ChrisColeTech/flux1-dev-turbo-GGUF

FLUX.1-dev Turbo v4 — GGUF build

12B FLUX.1-dev turbo-merge. Split GGUF layout.

What this repo is: a requantized, split-layout build of the upstream model — weights only, not a retrain. The recommended settings below are the values these weights are actually run with day to day, not copied from the upstream card.

---

Samples

**prompt:** `a red apple on a rustic wooden table beside a window, soft daylight, fine skin texture and wood grain, studio photograph, sharp focus` — 1024×1024, 30 steps, guidance 1.0, cfg 1.0, seed 12345	**prompt:** `portrait of an older fisherman with a weathered face, natural window light, shallow depth of field, photorealistic` — 1024×1024, 30 steps, guidance 1.0, cfg 1.0, seed 99
**prompt:** `a rain-slicked Tokyo street at night, neon signs reflecting in puddles, cinematic, 35mm photograph` — 1024×1024, 30 steps, guidance 1.0, cfg 1.0, seed 1234

---

img2img

This model does img2img as well as txt2img. All four results below use the same prompt; the captions give the settings that produced each.

**source** — a photograph, used as the img2img input	**img2img prompt:** `transform into oil painting with thick visible brushstrokes` — 1024×1024, **30 steps**, **guidance 1.0**, **strength 0.5**, seed 42 — about 52 s on an A100 80GB	**img2img prompt:** `transform into oil painting with thick visible brushstrokes` — same source and seed at **guidance 3.5**, 20 steps, strength 0.7
**source** — a synthetic gradient, the form `create_synthetic_test_image()` produces	**img2img prompt:** `transform into oil painting with thick visible brushstrokes` — 1024×1024, **20 steps**, **guidance 3.5**, **strength 0.9**, seed 42

Keep guidance at 1.0

This repo holds the turbo v4 merge. Flux uses embedded guidance — the scale is a conditioning input to the transformer, not a second unconditional pass — and this merge is distilled for a low value. engines/flux1.py presets guidance 1.0 and records degradation above roughly 1.5.

Raising it does not sharpen prompt adherence; it adds content that is not in the source. The third image above is the same source and seed at guidance 3.5 — the figure used throughout tests/flux1-dev, which was written for the full flux1-dev checkpoint — and it contains a person who does not exist in the input.

Across the sweep at strength 0.7, guidance 1.0 / 2.0 / 3.5 gave mean pixel differences of 27.5 / 27.6 / 30.5 with luminance flat at ~112: the parameter barely moves tone, it changes content. Invented figures appeared at 3.5 and 5.0, not at 1.0 or 2.0.

strength 0.5 preserves the scene; 0.9 replaces it

strength	result on a photographic source
0.5	composition preserved — ridgeline, valley, sun position and cloud structure hold
0.7	preserved, with more freedom in the foreground
0.9	scene replaced — the mountain source returned a coastline

Steps

12, 20 and 30 all restyle. 30 is the preset and gives the cleanest result; 12 suffices when speed matters. Step count does not change tone (luminance ~112 across all three), only refinement.

On the synthetic source

The bottom row uses a generated gradient, the source form tests/flux1-dev builds at runtime via create_synthetic_test_image().

It needs strength 0.9 to transform at all. At the photographic preset (strength 0.5) the same source returns a mean pixel difference of 3.7 — visually identical to the input — because a uniform circle and smooth gradient carry no detail for the sampler to work from. At 0.9 the difference is 47.5 and the gradient becomes radial paint strokes with drips and impasto ridges, though the flat circle still reproduces flat.

So the strength that suits a photograph and the strength that suits this fixture are not the same number. Subject preservation also cannot be observed against a source with no subject — validate img2img against a real image as well.