<h1 id="see-krea2-genesis-or-personal-anime-checkpoint-on-krea-2-turbo">🎨 See-Krea2-Genesis | Personal Anime Checkpoint on Krea 2 Turbo</h1>8-Step Turbo • Clean Anime Aesthetics • Natural Language Prompting • Qwen3-VL Text EncoderDiffusion Model • 1–2 MP Native • LoRA-friendly • ComfyUI-ready<hr /><h2 id="what-is-see-krea2-genesis">✨ What is See-Krea2-Genesis?</h2>See-Krea2-Genesis is my personal anime checkpoint built on Krea 2 Turbo. The goal was simple: keep everything that makes Krea 2 Turbo good — fast 8-step generation, strong prompt comprehension, clean rendering — and push the default look firmly toward anime illustration.What you get is a model that produces anime art as its natural output instead of needing to be argued into it, while the 8-step turbo behaviour stays fully intact.It ships in five quantised builds so you can match it to your GPU instead of running whatever happens to fit. Pick one from the table below — the rest of this page applies to all of them equally.<hr /><h2 id="key-features">🎯 Key Features</h2><ul><li>✅ Anime-focused checkpoint on the Krea 2 Turbo base</li><li>✅ 8 steps at CFG 1.0 — full images in seconds</li><li>✅ Turbo distillation fully preserved, no step-count regression</li><li>✅ Strong natural-language prompt understanding through the Qwen3-VL text encoder</li><li>✅ Handles long, detailed scene descriptions — pose, camera angle and lighting all respond</li><li>✅ Clean line art and painted texture without heavy post-processing</li><li>✅ Works with the standard Krea 2 ComfyUI workflow</li><li>✅ Drop-in replacement for <code>krea2_turbo_bf16.safetensors</code></li><li>✅ LoRA-friendly — same architecture, existing Krea 2 LoRAs stack normally</li><li>✅ BF16, no quantisation artefacts — quantised variants can be derived from this file</li></ul><hr /><h2 id="which-version-should-i-download">📦 Which version should I download?</h2><h3 id="genesis-int8_convrot">⭐ Genesis-int8_convrot</h3>Size: 13.2 GB Weight error: 0.99%Recommended for: Best overall quality.Works on any GPU that supports INT8, including RTX 30-series and newer.➡️ Choose this version if you are unsure which one to download.<hr /><h3 id="genesis-fp8_scaled">Genesis-fp8_scaled</h3>Size: 13.1 GB Weight error: 2.66%Recommended for: RTX 40-series (Ada) and newer GPUs.Uses the native FP8 path and is a good choice for GPUs with strong FP8 support.<hr /><h3 id="genesis-int8">Genesis-int8</h3>Size: 13.2 GB Weight error: 1.55%Recommended for: Users who want INT8 with slightly faster runtime performance.Similar size and quality to <code>Genesis-int8_convrot</code>, but optimized more toward inference speed.<hr /><h3 id="genesis-nvfp4">Genesis-nvfp4</h3>Size: 7.7 GB Weight error: 9.41%Recommended for: RTX 50-series / Blackwell GPUs.Uses native NVFP4 and requires significantly less storage and VRAM than the INT8/FP8 versions.➡️ A great option if you have a Blackwell GPU and want a much smaller model.<hr /><h3 id="genesis-int4_convrot">Genesis-int4_convrot</h3>Size: 6.9 GB Weight error: 16.43%Recommended for: Low-VRAM systems or users who want the smallest possible download.This is the smallest Genesis build, but it also has the highest quantization error.➡️ Choose this version when VRAM or disk space matters more than maximum quality.<hr /><h2 id="how-to-read-the-&quot;weight-error&quot;">⚠️ How to read the "Weight Error"</h2>Lower is better.The weight error indicates how much the quantized model weights differ from the original model.<ul><li>0.99% → Very close to the original weights</li><li>1–3% → Excellent quality / very small difference</li><li>~9% → More aggressive compression</li><li>~16% → Significant compression, mainly intended to save VRAM and disk space</li></ul>A higher weight error does not automatically mean that generated images will be worse by the same percentage. It only describes the difference between the original and quantized model weights.<h3 id="quick-recommendation">Quick recommendation</h3>⭐ Best quality: <code>Genesis-int8_convrot</code> ⚡ RTX 40-series: <code>Genesis-fp8_scaled</code> 🚀 Faster INT8: <code>Genesis-int8</code> 💾 RTX 50-series / smaller model: <code>Genesis-nvfp4</code> 🪶 Smallest possible version: <code>Genesis-int4_convrot</code><h3 id="read-the-&quot;weight-error&quot;-column-correctly">⚠️ Read the "weight error" column correctly</h3>That number is the measured deviation of the weights from the BF16 master, not of the images. The two are only loosely related.In side-by-side testing all five builds look good, and the differences are small enough that you generally have to put two images next to each other to spot them. So do not rule out the small builds because of the number — if 6.9 GB is what fits your card, try it before assuming it is a downgrade.What the numbers are actually good for is ranking the formats against each other, and explaining why <code>int8_convrot</code> beats <code>fp8_scaled</code> at the same file size: FP8 E4M3 has three mantissa bits, INT8 has 256 evenly spaced steps, and the ConvRot rotation spreads outliers before quantising instead of letting one large value eat the scale. <hr /><h2 id="recommended-settings">⚙️ Recommended Settings</h2>The settings I personally use and recommend as a starting point:<pre><code>Steps: 8
CFG: 1.0
Sampler: euler
Scheduler: simple
Denoise: 1.0
Resolution: ~1 MP, for example 1024 × 1024</code></pre><h3 id="sampler-alternatives">Sampler alternatives</h3><ul><li>euler + simple — my default. Neutral, clean, predictable.</li><li>er_sde + simple — slightly more texture and character, works well at 10–12 steps.</li></ul><h3 id="cfg-must-stay-at-1.0">⚠️ CFG must stay at 1.0</h3>This is a turbo model. Raising CFG above 1 will burn the image — that is not a style choice, it breaks.Two consequences people trip over:<ul><li>Your negative prompt does nothing. At CFG 1.0 the unconditional branch is never computed, so negative text has no mathematical effect on the result. Whatever you put there is ignored.</li><li>Exclusions must be phrased positively. If you do not want gloves, write <code>bare hands</code> in the positive prompt. Writing <code>gloves</code> in the negative field changes nothing.</li></ul><hr /><h2 id="prompting-guide">💡 Prompting Guide</h2>This model rewards written description, not tag dumps. The text encoder is a language model, so it actually reads sentences — and that is where most of the control lives.<h3 id="prompt-structure-that-works">Prompt structure that works</h3><pre><code>[medium and style] → [camera and framing] → [pose and action] →
[face and hair] → [body] → [clothing in detail] → [background] → [lighting]</code></pre>You do not need to label the sections. Just write them in that order as flowing sentences.<h3 id="do">✅ Do</h3><ul><li>Write full sentences. 80–200 words is the sweet spot.</li><li>Name the camera angle explicitly — low-angle shot, close-up portrait, full-body view from behind.</li><li>Describe both hands separately if the pose matters. "one hand raised to her forehead, the other resting on her waist" lands far more reliably than "posing".</li><li>Describe clothing in layers and materials, not just as a name.</li><li>Watch your pronouns. This encoder reads context — a stray <code>his</code> in a single-character prompt can summon a second person.</li></ul><h3 id="avoid">❌ Avoid</h3><ul><li>Short tag dumps like <code>anime girl, red hair, desert</code>. Tags still work, but you lose almost all control over pose and camera.</li><li>Weight syntax such as <code>(red hair:1.3)</code>. Qwen3-VL encodes tokens in context, so scaling embeddings distorts the meaning rather than emphasising it. If a detail is not showing up, describe it more specifically instead of weighting it.</li><li>Filling the negative prompt. See the CFG note above — it is ignored.</li></ul><hr /><h2 id="example-prompts">🖼️ Example Prompts</h2>Five prompts covering very different territory. Each one is written the way the model likes to be talked to — use them as templates and swap the content.<h3 id="1-dark-sci-fi-portrait">1 · Dark sci-fi portrait</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6ebf578c-f55f-453c-977a-b33239be9758/original=true/6ebf578c-f55f-453c-977a-b33239be9758.jpeg" /><pre><code>A striking dark sci-fi anime portrait shows a cybernetic girl against a fractured
digital void. The illustration uses high-contrast monochrome tones with glowing cyan
glitch accents, precise line art and a cold surreal aesthetic. A tight close-up frames
her head and shoulders from a slightly low angle. She holds perfectly still, chin
lifted, staring past the viewer. Her short white hair is cut in a sharp asymmetric bob
with one side shaved, and thin luminous seams trace her cheekbone and temple. Her eyes
are solid cyan with no visible pupil. She wears a matte black high-collared bodysuit
with exposed cabling along the neck. Fragments of broken screen geometry drift behind
her. Rim lighting picks out the edge of her jaw against the darkness.</code></pre><h3 id="2-neon-graphic-full-body">2 · Neon graphic full-body</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e62cd5fb-73cc-4bdc-8cab-87f167face36/original=true/e62cd5fb-73cc-4bdc-8cab-87f167face36.jpeg" /><pre><code>A stylised graphic anime artwork depicts a mischievous spirit girl rendered entirely
in liquid chrome, outlined in electric magenta. The full-body composition places her
mid-leap above a glowing hexagonal sigil, one arm thrown back and the other reaching
forward with fingers splayed, grinning with her eyes squeezed shut. Her form is a
glossy silver silhouette with flowing hair that breaks apart into floating droplets.
High-contrast flat colour, thick neon outlines and hard shadow shapes define the style.
The background is deep violet with radiating light streaks. Magenta reflections pool
beneath her on a dark mirrored floor.</code></pre><h3 id="3-fantasy-scene-with-heavy-background">3 · Fantasy scene with heavy background</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/11dbbc2f-9d52-419d-b8f5-bca669bd6006/original=true/11dbbc2f-9d52-419d-b8f5-bca669bd6006.jpeg" /><pre><code>An anime illustration shows a sorceress standing at the centre of a flooded jungle
temple, framed by enormous carved stone faces half-swallowed by roots. A wide full-body
shot from slightly below places her small against the architecture. She stands ankle
deep in still water, one arm outstretched with palm turned upward, looking calmly
toward the viewer. Her long moss-green hair falls past her waist and her eyes glow
pale gold. She wears layered ivory robes bound with braided cord, a heavy bronze
pectoral collar and bare feet. Vines and hanging orchids drape the ruins behind her.
Shafts of warm daylight break through the canopy and scatter across the water surface.</code></pre><h3 id="4-slice-of-life">4 · Slice of life</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9458b63d-a32e-4943-9eba-f160c809107b/original=true/9458b63d-a32e-4943-9eba-f160c809107b.jpeg" /><pre><code>A soft anime illustration of a young florist arranging a bouquet on a wooden workbench
in a small sunlit shop. A relaxed medium shot from across the counter catches her
mid-motion, holding a stem of white ranunculus in one hand and secateurs in the other,
head tilted in concentration. Her wavy honey-blonde hair is pinned up loosely with
a few strands escaping, and she has warm grey eyes and a light scattering of freckles.
She wears a soft blue linen shirt with the sleeves rolled to the elbow under a canvas
apron marked with plant stains. Buckets of flowers, brown paper and twine crowd the
bench around her. Late afternoon light comes through the shop window and warms the
whole scene.</code></pre><h3 id="5-bright-character-portrait">5 · Bright character portrait</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/aa5ed5a9-93b8-406c-9a9e-c20a6ef55620/original=true/aa5ed5a9-93b8-406c-9a9e-c20a6ef55620.jpeg" /><pre><code>A cheerful anime adventurer poses in a crisp modern digital portrait with vibrant
colours and soft skin shading. A waist-up framing angled slightly from the side catches
her turning toward the camera, one hand raised in a small wave, the other steadying
a satchel strap on her shoulder. Her round face shows flushed cheeks, bright teal eyes
and a wide open smile. Her dark auburn hair is tied into two short braids with orange
ribbons. She wears a rust-coloured travelling cloak over a cream tunic, leather bracers
and a compass on a cord around her neck. Autumn woodland blurs softly behind her.
Golden late-day light catches the loose strands of her hair.</code></pre><hr /><h2 id="nsfw">🔞 NSFW</h2>Straight answer, so nobody downloads this expecting the wrong thing:<ul><li>✅ Nudity works. Artistic nude and pin-up style generations come out cleanly without extra help.</li><li>❌ Explicit sexual acts do not. You will need an additional NSFW LoRA on top for that.</li></ul>No NSFW LoRA is baked into this checkpoint. If you want the explicit range, Krea 2 NSFW V4 stacks well and is very light, so it adds the capability without pulling the anime style off course. Start at strength 0.6–0.8.<hr /><h2 id="installation">🔧 Installation</h2>All five versions are diffusion-model files and use the standard Krea 2 three-file setup. Download one of them.<pre><code>ComfyUI/models/diffusion_models/&lt;the version you picked&gt;.safetensors
ComfyUI/models/text_encoders/qwen3vl_4b_bf16.safetensors
ComfyUI/models/vae/qwen_image_vae.safetensors</code></pre>Then load them with the standard Krea 2 workflow:<ul><li>Load Diffusion Model → the version you downloaded</li><li>Load CLIP → <code>qwen3vl_4b_bf16.safetensors</code>, type <code>krea2</code></li><li>Load VAE → <code>qwen_image_vae.safetensors</code></li></ul>If you already run Krea 2 Turbo, you have the text encoder and VAE already — just drop in the checkpoint and point the loader at it.The quantised builds need a ComfyUI recent enough to carry <code>comfy_quant</code> support. If a version fails to load with scale-related errors, update ComfyUI first, or fall back to <code>Genesis-fp8_scaled</code>, which has the widest loader support.<hr /><h2 id="version-history">📈 Version History</h2><h3 id="v1.0-initial-release">v1.0 — Initial Release</h3><ul><li>Anime-focused checkpoint on Krea 2 Turbo</li><li>8-step generation at CFG 1.0, turbo behaviour preserved</li><li>Five quantised builds: INT8 ConvRot, INT8, FP8 scaled, NVFP4, INT4 ConvRot</li><li>Compatible with the standard Krea 2 ComfyUI workflow</li></ul><hr /><h2 id="credits">🙏 Credits</h2><ul><li>Base Model: Krea 2 Turbo by Krea</li><li>Text Encoder: Qwen3-VL-4B</li><li>VAE: Qwen-Image VAE</li><li>Checkpoint: SeeSee</li></ul><hr /><h2 id="license">📜 License</h2>This is a modified version of the Krea 2 model. It is not an official Krea product and is not endorsed by Krea.Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit <a target="_blank" rel="ugc" href="https://krea.ai/krea-2-licensing">https://krea.ai/krea-2-licensing</a>.Use of this checkpoint is subject to that agreement, which applies to derivatives as well. By downloading, you agree to be bound by its terms — including the commercial-use revenue threshold and the content-filtering requirements.<h3 id="my-own-terms-on-top-of-the-above">My own terms, on top of the above</h3><ul><li>Commercial use — yes</li><li>Merging and derivative models — yes</li><li>Selling generated images — yes</li><li>Credit — appreciated</li></ul>Nothing here restricts any right the Krea 2 Community License already grants you. If you merge this into something of your own, a mention of where the anime part came from is all I ask.See-Krea2-Genesis — a personal anime checkpoint on Krea 2 Turbo. 🎨

See-Krea2

FP8 E4M3, per-tensor scaled — 13.1 GB. The build with the widest loader support, and the format Krea themselves ship.Output-side projections (attention output, attention gate, MLP down) are flagged for full-precision matrix multiply, so error does not accumulate where it matters most; the input projections use the fast FP8 path. That split is not my invention — it is the policy read directly out of Krea's own official FP8 release.Measured weight deviation: 2.66 %. That is simply what FP8 E4M3 costs — three mantissa bits do not give more, and Krea's own FP8 build measures the same.Runs on: RTX 40xx (Ada) and newer for the native FP8 path. Older cards can still load it with software dequantisation. Pick this if: you are on Ada, or if another build fails to load and you want the safest option.Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP.

Files with SHA256: c90f03ab41bd135c4a24bc7c2c359b04d1f8a79a7c8b2f97764e8e35fac4ee88