CivArchive
    Ari — Krea 2 OC - v1.0
    NSFW
    Preview 145644295
    Preview 145644294
    Preview 145644300
    Preview 145644301
    Preview 145644303
    Preview 145644304

    Ari is a character LoRA for Krea 2, made for casual selfies, portraits and outfit photos.

    The gallery images were generated with Lustify v10 Krea 2 Turbo (INT8 ConvRot), using the selected Ari E20 checkpoint.

    • LoRA strength: 1.0

    • 8 steps · CFG 1

    • Euler / simple, shift 1.15

    • No additional style LoRAs, Turbo adapter or reference-image conditioning.

    Ari is a rank 32 / alpha 32 BF16 LoRA, approximately 235 MB.

    Describe the expression, clothing fit, camera angle, setting and lighting you want. Keep hair and makeup descriptions consistent when making a matching set. Facial details, hands and exact poses can vary between seeds and prompts.


    Hey! Could really use some extra money rn, so I’m offering paid help with ComfyUI and local AI stuff:

    • Fix broken workflows and custom nodes

    • Build custom workflows and small custom nodes

    • Train LoRAs to order and prepare datasets

    • Set up ComfyUI, models and dependencies

    • Optimize speed and VRAM usage where possible

    • Automate batch generation and repetitive tasks

    • Set up upscaling, inpainting and background removal

    DM me with what you need + your GPU. I’ll confirm what I can do and the price before starting 🙂

    Description

    Initial release: rank 32 / alpha 32, BF16, approximately 235 MB.

    Selected epoch 20 after portrait, wider-shot and pose comparisons. Tested with Krea 2 Turbo at strength 1.0, 8 steps, CFG 1, Euler / simple and shift 1.15.

    The six showcase images use Ari alone, without additional style LoRAs.

    FAQ

    Comments (5)

    mqblbhaiOct 14, 2026· 1 reaction
    CivitAI

    Lovely character, but it seems under trained. I suggest you to train for more epochs/steps for further characters and if you have time release v2 for these characters.

    hestiacummings
    Author
    Oct 14, 2026

    I did use epoch 20/20 and I'm still not sure about trigger words as I'm getting mixed results with and without. I'll give it another go at 30-35 and see where it goes.

    hestiacummings
    Author
    Oct 14, 2026· 1 reaction

    Trained again with epoch 18/30 winning by a small margin and losing to v1 20/20 by a small margin in likeness comparison to the source dataset. 6 prompts, 60 seeds, 60 images with very good character consistency across the board and very few outliers on both krea 2 turbo and lustify. I don't know what's happening to you, but i'm willing to bet its your prompts or model and not my lora.

    mqblbhaiOct 14, 2026

    @hestiacummings Thanks for responding, I will check on my end again

    hestiacummings
    Author
    Oct 14, 2026

    @mqblbhai No worries :D Copy my prompt, run a batch of 20 with your usual models and settings and see what's going on, maybe i'm missing something too.

    LORA
    Krea 2

    Details

    Downloads
    297
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/12/2026
    Updated
    11/3/2026
    Deleted
    -

    Files

    Ari.safetensors

    Mirrors

    CivitAI (1 mirrors)