CivArchive
    K2 | ARSA - INT8 Conv [Scaled]
    NSFW
    Preview 146693442
    Preview 146693129
    Preview 146693123
    Preview 146693125
    Preview 146693110
    Preview 146693120
    Preview 146693113
    Preview 146693117
    Preview 146693099
    Preview 146693116
    Preview 146693127
    Preview 146693105
    Preview 146693101
    Preview 146693907
    Preview 146693112
    Preview 146693111
    Preview 146693109
    Preview 146693107
    Preview 146693115
    Preview 146693122

    🎨 Krea 2 | ARS-A [Turbo]

    Styled version


    📖 About the Model

    Base model: Krea 2 Turbo
    Precision: Full Precision BF16
    Model size: 24 GB
    Prompting: Natural Captions & Booru Tags

    Full BF16 precision means full compatibility with LoRA. The model works best with them: characters, and more.

    • 🎯 YES, baked-in ARSA styles. The model has mixed-in styles.

    • 🖼️ Baked LoRA Tune. The dataset was expanded with a tuned set of 350 high-quality images.

    • 🏷️ Flexible prompting. Works with both Captions and Booru Tags.

    • 🔁 A good replacement for Illustrious or Anima models.

    This is third release, so I haven't run any big experiments with the model yet.


    🆕 NEW

    1. 🖌️ Aimed at artistry and style

    This model is aimed at artistry and style. Don't expect ultra-detailed images from it, since a LoRA is mixed in here that I trained specifically for art-stylized images — semi-brushstrokes, artistry, brushes, gouache, etc.

    Very beautiful landscapes come out.

    2. 🔞 Regarding Uncensored

    I generated [R, R+, X] content perfectly fine, but not [XXX]. Even the Base model won't let you do that.

    To generate full-fledged 18+ content, there are LoRAs to unlock it, for example this one:

    I won't bake them in, this is your own choice.

    (Breasts, nudity, genitals — pass generation and censorship on this model.)

    3. 🖼️ Regarding preview images and the lack of prompts

    I won't change my opinions, but I'll make a concession: the first 5 images will always have prompts for my resources.

    (I don't like sharing them. It's my personal decision, and it will remain so.)

    If you don't like this approach, or you have an obsessive mania that you're being deceived, then look for other authors and their models. I'm not going to deceive anyone here or even try to.

    I don't use any LoRAs or modifications in the previews to influence the final result — just simple and direct image generation.

    4. 🛡️ Merge Authorship Protection

    • Merging: prohibited, if the model does not declare permission for such actions on Civit or other platforms.
      (Many people don't even read the permissions and prohibitions, so I integrated a small protection.)

    • Quantization: remains and will always be available to you.

    5. 📦 Personal quantization requests

    If many people ask for personal quantization into certain formats that you can't do yourself, and the request is supported, of course I will release that model as soon as possible.

    My Guide for Quantization of my Models: https://civarchive.com/posts/31786656


    My settings for Krea 2 Turbo

    • CLIP Skip: None (or 1)

    • CFG Scale: 1 – 1.5 (max)

    • Sampler: Euler

    • Scheduler: SGM_Uniform, Beta, or Simple ⭐ best way!

    • Steps: 8 – 10 (max)

    • Hi-Res Fix: ❌ I don't use it, but you can if you want.

    • ADetailer: ❌ I don't use it, but you can if you want.


    💬 Feedback and example images are very welcome. Please leave a review and share your results!

    📜 License

    The model follows the official license from the creators, without violating its copyright. You can review the official license here:

    Description

    INT8 Convrot version release.

    FAQ

    Checkpoint
    Krea 2

    Details

    Downloads
    116
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/24/2026
    Updated
    10/27/2026
    Deleted
    -

    Files

    k2ARSA_int8ConvScaled.safetensors

    Mirrors