<h1>📷 LegacyKrea</h1>
Photographic realism for Krea 2, with instruction editing built in.
LegacyKrea is a Krea 2 merge made for realistic photography: natural skin texture, believable light, and women, build and background as its strongest subject. It renders portraits, fashion, documentary, cinematic stills, wildlife, vehicles, space, technology and landscapes, and it handles mature content when you ask for it.
It is also an image → image model. The same file that renders from text can take a picture and change it by instruction: "change her coat to a red raincoat", "turn the scene into night", "give her round glasses and a straw hat". The showcase pairs originals with their edits.
No trigger words. Describe the photograph you want.
<hr />
<h2>🧪 The recipe</h2>
A merge over the official Krea 2 Turbo base
<h2>⚙️ Settings</h2>
<ul>
<li>Steps: 8 (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: Euler + 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 encoder, its vision projector and the VAE are attached as optional components. Skip them if you already have Krea 2's.
<h2>✍️ Prompting</h2>
Write it like a photo brief: the subject and her details first, then pose and action, the setting, the light, and finally the camera (lens, depth of field, film grain).
<ul>
<li>Full sentences beat tag lists. Weight syntax like <code>(word:1.2)</code> is read as literal text.</li>
<li>Age: faces lean toward the twenties, and a number on its own barely moves them. Describe age you can see instead: silver or grey hair, laugh lines, wrinkles.</li>
<li>Men and couples: it leans female, so a "couple" can come back as two women. Write "a man and a woman" and describe him too.</li>
<li>2.5D: LegacyKrea is a photographic model first. Asking for a 2.5D or 3D render usually still gives you a photo, so pick a stylized model for illustration work.</li>
<li>Text you want in the picture goes in double quotes, with where it appears.</li>
</ul>
<h2>🪄 Image → image</h2>
Give it a picture and write the prompt as an instruction: what to change, and what must stay ("Change her gown to deep emerald-green silk. Keep her pose, her face and the desert unchanged.").
<ul>
<li>Works well at the normal 8 steps / CFG 1: clothing and hair changes, recolours, adding accessories, changing the time of day.</li>
<li>Weak: big whole-scene changes such as covering a whole street in snow, and restyling a photo into another medium. Those tend to come back nearly unchanged.</li>
<li>Removing an object is the hardest edit for distilled Krea 2 edit models. If it keeps painting the object back, raise CFG to about 3 at around 20 steps and name the object in the negative prompt.</li>
<li>In ComfyUI, editing needs the Krea 2 edit nodes (search the Manager for Krea 2 Identity Edit). A reference grounding of 768 px is a good default.</li>
</ul>
<h2>📦 Which file?</h2>
<ul>
<li>BF16 · <code>legacykrea-v1.0-bf16.safetensors</code> · 25.6 GB: the full-precision master, for merging, training or making your own quants.</li>
<li>FP8 (scaled) · <code>legacykrea-v1.0-fp8_scaled.safetensors</code> · 13.5 GB: a lighter safetensors file for ComfyUI.</li>
<li>Q8_0 · <code>legacykrea-v1.0-q8_0.gguf</code> · 13.6 GB: near-lossless GGUF. Every showcase image was rendered with this exact file.</li>
<li>Q5_K · <code>legacykrea-v1.0-q5_K.gguf</code> · 8.8 GB: for tighter VRAM, at a small quality cost.</li>
<li>Q4_K · <code>legacykrea-v1.0-q4_K.gguf</code> · 7.2 GB: the smallest GGUF; some fine detail is lost.</li>
<li>INT8 · <code>legacykrea-v1.0-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>legacykrea-v1.0-int4_convrot.safetensors</code> · 9.4 GB: ComfyUI INT4 W4A4. Experimental, and it does not run in stable-diffusion.cpp yet.</li>
</ul>
Every format is converted directly from the BF16 master, never from another quant. GGUF files need ComfyUI-GGUF (a build that knows Krea 2), Forge or stable-diffusion.cpp. Editing with a GGUF text encoder in stable-diffusion.cpp also needs the encoder's vision projector (the optional <code>mmproj</code> component).
<h2>🔞 Content note</h2>
LegacyKrea is NSFW-capable and renders nudity and adult themes when prompted. It also leans NSFW on its own: a clothed prompt can come back nude (an open coat, a missing top). When you need a safe-for-work image, spell the clothing out ("a buttoned coat", "fully clothed") and check the result before you post it.
Several showcase images contain nudity, including an intimate scene between two women; CivitAI marks them as mature, and the rest are SFW. Every person shown is an adult. Because faces lean young, describe adult features clearly when you prompt mature content. Please follow CivitAI's rules and the laws where you live.
<h2>🖼️ Cover prompt</h2>
<blockquote>
A snow leopard rests on a rocky ledge high in the Himalayas, its thick spotted grey coat dusted with snow and its pale eyes staring into the camera, jagged peaks soft in the background. Wildlife photograph, 600mm lens, crisp fur detail, cold natural light.
</blockquote>
8 steps · CFG 1 · Euler / Simple · 1216×832 · seed 886512173
Every showcase image keeps its full prompt and settings, edits included, so you can remix any of them. No LoRAs were used: one prompt still carries a <code>&lt;lora:…&gt;</code> tag from a copied prompt, but it had no effect.
<h2>🙏 Credits</h2>
Thank you to Krea team for Krea 2. LegacyKrea is built entirely on their work.

LegacyKrea

v1.0: first release 📷
<ul>
<li>Add-difference merge over Krea 2 Turbo</li>
<li>Text → image and instruction editing in one file</li>
<li>Seven formats: BF16, FP8 (scaled), Q8_0, Q5_K, Q4_K, INT8 and W4A4 (INT4 convrot), all converted from the BF16 master</li>
<li>Optional components: the Qwen3-VL 4B text encoder (Q4_K_M), its vision projector, and the Qwen Image VAE</li>
<li>Showcase rendered with the Q8_0 file; originals are paired with their edits</li>
<li>Recommended: 8 steps · CFG 1 · Euler / Simple · 832×1216</li>
</ul>

Files with SHA256: 447be7ba0c6cd4cdc2908865d945046ee59b06256f586804f17a1dcfa3b9d305