ChrisColeTech/flux2-dev-turbo-GGUF

FLUX.2 dev Turbo — merged GGUF

The 32B FLUX.2 dev transformer with the fal Turbo LoRA merged into the weights — one GGUF file per quant tier, 12 steps, ~85 s per 1024² image on an RTX 5090.

What this repo is: true merged turbo checkpoints (base ⊕ Turbo LoRA at full precision, then quantized), plus the Ministral text encoder and VAE they need — weights only, not a retrain. No runtime LoRA required: load the transformer and generate. The Q4-class file uses a mixed-precision recipe that exists for a measured reason — see the gotchas.

---

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, 12 steps, guidance 2.5, seed 12345	**prompt:** `portrait of an older fisherman with a weathered face, natural window light, shallow depth of field, photorealistic` — 1024×1024, 12 steps, guidance 2.5, seed 99
**prompt:** `a rain-slicked Tokyo street at night, neon signs reflecting in puddles, cinematic, 35mm photograph` — 1024×1024, 12 steps, guidance 2.5, seed 1234

---

img2img

One source (generated by this same build), three prompts, identical settings — the difference you see is the prompt.

**source** — txt2img output of this build (`a dramatic mountain landscape at golden hour, snow-capped peaks, alpine lake reflection, photorealistic, 35mm photograph`, seed 42)	**img2img prompt:** `An oil painting of a dramatic mountain landscape at sunset, thick impasto brushstrokes` — 1024×1024, 12 steps, **guidance 3.5**, **strength 0.85**, seed 42	**img2img prompt:** `A watercolor painting of a mountain valley at dawn, soft wet-on-wet washes` — 1024×1024, 12 steps, **guidance 3.5**, **strength 0.85**, seed 42	**img2img prompt:** `A detailed pencil sketch of a mountain landscape in warm sepia and sanguine tones` — 1024×1024, 12 steps, **guidance 3.5**, **strength 0.85**, seed 42

strength 0.7 (the production default) preserves more of the source and reads as a lighter restyle; 0.85 is the value for a genuine change of medium. At either value the composition — peaks, valley, lake, reflection — survives.

Recommended settings

Values these weights are run with in practice.

Parameter	Production value	Meaning
width	1024	Output width in pixels
height	1024	Output height in pixels
steps	12	Denoising steps (dynamic-shifted schedule)
guidance	2.5	Embedded guidance (txt2img)
guidance (img2img)	3.5	Stronger prompt adherence for restyling
strength	0.7	img2img denoise strength (0.85 for style transfer)

Supported modes: txt2img, img2img

Notes and gotchas

• Why the Q4-class file is "mixed": a uniform Q4_K bake of merged turbo weights is broken. The merged turbo delta averages ~0.065% of weight magnitude; Q4_K's per-element quantization step on these tensors is ~3–7%. Quantizing the merged weights uniformly at Q4_K shreds the delta into a structured-noise residue — a fine crackle/debris texture across flat image regions (walls, skin, sky). Uniform Q6_K is clean but does not fit a 32 GB card at 1024². flux2-dev-turbo-Q4_K_M-mixed.gguf therefore keeps the Q4_K_M bulk but holds the artifact-carrying tensors — single-block linear2, the attention out-projections, and every modulation/embedding/conditioning tensor — at Q6_K/Q8_0. Verified equivalent to an exact-delta (runtime-LoRA) reference in same-seed A/B tests; 21.0 GB, ~22.6 GB VRAM peak at 1024².
• 12 steps on the standard dynamic schedule, not 8 turbo sigmas. With the published 8-step sigma schedule this Q4-class build ghosts (double exposure); 12-step dynamic is the verified preset. Executed img2img steps are steps × strength — schedule more if you lower strength.
• The 14B Ministral text encoder runs on CPU in the reference setup — moving a GGML-loaded encoder to CPU after encode does not free its VRAM, so it lives on CPU permanently. Prompt encode costs a few seconds.
• Distilled weights are quantization-sensitive: a uniform Q3_K_M bake of dev+turbo produced pure noise mosaics, while the base model at Q3_K_M is fine. Do not requantize these files downward.

---

Layout

Components ship as separate files: the merged transformer GGUF (pick one quant tier), the text encoder GGUF (+ tokenizer under split/text_encoders/config/), and the VAE. Any GGUF-capable Flux2 loader that accepts explicit per-component paths can consume this directly — no LoRA handling required.

A generic diffusers snippet is deliberately omitted. This is a split layout with a substituted encoder, so an upstream example will not load it unmodified.

---

Files

File	Size	Role
split/gguf/flux2-dev-turbo-Q4_K_M-mixed.gguf	21.02 GB	merged turbo transformer, mixed-precision Q4-class — fits 32 GB at 1024²; the tier these settings were measured on
split/gguf/flux2-dev-turbo-Q6_K.gguf	26.73 GB	merged turbo transformer, uniform Q6_K (needs >32 GB at 1024²)
split/gguf/flux2-dev-turbo-Q8_0.gguf	34.47 GB	merged turbo transformer, uniform Q8_0 (server/RunPod tier)
split/text_encoders/Ministral-3-14B-Instruct-2512-Q3_K_S.gguf	6.07 GB	text encoder
split/text_encoders/mmproj-Ministral-3-14B-Instruct-2512-F16.gguf	878 MB	multimodal projector (vision path, optional)
split/vae/diffusion_pytorch_model.safetensors	336 MB	VAE
split/text_encoders/config/	~34 MB	tokenizer + encoder configs

---

Provenance

• Upstream base model: black-forest-labs/FLUX.2-dev (upstream license: FLUX.2-dev Non-Commercial License)
• Turbo LoRA: fal/FLUX.2-dev-Turbo, merged at full precision before quantization
• This build: base ⊕ LoRA merged in float32 from a Q6_K dequant of the base, written to an F16 master, then quantized per tier (mixed-precision recipe for the Q4-class file). Weights are not retrained.
• License: left as unknown in this repo's metadata. Refer to the upstream licenses for redistribution and commercial-use terms — the FLUX.2-dev base is non-commercial.