CivArchive
    CyberRealistic Krea 2 - v1.0
    NSFW
    Preview 138730396
    Preview 138730413
    Preview 138730414
    Preview 138730377
    Preview 138730399
    Preview 138730366
    Preview 138730364
    Preview 138730379
    Preview 138730388
    Preview 138730391
    Preview 138730404
    Preview 138730409
    Preview 138730416
    Preview 138730405
    Preview 138730408
    Preview 138730410
    Preview 138730411
    Preview 138730415
    Preview 138730417
    Preview 138730418

    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.

    👉 Join on Whop
    💬 Join the community for support, free tools and early news on Discord

    What is CyberRealistic Krea 2

    CyberRealistic Krea 2 is a full finetune of Krea 2 Turbo (Krea.ai, a 12.9B diffusion transformer with a Qwen3-VL-4B text encoder and a Qwen VAE). The base model is aesthetic first and deliberately broad. It explores medium, texture and mood rather than locking onto one look. This finetune narrows that down toward grounded photographic realism, with more consistent skin, faces and available-light rendering, and less of the base model's drift into editorial gloss.

    It's still a photographic model, not a stylization engine. Krea 2's style LoRAs and style-reference system still work on top of it, so the illustration and print looks are still there when you want them. What changed is the default: prompt it plainly and you get a photograph, not an aesthetic interpretation of one.

    What's different from base Krea 2

    • Tuned toward grounded photographic realism out of the box, with less aesthetic drift on short prompts

    • Better skin, faces and hands, with a stronger response to material and lens language

    • More stable across seeds. The base model's variety is a feature, but it makes locking a look harder

    • Full compatibility with the official Krea style LoRAs and with style reference images

    How Krea 2 prompting works

    Worth reading if you're coming from SDXL, Illustrious or Pony.

    Krea 2 isn't a CLIP model. The Qwen3-VL encoder reads your prompt as language, so almost every habit from SDXL and Pony works against you here.

    • Write sentences, not tags. Grammar parses. Spatial relationships, possessives and counts survive. Comma-separated tag lists get flattened into mush.

    • Quality tags do nothing. masterpiece, best quality, 8k, ultra detailed, absurdres, score_9, source_anime: all dead weight. Delete them.

    • Numeric weighting is broken here. (word:1.3) is CLIP syntax. On Krea 2 it drags the whole conditioning around instead of lifting one word, and it falls apart above roughly 1.2. Use LoRA strength if you need a real numeric knob.

    • Order carries the emphasis. The encoder front-loads, so whatever comes first reads as the subject. "A brass compass on a nautical chart" keeps the compass. "A nautical chart with a brass compass on it" hands the frame to the chart.

    • To emphasize something, restate it or get specific. Instead of (rust:1.4), say it twice in different words: "a rusted iron gate, orange corrosion eating through the hinges." And vocabulary is the real weight knob: oxblood beats red, sodium vapor beats orange light.

    • Always name the light source. Prompts that come back mushy almost always have two competing actions or no described light.

    • One action per subject. Two actions at once is the most common cause of a broken pose.

    Prompt structure

    subject, setting, composition and camera, lighting, mood, medium and style, texture detail

    Write it as flowing clauses, not a comma dump. If the medium is the point (risograph poster, ukiyo-e print, crayon drawing), lead with the medium instead.

    Length targets:

    • Exploration, letting the model decide: 5 to 20 words

    • Controlled image: 30 to 80 words

    • Complex scene with a specified palette: 80 to 140 words

    Past roughly 150 words the style adjectives start fighting each other. If a style reference image is active, cut the style language from the text and let the reference carry the look.

    Style LoRAs

    Krea shipped nine official style LoRAs with the base weights. All were trained on Raw, all run on Turbo, and they work on this finetune too. Each needs its trigger phrase in the prompt text, and the phrases have nothing to do with the file names.

    LoRATrigger phraseDarkbrushmonochrome ink wash styleDotmatrixmonochrome stippling styleKidsdrawingnaive expressive sketch styleNeondriptextured abstract styleRainywindowrainy window styleRetroanimepurple retro anime styleSoftwatercolorart deco watercolor styleSunsetblurethereal motion blur styleVintagetarotvintage tarot style

    Put the trigger phrase early, then keep the rest of the prompt about content, not style. Most "this LoRA does nothing" reports turn out to be a prompt fighting the LoRA with competing style words. Don't ask for photorealism while a stylized LoRA is loaded.

    Style reference images

    Krea 2 accepts style reference images alongside the text, each with its own strength. Keep the text about content and let the reference carry palette, film stock, lens feel and print texture. A polaroid as reference beats writing "shot on Polaroid 600, instant film aesthetic." Only stack multiple references if they agree visually.

    • Resolution: generate at full size. Krea 2 was trained at 2048x2048, so generating at 1024 and upscaling throws away capability you already downloaded. Good ratios: 2048x2048, 1728x2304, 2304x1728.

    • Steps: 8. This is a Turbo finetune, so more steps won't help.

    • CFG: 0.0. Because of this, negative prompts barely do anything. Skip the big negative block.

    • Mu: 1.15.

    • Quantization: fp8 is the sweet spot at about 12 GiB and fits a 16GB card with the encoder loaded. BF16 runs around 24.8 GiB, and the quality gap between them is close to invisible.

    • No custom nodes needed. Diffusion model, qwen3vl_4b text encoder, qwen_image_vae, and a LoRA loader in between if you want a style.

    Example prompts

    Available-light documentary:

    a fishmonger in a yellow rubber apron arranging silver mackerel on crushed ice at a covered market stall, early morning, cold blue daylight mixing with warm tungsten bulbs overhead, wet reflective concrete floor, photorealistic, slight motion blur on his hands

    Natural-light portrait:

    a woman laughing with her eyes closed against a pale blue sky, loose windblown strands of dark hair across her face, sleeveless white lace top, shot from a low angle in soft natural daylight, minimal background, medium format colour photograph, natural skin texture

    Low-light interior:

    a nervous teenager in a dim convenience store aisle at 2 a.m., handheld documentary photograph, 35mm lens, harsh fluorescent ceiling lights, slight motion blur, visible film grain, realistic skin texture, muted colors

    Product:

    a matte black perfume bottle on wet obsidian stone, close studio product photograph, single softbox reflection down the left edge, subtle rim light, shallow depth of field, crisp material texture

    Notes

    Short prompts are for exploration, not control. Turbo at 8 steps is fast enough to sketch, so generate a few directions first, then lock the one you want with camera, light and material detail.

    The base Krea 2 weights come under the Krea 2 Community License, not Apache. Read the actual license text before commercial work.

    Text rendering inside images sits somewhere between SDXL and Ideogram. Fine for a shop sign, weak for a poster headline.

    Description

    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.
    👉 Join on Whop

    Use the CyberRealistic Krea 2 workflow: https://civitai.red/models/2840789/cyberrealistic-krea-2-comfyui-workflow

    Checkpoint
    Krea 2
    by Cyberdelia
    Download (Beta) View on CivitAI
    realistic
    cinematic
    sf
    fantasy
    photorealistic
    base model

    Details

    Downloads
    48
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/4/2026
    Updated
    8/10/2026
    Deleted
    -

    Files

    qwen_image_vae.safetensors

    Size:
    242.05 MB
    SHA256:
    a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f

    Mirrors

    HuggingFace (219 mirrors)
    CivitAI (160 mirrors)
    TensorFiles (1 mirrors)
    ModelScope (3 mirrors)
    ModelScope CN (5 mirrors)

    cyberrealisticKrea2_v10.safetensors

    Size:
    24.48 GB
    SHA256:
    7236d96deb53a8015102e03cf72bae846947ca6f4d2951a1234f104e67c18730

    Mirrors

    CivitAI (1 mirrors)

    cyberrealisticKrea2_v10_txt.safetensors

    Size:
    4.88 GB
    SHA256:
    54bd5144df0bbc25dd6ccadfcb826b521445a1b06ae5a42570bdd2974ca87094

    Mirrors

    HuggingFace (31 mirrors)
    CivitAI (5 mirrors)
    ModelScope (1 mirrors)

    cyberrealisticKrea2_v10.safetensors

    Size:
    12.83 GB
    SHA256:
    8053facd190599b162efd3b5a85dff15c846063cf1ea4f47ae62ade2b7f03264

    Mirrors

    CivitAI (1 mirrors)

    cyberrealisticKrea2_v10.safetensors

    Size:
    7.61 GB
    SHA256:
    fcc98fbb199b225aaab079189cf032bd40b03373b56aa57c191e186ffbcdb1ba

    Mirrors

    CivitAI (1 mirrors)

    cyberrealisticKrea2_v10.safetensors

    Size:
    12.84 GB
    SHA256:
    e66d2d9ea05a7abb2036234b473b5c8bbce960be7da4852b376e6f2ad0f6f3e3

    Mirrors

    CivitAI (1 mirrors)

    cyberrealisticKrea2_v10_txt.safetensors

    Size:
    8.27 GB
    SHA256:
    36f3ff447ef59201722e8f9ce6020c9819fdcfba6aa2608c4e09b1c0ce114e34

    Mirrors

    HuggingFace (17 mirrors)
    CivitAI (2 mirrors)
    ModelScope (2 mirrors)
    ModelScope CN (1 mirrors)