CivArchive
    Agrat - (OC) Original Character (DEMON) by X - v1.0
    NSFW
    Preview 146729637
    Preview 146729687
    Preview 146729700
    Preview 146729759
    Preview 146729836
    Preview 146729957
    Preview 146730029
    Preview 146730093
    Preview 146730103

    MODEL IS PROVIDED AS-IS, I made this for myself but thought I should share :)

    ======

    Original Character. I chose the name Agrat "Dancing Witch" (according to https://mythopedia.com/name-generator/succubus-names/)

    I wanted to test training the smoke clothing into a model, and this is the result

    NOTE: The horns are very persistent, but the wings and tail are not as bad. You can view the attached images for the prompts I used. The wings are a slight issue sometimes, but when they generate properly they look fantastic. Please check the attached images for the prompts I used to generate each look.

    ======

    Workflow used "KREA2BYENZINOv2"

    reference images generated in COMFYUI using a check point merge:

    "Krea 2 Turbo"

    https://civarchive.com/models/2732656/krea-2-turbo?modelVersionId=3072332

    The following LORAs were used:

    "Realisitc Good Girl - Krea2"

    https://civarchive.com/models/2898581/realisitc-good-girl-krea2?modelVersionId=3277262

    Krea2 TextFusions Refusal-Reduction LoRA (Updated)

    "KreaShiver"

    https://civarchive.com/models/2833010/kreashiver?modelVersionId=3197004

    ======

    LORA was trained in AI-TOOLKIT for the FLUX1-DEV model in 2250 steps using 20 images at 816x1088 resolution.

    Image generation was tested for 816x1088, 1024x1024 and 1440x816, it was acceptable for my purposes.

    ======

    WHAT THIS DOES WELL:

    I use my character LORAs for SFW generation, and for NSFW I use them to polish an image via img2img. I build my prompts to work in both the illustrious and the FLUX models that I use.

    WHAT THIS DOES NOT DO WELL:

    NSFW content. Nudity will still require another LORA, though I did use some NSFW images in training. (obviously try and see if you need an inpainting or img2img touchup)

    Description

    LORA was trained in AI-TOOLKIT for the FLUX1-DEV model in 2250 steps using 20 images at 816x1088 resolution.

    Image generation was tested for 816x1088, 1024x1024 and 1440x816, it was acceptable for my purposes.

    LORA
    Flux.1 D

    Details

    Downloads
    16
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/25/2026
    Updated
    10/26/2026
    Deleted
    -
    Trigger Words:
    agrat

    Files

    agrat_flux.safetensors

    Mirrors

    CivitAI (1 mirrors)