CivArchive
    Preview 148782318
    Preview 148782315
    Preview 148782314
    Preview 148782317
    Preview 148782319
    Preview 148782321
    Preview 148782323
    Preview 148782322
    Preview 148782324
    Preview 148782325
    Preview 148782327
    Preview 148782328
    Preview 148782331
    Preview 148782332
    Preview 148782335
    Preview 148782336
    Preview 148782342
    Preview 148782345
    Preview 148782350
    Preview 148782351

    KREA-V3:

    This is the result of much trial and error during the training process, using 370 carefully selected, high-quality, and highly diverse images. Each image was captioned—and subsequently rewritten by hand, multiple times, by me personally—because none of the automated captioning neural networks I tested were able to provide sufficiently accurate descriptions.

    Write your prompts correctly, and you’ll get exactly what you want. If you just copy some junk prompt you found somewhere on Civitai, you’ll end up with junk—which only makes sense.

    As time permits, I will share examples of all the concepts that the model V3 has successfully learned.

    MOSS-V0.1:

    I’m getting flashbacks to the early Flux models—it’s an interesting experience, but either I lack the necessary expertise, or Moss is just like Flux, and that’s something that can’t be fixed. I’ll keep an eye on what the community comes up with. For now, I prefer Krea 2 and will continue working on that.

    KREA-V2:

    The styles are just as relevant as they were in V1.

    Precise breast sizes have been reduced to basic knowledge.

    Anatomy and prompt adherence in complex scenarios have been improved (examples attached in the preview).

    P.S. Before moving on to KREA-V3, I need to thoroughly examine the information the model has acquired. I will publish my findings in the future.

    __________________________________________________________________________________________________

    All the images in the preview were generated using the same seed.

    KREA-V0.1 can be viewed simply as a decent model.

    KREA-V1 is a completely different story, utilizing a different training approach.

    You can see the model's capabilities in the preview. These are distinct styles that can be combined; I’ve also included lighting presets in the STYLES.txt that can be mixed with the styles.

    The generation quality in BF16 and INT8 formats differs hardly at all.

    I don't think there will be a future version, as I’m afraid of breaking what has already been achieved. You can see for yourself the wide range and high quality of images the model can produce. For now, I’ll just keep releasing LoRAs.

    I never managed to fully resolve issues with certain micro-concepts—such as distinguishing between vaginal and anal penetration—nor did I quite succeed in reliably controlling spatial positioning (i.e., which side faces what).

    Description

    FAQ

    Checkpoint
    Krea 2

    Details

    Downloads
    13
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/12/2026
    Updated
    11/20/2026
    Deleted
    -

    Files

    EXERCITUS_V3_INT8.safetensors

    Mirrors