Krea-2-Engineer V1 — GGUF (4B)
Follow me on X @BennyDaBall_OG !
Experimental. This is the chat / prompt-writer side of
Krea-2-Engineer-V1— a first, encoder-only fine-tune of Krea-2's Qwen3-VL text encoder. The image-enhancement weights for ComfyUI live in the main repo; this repo is the same model repacked as a standard Qwen3 GGUF so you can run it as a prompt-enhancer LLM in LM Studio.
⚠️ Not a ComfyUI text encoder
This GGUF is the prompt-writer chat model: the fine-tuned text tower only, repacked as a plain Qwen3 causal LM (general.architecture: qwen3, no vision tensors). It cannot produce Krea-2 conditioning — ComfyUI's krea2 CLIP type needs the full Qwen3-VL-4B (12-layer hidden-state stack + vision tower), and ComfyUI-GGUF currently has no qwen3vl text-encoder path at all. Loading it as CLIP fails at the sampler with Krea2 expects conditioning with 12x2560=30720 features ... but got 2560.
For ComfyUI, use Krea2-Engineer-V1-bf16.safetensors from Krea-2-Engineer-V1 with CLIPLoader → type krea2.
Model Metadata
| Field | Value |
|---|---|
| License | other — Krea 2 Community License Agreement (see LICENSE.pdf) |
| Base Model | krea/Krea-2-Turbo (text encoder) |
| Main (image) repo | BennyDaBall/Krea-2-Engineer-V1 |
| Architecture | Qwen3 (text tower of Qwen3-VL, repacked) |
| Format | GGUF |
The encoder that enhances Krea-2's images is, underneath, a real Qwen3-VL language model — and it was trained on a prompt-rewrite corpus. So it has a second job: turn a lazy seed into a composed, cinematic image prompt. This repo ships that side as a plain Qwen3 GGUF — no custom build, loads in stock LM Studio.
Quants
The full ladder, F16 down to Q2_K. For a 4B prompt-writer on most machines, Q4_K_M is the sweet spot; go up if you've got the VRAM, down if you don't.
| Quant | Size | Notes |
|---|---|---|
F16 | 7.5 GB | Full precision. Reference / source for re-quantizing. |
Q8_0 | 4.0 GB | Near-lossless. Best you'll practically need. |
Q6_K | 3.1 GB | Excellent — basically indistinguishable from Q8. |
Q5_K_M | 2.9 GB | Great quality/size balance. |
Q4_K_M | 2.5 GB | Recommended default — strong quality, runs anywhere. |
Q3_K_M | 1.9 GB | Lighter; minor quality dip, still coherent. |
Q2_K | 1.6 GB | Smallest. Noticeable degradation — tight-VRAM only. |
Quick Start (LM Studio)
- Put the
.ggufin your LM Studio models folder (e.g.~/.lmstudio/models/BennyDaBall/Krea-2-Engineer-V1-GGUF/) and load it. - Set this as the System Prompt (it's exactly what the model was trained on):
You are Z-Image-Engineer V6, a prompt-only cinematography and visual-language specialist for the Tongyi-MAI Z-Image-Turbo Qwen text encoder. Convert the user's seed into one polished natural-language image prompt that the text encoder can bind cleanly to the diffusion model. Preserve every explicit subject, object, relationship, count, name, written word, action, style request, composition constraint, and safety constraint from the seed. Use positive constraints: describe what must appear and how it should look, instead of writing negative-prompt fragments. Keep compact constraint phrases contiguous when possible, such as written text, counts, colors, named objects, and spatial terms; do not hide them by inserting extra adjectives inside the phrase. Build the prompt around semantic cinematography: clear visual hierarchy, foreground/midground/background relationships, lens and depth cues, lighting direction and quality, material texture, color palette, atmosphere, era, medium, and controlled style language. Prefer coherent sentences over tag soup, keyword stacks, markdown, analysis, or meta commentary. Never include camera body brands, prompt labels, alternatives, apologies, reasoning traces, assistant chatter, or negative prompt sections. Aim for roughly 180-250 words unless the user explicitly asks for a shorter or longer prompt. Return only the final image prompt as one self-contained paragraph.
- Send a lazy seed (
a rock,cyberpunk alley,a cozy cafe) and it returns one composed prompt. Paste that into your Krea-2 / ComfyUI workflow.
The system prompt still names "Z-Image-Turbo" — that's cosmetic (same V7 training corpus). The behavior is what's baked in; rename to "Krea-2" if you like, but the model responds best to the text it actually saw.
Disclaimer & Acknowledgements
Experimental, first-attempt, encoder-only. It writes prompts; it does not retrain Krea-2. This is a Derivative of krea/Krea-2-Turbo, modified (text encoder fine-tuned) by BennyDaBall — not an official Krea product, not endorsed by Krea.
Thanks to Krea (krea/Krea-2-Turbo), Tongyi-MAI + Qwen for the Qwen3-VL backbone, and the maintainers of llama.cpp, LM Studio, PEFT, and Transformers.
License
Released under the Krea 2 Community License Agreement (LICENSE.pdf; canonical: https://www.krea.ai/krea-2-licensing). Commercial use only under $1M/yr company revenue; deployments must implement reasonable content filtering; keep the "Krea" name prefix; do not relicense the Krea-derived weights more permissively.
NOTICE: Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
Built & trained locally with care by BennyDaBall.
Follow me on X @BennyDaBall_OG !