CivArchive
    Runaway LoRA Training - Deconstructed Version
    Preview 64750384Preview 64746651Preview 64746653Preview 64746655Preview 64746658Preview 64746652Preview 64746654Preview 64750388Preview 64750383Preview 64750382

    Several of my LoRAs have been trained using epoch sample images from the training sessions of earlier LoRAs. In this case a training session got stuck in two simultaneus loops, my browser flipping back and forth between them every couple of minutes, and both restarting at the first epoch every time it got to around the 12th, 13th, or 14th. I had to contact Civitai to get it shut down, but of course saved interesting sample images the whole time.

    Now I've used all those images to train this LoRA, and used Civitai's new auto-tagging feature to caption the images. Who knows what it'll do?

    Enjoy the fruits of error!

    Some of the training images are here:

    Runaway LoRA training images, pt. 1 | Civitai

    Runaway LoRA training images, pt. 2 | Civitai

    Runaway LoRA training images, pt. 3 | Civitai

    Description

    This version was trained on the same images, but I randomly shifted the autotags between different images. At a strength of 1 the effect may be unnoticeable, but put it at 1.25 and all sorts of weird things can happen. At 1.5 you might not recognize what you prompted for!

    LORA
    SDXL 1.0

    Details

    Downloads
    17
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/20/2025
    Updated
    9/27/2025
    Deleted
    -

    Files

    Deconstructed_RLoRAT.safetensors

    Mirrors