<h1><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1bd247aa-d68b-4307-b140-11e8d7c39cbd/original=true/1bd247aa-d68b-4307-b140-11e8d7c39cbd.jpeg" /></h1>You can get this model through Civitai Early Access, or grab it along with many others by joining The Tinkerer on Whop. Membership gets you early releases, private tools and members-only pages. There are free pages too, no membership needed.👉 <a target="_blank" rel="ugc" href="https://whop.com/cyberdelia-ai-lab">Join on Whop</a> 💬 <a target="_blank" rel="ugc" href="https://discord.gg/88R2hdjb22">Join the community for support, free tools and early news on Discord</a><hr /><h2 id="cyberrealistic-krea-2">CyberRealistic Krea 2</h2>CyberRealistic Krea 2 is a full finetune of Krea 2 Turbo, <a target="_blank" rel="ugc" href="http://Krea.ai">Krea.ai</a>'s 12.9B diffusion transformer with a Qwen3-VL-4B text encoder and the Qwen Image VAE.Base Krea 2 is aesthetic first. It explores medium, texture and mood instead of settling on one look. Great for creative work, less useful when you want a photograph.This finetune moves the default toward photography. Natural skin, believable faces, available light, real materials and camera language, with less of the editorial gloss the base model drifts into.The creative side didn't go anywhere. The official style LoRAs still work, style reference still works, and the model will leave photography behind when you ask it to. What changed is where it starts.Base Krea 2 likes to interpret. CyberRealistic Krea 2 leans toward the camera.<h2 id="what-changed">What changed</h2><ul><li>Photographic by default, no stack of realism keywords needed first</li><li>Less aesthetic drift on short or simple prompts</li><li>More natural skin texture, less plastic rendering</li><li>Better consistency in faces and hands</li><li>Stronger response to camera, lens, lighting and material descriptions</li><li>More predictable when you're trying to lock in a specific look</li><li>Still moves into illustration, print or anime when you prompt for it</li><li>Works with the official Krea 2 style LoRAs in my testing</li></ul>Seed variation is still there. I didn't try to remove it, since that's part of what makes Krea 2 interesting. CyberRealistic just makes it easier to stay inside one visual language.<h2 id="which-version-should-i-download">Which version should I download?</h2>CyberRealistic Krea 2 ships in five precision and quantization formats. All five contain the same finetune. What changes is file size, memory use, hardware requirements and, to a smaller extent, the image itself.If you're not sure, FP8 is a starting point. INT8 ConvRot is also worth trying if your hardware and backend support its accelerated inference path.<h3 id="bf16-reference-version">BF16 · reference version</h3>The original full-precision release and the version the other four are built from. Highest numerical precision and no quantization loss, but the largest files and the highest memory requirement.Best for:<ul><li>Maximum quality</li><li>Merging, converting or further training</li><li>Systems with plenty of VRAM and RAM</li><li>Anyone who wants the master version</li></ul>For normal image generation you don't need BF16.<h3 id="fp8-recommended-for-most-users">FP8 · recommended for most users</h3>FP8 reduces weight storage to roughly half that of BF16 for the quantized layers. Actual file size and VRAM use depend on how many layers remain in higher precision and how your workflow loads the model. It's a practical option, but not a guarantee of the highest speed or closest match to BF16: INT8 ConvRot can outperform it on supported setups.Best for:<ul><li>Most ComfyUI users</li><li>GPUs with limited VRAM</li><li>Everyday generation</li><li>BF16-like results without BF16 requirements</li></ul>Start here if you don't want to think about it.<h3 id="int8-convrot-efficient-int8">INT8 ConvRot · efficient INT8</h3>INT8 ConvRot has shown lower deviation from BF16 and faster inference than the FP8 versions compared in published community benchmarks. Those results don't establish a universal ranking for every model and setup. Hardware acceleration, backend support and workflow compatibility still matter, so both formats are provided.Best for:<ul><li>Low VRAM systems</li><li>GPUs with strong INT8 performance</li><li>A smaller model without giving up much quality</li></ul>INT8 isn't automatically faster than FP8. That depends on your GPU and backend, so FP8 stays the safer general-purpose pick.<h3 id="mxfp8-blackwell-optimized-fp8">MXFP8 · Blackwell-optimized FP8</h3>MXFP8 is microscaling FP8. Instead of one large shared scaling range it scales small blocks of values, so the FP8 data keeps more accuracy. The real payoff comes from hardware with native MXFP8 acceleration, mainly NVIDIA Blackwell.Best for:<ul><li>RTX 50-series / Blackwell</li><li>Setups with native MXFP8 support</li><li>Anyone who wants an efficient modern FP8 format</li></ul>On older cards MXFP8 gets emulated, so expect little or no gain over normal FP8.<h3 id="nvfp4-smallest-and-most-aggressive">NVFP4 · smallest and most aggressive</h3>NVIDIA's 4-bit floating point format for Blackwell. By far the most compact of the five and the lowest memory footprint. It's also the most aggressive quantization here, so differences from BF16 and FP8 turn up more often depending on the prompt and settings.Best for:<ul><li>Blackwell / RTX 50-series</li><li>Minimum VRAM use</li><li>Minimum model size</li><li>Efficiency over absolute precision</li></ul>Don't pick NVFP4 just because the file is small. On hardware that can't use it you lose most of the reason for it.<h3 id="quick-pick">Quick pick</h3><ul><li>Plenty of VRAM and want the original: BF16</li><li>A general starting point: FP8</li><li>Supported INT8 acceleration and looking for better speed or fidelity: Try INT8 ConvRot</li><li>RTX 50-series / Blackwell GPU: MXFP8</li><li>Blackwell and want the smallest file: NVFP4</li></ul>Differences between the quantized versions shift with your GPU, backend, workflow, resolution and prompt. No single format wins everywhere.<h2 id="recommended-settings">Recommended settings</h2>These apply to Krea 2 Turbo and CyberRealistic Krea 2.<ul><li>Sampler: Euler</li><li>Scheduler: Simple</li><li>Steps: 8</li><li>CFG (ComfyUI KSampler): 1.0</li><li>Guidance (official Krea code): 0.0</li><li>Mu / timestep shift: 1.15</li><li>Resolution: 1K to 2K</li></ul><h3 id="about-that-cfg-value">About that CFG value</h3>Two conventions exist here and they cause a lot of confusion. Krea's own inference code uses guidance 0.0, which means CFG is off. The ComfyUI equivalent of "off" on the standard KSampler is CFG 1.0.Don't set the regular KSampler to 0.0. That breaks generation.Turbo runs without CFG, so traditional negative prompts do very little in the standard workflow. The official ComfyUI workflow simply zeroes the negative conditioning.<h3 id="steps-and-resolution">Steps and resolution</h3>Turbo is an 8-step distilled model. More steps are not automatically better.Krea 2 Turbo is built for roughly 1K to 2K, so there's no need to generate everything at 1024 first if your hardware can handle more. Useful starting points:<ul><li><code>2048 × 2048</code></li><li><code>1536 × 2048</code></li><li><code>2048 × 1536</code></li><li><code>1152 × 2048</code></li><li><code>2048 × 1152</code></li><li><code>1440 × 1920</code></li></ul>Lower resolutions are fine when you want speed or need the VRAM.<h3 id="required-models">Required models</h3><ul><li>CyberRealistic Krea 2 diffusion model</li><li><code>qwen3vl_4b</code> text encoder</li><li><code>qwen_image_vae</code></li></ul><h2 id="prompting">Prompting</h2>If you're coming from SDXL, Pony or Illustrious, read this part.Krea 2 has no CLIP text encoder. It uses Qwen3-VL-4B, so it reads your prompt much more like normal language. You don't have to forget everything you know, but most old Stable Diffusion habits stop paying off here.<h3 id="describe-don't-tag">Describe, don't tag</h3>Natural language works very well. Comma-separated clauses are fine, and Krea's own examples use them, but each clause should describe something instead of being a loose tag.Instead of:<code>woman, street, night, realistic, masterpiece, best quality, 8k, detailed skin</code>try:<code>A woman waiting alone on a quiet city street at night, photographed under the warm light of a shop window, with natural skin texture and wet pavement reflecting the streetlights.</code>The second prompt tells the model what the image actually is.<h3 id="quality-tags-do-nothing-special">Quality tags do nothing special</h3>Words like <code>masterpiece</code>, <code>best quality</code>, <code>ultra detailed</code>, <code>absurdres</code>, <code>score_9</code> and <code>source_anime</code> aren't quality switches here. Qwen reads them as ordinary language. They might nudge the result, but they're usually taking up space you could spend on something useful.Describe the detail you want instead.<h3 id="skip-numeric-weighting">Skip numeric weighting</h3>ComfyUI can parse <code>(rust:1.4)</code> syntax, but I wouldn't build Krea 2 prompts around it. Stronger wording is more predictable:<code>a heavily rusted iron gate, its hinges covered in rough orange corrosion</code>If you're running a LoRA, LoRA strength is a much better numeric control.<h3 id="put-the-subject-first">Put the subject first</h3>Prompt order still matters in practice. If the image is about a brass compass, start with the compass:<code>A weathered brass compass resting on an old nautical chart...</code>No need to obsess over exact word order. Just make the subject and the main composition clear early.<h3 id="be-specific">Be specific</h3><ul><li><code>red dress</code> becomes <code>deep oxblood satin dress</code></li><li><code>orange light</code> becomes <code>warm sodium-vapor street lighting</code></li><li><code>old wall</code> becomes <code>sun-bleached plaster wall with hairline cracks and flaking paint</code></li></ul>Specific words carry more visual information than repeating "detailed" and "realistic."<h3 id="describe-the-actual-light">Describe the actual light</h3>This is the biggest single win for photographic prompts. Instead of <code>cinematic lighting</code>, say where the light comes from:<code>soft daylight entering through a north-facing window</code><code>a single tungsten bulb hanging above the table</code><code>late-afternoon sunlight coming through venetian blinds</code><code>cold fluorescent ceiling lights</code><h3 id="keep-actions-simple">Keep actions simple</h3>Complex actions work, but if a pose starts breaking, simplify. One clear action per subject is a good rule. Get that working first, then add the secondary details.<h3 id="quote-your-text">Quote your text</h3>If you want words in the image, put the exact text in quotation marks:<code>a small neon sign reading "OPEN ALL NIGHT"</code>Krea specifically recommends this.<h3 id="a-structure-that-works">A structure that works</h3>subject → setting → composition/camera → lighting → mood/style → material and texture<code>A middle-aged fisherman repairing a red net on a wooden dock, photographed from waist height with a 50mm lens, grey sea behind him, soft overcast morning light, muted natural colors, weathered hands, damp wool sweater and rough salt-stained wood.</code>It doesn't have to read like literature. Dense descriptive clauses work fine. The point is that the words describe the image instead of listing magic tokens.<h3 id="how-long">How long?</h3>My practical ranges, not model limits:<ul><li>5 to 20 words: exploration, let the model decide</li><li>30 to 80 words: good balance of control and freedom</li><li>80 to 150 words: complex scenes, precise styling, detailed composition</li></ul>Longer prompts work. Length isn't the problem, contradiction is. Ask for soft window light, hard flash, dreamy pastels, deep black shadows and flat commercial lighting in one prompt and the model has to pick a winner. More words only help when they add information.<h2 id="example-prompts">Example prompts</h2><h3 id="available-light-documentary">Available-light documentary</h3><pre><code>A fishmonger in a yellow rubber apron arranging silver mackerel on crushed ice at a covered market stall in the early morning. Cold blue daylight enters from the open side of the market and mixes with warm tungsten bulbs above the counter. The concrete floor is wet and reflective, with slight motion blur on his hands and natural texture in the skin, rubber and fish scales.
</code></pre><h3 id="natural-light-portrait">Natural-light portrait</h3><pre><code>A woman laughing with her eyes closed against a pale blue sky, loose dark hair blowing across her face, wearing a sleeveless white lace top. Photographed from a slightly low angle in soft natural daylight with a minimal background, medium-format color photograph, shallow depth of field and natural skin texture.
</code></pre><h3 id="low-light-interior">Low-light interior</h3><pre><code>A nervous teenager standing alone in a convenience-store aisle at two in the morning, photographed handheld with a 35mm lens. Harsh fluorescent ceiling lights fall across the shelves and face, with slight motion blur, visible film grain, muted colors and realistic skin texture.
</code></pre><h3 id="product-photography">Product photography</h3><pre><code>A matte black perfume bottle standing on wet obsidian stone, photographed close-up in a dark studio. A single large softbox creates a narrow reflection along the left edge of the bottle while a subtle rim light separates it from the background. Shallow depth of field, crisp glass and stone texture, controlled specular highlights.
</code></pre><h2 id="a-few-last-things">A few last things</h2>Short prompts are great for exploring. Turbo is fast enough that generating five directions first and then adding camera, light and material detail to the one you like beats writing a perfect prompt blind.Text rendering is much better than older diffusion models, but it still isn't a typography engine. Quote the exact text and keep signs short.And the licensing. Krea 2 is not Apache 2.0. The weights and everything derived from them fall under the Krea 2 Community License, which covers commercial use, redistribution, attribution and derivative naming. Commercial use is free below the stated annual revenue threshold. Above it you need an Enterprise License from Krea. The license also requires derivative model names to start with "Krea", which is why this one is called CyberRealistic Krea 2 and not the other way around.If you're using or redistributing this model, read the current Krea 2 license instead of assuming the usual open-model terms.

Krea 2 CyberRealistic

Grab this model and many more by joining The Tinkerer on Whop. Membership gets you early releases, private tools and members-only pages, plus a lot more. 👉 <a target="_blank" rel="ugc" href="https://whop.com/joined/cyberdelia-ai-lab/products/the-tinkerer/">Join on Whop</a> CyberRealistic Krea 2 V4 brings another round of refinements, with improved realism, consistency, detail and overall image quality.It’s also a little less shy when things get… interesting. And yes no LoRAs were harmed, borrowed or secretly sacrificed in the making of the sample images.

Files with SHA256: abb1af0bdb2245b969ad0e202e05e1eda8be28f74700b689cbae7f95d5bf36f8