ZeroWw/krea2-kroma-q6k-cpu

Krea 2 Kroma v0.3.1 OPD (Q6_K) + Qwen3-VL-4B text encoder, CPU bundle

A ready-to-run set of files for generating images with stable-diffusion.cpp (sd-cli). It was tested on CPU only (2 threads, no GPU).

960x544, 8 steps, seed 42. About 11 minutes on 2 CPU threads.

Files

File	What it is
kroma-v0.3.1-turbo-opd-Q6_K.gguf	Diffusion transformer: lodestones/Kroma kroma-v0.3.1-turbo-opd, a Krea 2 fine-tune, quantized to Q6_K [edit: how you made it, e.g. sd-cli -M convert --type q6_K]
Huihui-Qwen3-VL-4B-Instruct-abliterated-Q6_K.gguf	Text encoder ("LLM"): an abliterated (refusal-direction-removed) Qwen3-VL-4B-Instruct, Q6_K. It is used only to produce prompt embeddings; it does not generate text here. [edit: source of your GGUF]
wan_2.1_vae.safetensors	Wan 2.1 VAE
test.png	Example output

Krea 2 conditions on hidden states of Qwen3-VL-4B, so the encoder cannot be swapped for a different LLM.

Usage

Build or download stable-diffusion.cpp, then:

PROMPT='Cinematic portrait of a young woman standing on a rain-soaked rooftop in Tokyo at blue hour, wearing a deep emerald silk trench coat with a high collar, long black hair blown to one side by the wind, soft freckles, a thin gold earring, looking over her shoulder at the camera with a calm, confident expression. Behind her, a blurred skyline of glowing magenta and cyan neon signs, one reading "MIDNIGHT" in white letters, reflected in shallow puddles on the concrete. Warm rim light from the left, cool fill from the neon, fine raindrops catching the light, shallow depth of field, 85mm lens, rich film grain, high detail in skin and fabric texture.'

sd-cli --diffusion-model kroma-v0.3.1-turbo-opd-Q6_K.gguf \
       --llm Huihui-Qwen3-VL-4B-Instruct-abliterated-Q6_K.gguf \
       --vae wan_2.1_vae.safetensors \
       -W 960 -H 544 -s 42 --steps 8 --cfg-scale 1.0 \
       --flash-attn off,diffusion=on --vae-tiling -t "$(nproc)" \
       -p "$PROMPT" -o test.png

Notes:

• 8 steps and CFG 1.0 are for the Turbo-speed (OPD) checkpoint.
• On a GPU build you can raise the resolution (for example 1024x1024) and drop --vae-tiling.
• Use a recent stable-diffusion.cpp master; Krea 2 support is new and flags change often.
• Smaller text-encoder quants (Q4_K_M and below) follow detailed prompts noticeably less well.

Credits and licenses

Not original work: this repo only repackages and quantizes the following. Check each license before use or redistribution.

• Krea 2: krea, released under the Krea 2 Community License.
• Kroma fine-tune: lodestones/Kroma.
• Qwen3-VL-4B-Instruct: Qwen team; abliterated variant by Huihui [edit: link to the exact source].
• Wan 2.1 VAE: Wan-AI.
• stable-diffusion.cpp: leejet (MIT).

Responsible use

Do not use these files to create sexual or deceptive images of real people without their consent, and follow the Krea 2 Community License.