs still leaves a trail on disk. Data residency is about location; log policy is about how long that location remembers.

How do you reproduce these numbers?#

Latency came from bench_harness.probe_unified_latency; the model inventory came from bench_harness.probe_comfy_models. The 7-day HTTP P95 of 694 ms (measured 2026-07-26) reflects Hax operations, not per-image render time. Re-run the same probes on your hardware to get your own baseline.

Note: figures reflect the 2026-07-03 to 2026-07-26 window and will drift as models and hardware change; re-measure before quoting.

Related reading: 처음 설치하는 FLUX Schnell 5분 퀵스타트와 실패 지점 판단법, FLUX Schnell 응답속도, p50·p95로 체감 판단하기

Full guide: 노트북에서 돌리는 AI 모델, 흔한 함정과 해결법

References: Hax data.