<h1>🌙 Lunakrea</h1>
Warm light, clean lines, and a soft spot for the moon.
Lunakrea is a Krea 2 merge that leans toward luminous, storybook anime with a semi-real finish. It takes the bright, warm palette of Krea CinnamonSaphire, folds in the anime detail of VXP Krea2 Beta 4 (Anime), and has the TextFusion Refusal-Reduction LoRA baked straight into its weights. The result follows long, natural-language prompts closely without any extra LoRAs.
It loves a glowing light source (lanterns, neon, moonlight, oven fire), and it renders quoted text cleanly, so posters, shop signs and chalkboards are fair game.
No trigger words. Just describe the scene.
<hr />
<h2>🌒 The recipe</h2>
<ol>
<li>Base (A): <a rel="ugc" href="https://civarchive.com/models/2813776">Krea CinnamonSaphire v2.0</a> by Doughboy645</li>
<li>Blended in (B): <a rel="ugc" href="https://civarchive.com/models/2938537">VXP Krea2 Beta 4 (Anime)</a> by VXP, via <a rel="ugc" href="https://civarchive.com/models/2968817">mikmod's INT8 release</a>, with SLERP at t = 0.35 (roughly 65 % A / 35 % B)</li>
<li>Baked in: <a rel="ugc" href="https://civarchive.com/models/2775340">Krea2 TextFusion Refusal-Reduction LoRA v2.0 Full Rank</a> by Capitan01R, at strength 1.0</li>
</ol>
SLERP blends along the arc between the two models instead of averaging in a straight line, so neither parent washes the other out. The LoRA only touches Krea 2's text-fusion layers: the look comes entirely from the merge, while the LoRA changes how the prompt is read. Merged, baked and converted locally with stable-diffusion.cpp.
<h2>🌓 Settings</h2>
<ul>
<li>Steps: 10 (8–12 works)</li>
<li>CFG: 1. It is a turbo-style model, so negative prompts have no effect at CFG 1.</li>
<li>Sampler / scheduler: ER SDE + simple</li>
<li>Resolution: 832×1216 portrait, 1216×832 landscape</li>
<li>Text encoder: the Krea 2 Qwen3-VL 4B encoder (the showcase used Ektome-Qwen3-VL-4Bi-PristinelyUncensored Q4_K_M)</li>
<li>VAE: <code>qwen_image_vae.safetensors</code></li>
</ul>
The text encoder and VAE are not part of this upload. Use the ones you already have for Krea 2.
<h2>🌔 Prompting</h2>
Write it the way you would describe a painting to a friend: the subject first, then what they are doing, where they are, how the scene is lit, and finally the style.
<ul>
<li>Full sentences beat tag lists. Weight syntax like <code>(word:1.2)</code> is read as literal text.</li>
<li>Put any text you want in the picture in double quotes and say where it goes: a chalkboard on the wall reads "LUNA BAKERY".</li>
<li>Lunakrea defaults to warm and bright. If you want dark and gritty, ask for it explicitly.</li>
</ul>
<h2>🌕 Which file?</h2>
<ul>
<li>BF16 · <code>lunakrea-v1-bf16.safetensors</code> · 25.6 GB: the full-precision master, for merging, training or making your own quants.</li>
<li>FP8 (scaled) · <code>lunakrea-v1-fp8_scaled.safetensors</code> · 13.5 GB: a lighter safetensors file for ComfyUI.</li>
<li>Q8_0 · <code>lunakrea-q8_0.gguf</code> · 13.6 GB: near-lossless GGUF. Every showcase image was rendered with this exact file.</li>
<li>Q5_K · <code>lunakrea-v1-q5_K.gguf</code> · 8.8 GB: for tighter VRAM, at a small quality cost.</li>
<li>Q4_K · <code>lunakrea-v1-q4_K.gguf</code> · 7.2 GB: the smallest file; some fine detail is lost.</li>
<li>INT8 · <code>lunakrea-v1-int8.safetensors</code> · 13.5 GB: for ComfyUI INT8 (convrot) workflows. It also runs natively in recent stable-diffusion.cpp builds.</li>
<li>W4A4 · <code>lunakrea-v1-int4_convrot.safetensors</code> · 9.4 GB: ComfyUI INT4 W4A4. Experimental, and it does not run in stable-diffusion.cpp yet.</li>
</ul>
The other formats are converted directly from the BF16 master, never from another quant. GGUF files need ComfyUI-GGUF, Forge or stable-diffusion.cpp.
<h2>🌖 Content note</h2>
Both parent models are NSFW-capable and the refusal-reduction LoRA is baked in, so Lunakrea handles mature content when you prompt for it. The last showcase image contains nudity; everything else is SFW. All characters are adults. Please follow CivitAI's rules and the laws where you live.
<h2>🌗 Cover prompt</h2>
<blockquote>
A young woman with long silver hair and pale violet eyes stands on the tiled rooftop of an old Japanese inn at night, holding a glowing paper lantern in one hand. She wears an indigo kimono patterned with small white crescent moons and a loose obi that flutters in the wind. Behind her an enormous full moon fills half the sky, soft clouds drifting across it, and the warm orange lantern light mixes with the cool blue moonlight on her face. Full-body shot from a slightly low angle, detailed anime illustration with soft painterly shading and fine linework.
</blockquote>
10 steps · CFG 1 · ER SDE / simple · 832×1216 · seed 7110422
Every showcase image keeps its full prompt and settings, so you can remix any of them.
<h2>🌘 Credits</h2>
Thank you to Doughboy645 (CinnamonSaphire), VXP (VXP Krea2) and mikmod (its INT8 release), Capitan01R (TextFusion Refusal-Reduction), and the Krea team for Krea 2. Lunakrea would not exist without their work. 💜

LunaKrea

v1.1: fixed for ComfyUI 🌙
<ul>
<li>The same weights as v1.0: the SLERP merge of Krea CinnamonSaphire v2.0 (A) and VXP Krea2 Beta 4 Anime (B), t = 0.35, with the Krea2 TextFusion Refusal-Reduction LoRA v2.0 Full Rank baked in at 1.0</li>
<li>Fixes the v1.0 files in ComfyUI. They stopped with "tuple index out of range": the conversion through stable-diffusion.cpp's GGUF writer had dropped a dimension from one tensor (<code>txtfusion.projector</code>), and the GGUFs also had their modulation tables quantized, which ComfyUI-GGUF cannot read. Every file is rebuilt and loads in both ComfyUI and stable-diffusion.cpp.</li>
<li>GGUFs in ComfyUI need a ComfyUI-GGUF build that knows Krea 2, such as molbal/ComfyUI-GGUF</li>
<li>Seven formats: BF16, FP8 (scaled), Q8_0, Q5_K, Q4_K, INT8 and W4A4 (INT4 convrot)</li>
<li>New showcase images: the v1.0 prompts on new seeds</li>
<li>Recommended: 10 steps · CFG 1 · ER SDE / simple · 832×1216</li>
</ul>

Files with SHA256: f711ef5d215857a9255d5f20ab7f760cd911fb6638a11a1ad8ebba7e09482011