Krea 2 Turbo (+ Identity Edit) — GGUF build

8-step distilled text-to-image and identity-preserving instruction editing — ≈35 s per 1024² image on an RTX 5090, guidance-free.

What this repo is: the Krea 2 Turbo Edit denoiser (identity LoRA baked in) as GGUF at three quant tiers plus a scaled-FP8 build, with the Qwen3-VL-4B text encoder and Qwen-Image VAE it needs — weights only, not a retrain. The settings below are the values these weights are actually run with day to day.

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, 8 steps, guidance 0.0, seed 12345}	_{prompt: `portrait of an older fisherman with a weathered face, natural window light, shallow depth of field, photorealistic` — 1024×1024, 8 steps, guidance 0.0, seed 99}

_{prompt: `a rain-slicked Tokyo street at night, neon signs reflecting in puddles, cinematic, 35mm photograph` — 1024×1024, 8 steps, guidance 0.0, seed 1234}

Identity Edit

The signature mode. One reference image plus a plain-language instruction — the subject's identity, pose, lighting and scene survive; only the instruction changes. No mask, no inpainting region.


_{reference — the seed-99 portrait above, fed back as the edit source}	_{edit instruction: `put a red knitted beanie hat on his head` — 8 steps, seed 7, ref_boost 1.0, grounding 768 px}	_{edit instruction: `make it rain outside the window, wet glass with raindrops` — 8 steps, seed 7, ref_boost 1.0, grounding 768 px}

Up to two reference images are accepted (subject + style/context reference). ref_boost above 1.0 weights the reference harder if identity drifts.

img2img

Conventional strength-based restyling also works. One source (generated by this same build), three prompts, identical settings.


_{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, 8 steps, strength 0.75, seed 42}	_{img2img prompt: `A watercolor painting of a mountain valley at dawn, soft wet-on-wet washes` — 1024×1024, 8 steps, strength 0.75, seed 42}	_{img2img prompt: `A detailed pencil sketch of a mountain landscape in warm sepia and sanguine tones` — 1024×1024, 8 steps, strength 0.75, seed 42}

Recommended settings

Values this build is run with in practice.

Parameter	Production value	Meaning
`width`	`1024`	Output width in pixels
`height`	`1024`	Output height in pixels
`steps`	`8`	Denoising steps
`guidance`	`0.0`	Fully distilled — CFG is not used
`strength`	`0.75`	img2img denoise strength
`ref_boost`	`1.0`	Edit mode: reference weighting
`grounding_px`	`768`	Edit mode: reference grounding resolution

Supported modes: txt2img, img2img, edit (identity-preserving, 1–2 reference images)

Notes and gotchas

guidance 0.0 is correct. This is a fully guidance-distilled build; the text conditioning arrives through the Qwen3-VL encoder, not CFG. Raising guidance is not the lever for prompt adherence.
The identity LoRA is baked into this denoiser (turbo_edit weights). There is no separate LoRA file to apply; plain txt2img and img2img use the same weights.
Edit mode wants an instruction, not a description — put a red beanie on his head, not a man wearing a red beanie. Both reference identity and scene survive the edit (verified: same face, same window, same shirt across edits at seed 7).
≈11 GB VRAM resident for the Q4_K_M tier at 1024²; txt2img ≈35 s, edits ≈105 s on an RTX 5090.

Layout

Components ship as separate files: the denoiser (pick one quant tier), the Qwen3-VL-4B text encoder, and the Qwen-Image VAE. Any loader that accepts explicit per-component paths for the Qwen-Image family can consume this directly.

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/krea2_turbo_edit-Q4_K_M.gguf`	7.26 GB	denoiser — the tier these settings were measured on
`split/gguf/krea2_turbo_edit-Q6_K.gguf`	10.58 GB	denoiser, higher-precision quant
`split/gguf/krea2_turbo_edit-Q8_0.gguf`	13.71 GB	denoiser, highest-precision quant
`split/fp8/krea2_turbo_edit-fp8_scaled.safetensors`	13.14 GB	denoiser, scaled-FP8 (Ada/Hopper native path)
`split/text_encoders/qwen3vl_4b_fp8_scaled.safetensors`	5.24 GB	Qwen3-VL-4B text encoder (scaled FP8)
`split/vae/qwen_image_vae.safetensors`	254 MB	VAE

Provenance

Upstream base model: Krea 2 Turbo Edit (Krea AI), Qwen-Image family architecture
This build: requantized into GGUF tiers and a scaled-FP8 build, redistributed with the encoder and VAE alongside. Weights are not retrained here.
License: left as unknown in this repo's metadata. Refer to the upstream model's license for redistribution and commercial-use terms.

Downloads last month: 4,575

GGUF

Model size

13B params

Architecture

qwen_image

Hardware compatibility

4-bit

6-bit

8-bit


_{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, 8 steps, strength 0.75, seed 42}	_{img2img prompt: `A watercolor painting of a mountain valley at dawn, soft wet-on-wet washes` — 1024×1024, 8 steps, strength 0.75, seed 42}	_{img2img prompt: `A detailed pencil sketch of a mountain landscape in warm sepia and sanguine tones` — 1024×1024, 8 steps, strength 0.75, seed 42}