CivArchive
    Preview 134663365
    Preview 134664083
    Preview 134664305

    HerPOV — First-Person Female POV

    -- Moss Image Base--

    V2.0

    Update: Reduced the data set to 69 very high quality images with significant improvements!

    Prompting Guide: Start with a "first-person POV looking down the subject’s own body" then naming whats in frame (ie breasts, nipples, stomach, belly button, shaved pubic area, and how the legs are posed). Define the bottom of the image (breasts, stomach, belly button), middle of the image (feet on the ground, bent legs), and top of image (background, activity, ect).

    Trained on Ostris AI Toolkit - batch size 4 - lr 0.0002

    v1.0 live, trained on the same data set.

    Needs some more finessing than I'd like. Play around with 0.6-1.0 lora strength.

    Trained on Ostris AI Toolkit - batch size 4 - lr 0.0002 - couple hundred images

    --Krea 2--

    V2 - Her POV

    • Increased diversity of data - (10x input training data)

    • Significant improvement in understanding - (multi person interactions, mirrors, cum)

    Suggested Strength: 0.7-1.3

    (0.7 for more complex backgrounds/actions/more zoomed out)

    (1.3 for more stable, close up view)

    Trigger word: Her POV or POV

    V1

    My first LoRA for Krea-2. Trained to give you a first-person point-of-view from a woman's perspective. The camera is her eyes, looking down at her own body (thighs, lap, hands, etc.).

    Trigger word: herpov

    Recommended strength: 1.0 – 1.5.

    Training details:

    - Trained with ostris/ai-toolkit

    - 3000 steps, lr 0.0001, adamw8bit

    - LoRA rank 32 / alpha 32, bf16

    - Multi-resolution (512 / 768 / 1024), batch size 1, ~40 sample photos

    Description

    FAQ

    Comments (4)

    KolARJun 24, 2026· 1 reaction
    CivitAI

    Thank you, that's awesome LORA.

    Dunc4n1dah0Jun 24, 2026
    CivitAI

    How long it takes to train compared to z-image?

    t2kardyyJun 24, 2026· 2 reactions

    From what I was able to test so far, almost (not quite) 2x the time of z-image training

    RandmeistJul 21, 2026

    @t2kardyy Not sure why author does not use automagic v3 which will make this lora done in 1000-1250 steps without risk of overcooking