CivArchive

    Elke, a fictional character.

    Trained on 30 Klein 9b generated images, from an initial JIB mix qwen generated image.

    minimax h3

    Trained with ai-toolkit, automagic v3 with Contrastive Guidance and training adapter v1; differential guidance scale = 1; rank 24. ai-toolkit configuration attached as minimaxh3-elke-v01.yaml

    krea 2

    Trained with diffusion-pipe, ProdigyPlusScheduleFree optimizer (https://github.com/tdrussell/diffusion-pipe/pull/483), lokr factor 4, masked loss

    ideogram 4

    v02

    Trained with diffusion-pipe, ProdigyPlusScheduleFree optimizer (https://github.com/tdrussell/diffusion-pipe/pull/483), rank 128, using plain text (non-JSON) captions. config-ideogram4-v02.zip contains training images, mask images, captions, validation images, and diffusion-pipe configuration used.

    v01

    Trained with ai-toolkit, automagic v3 optimizer, LR 0.00001, weight decay 0.00001, rank 32, with plain text (not JSON) captions. Likeness is slightly higher when using plain text during inference instead of JSON, but JSON still works. The config-ideogram4-v01.zip attachment contains the training images, masks, captions, validation images, and the ai-toolkit configuration.

    qwen 2512

    Trained with diffusion-pipe, this is a 128-rank lora trained on qwen image 2512, using the ProdigyPlusScheduleFree optimizer (https://github.com/tdrussell/diffusion-pipe/pull/483), on 30 1440x1440 images, using masked training with backgrounds masked to 10%

    The model was evaluated every 300 steps (10 repeats of each training image) using FaceNet512 to evaluate the face similarity, using the same settings, seed, etc for inference, calculating the average distance to a subset of 10 of the training images.

    The attached training data zip includes: the diffusion pipe config files; the training images, prompts, and masks; the validation images (subset of training images); and the output of MirrorMetrics evaluation of epoch15 vs. the validation set.

    Description

    Trained with ai-toolkit, 9,000 steps on 30 1440x1440 images

    FAQ

    Comments (2)

    xcessionAug 21, 2026
    CivitAI

    is the H3 version a checkpoint from the 9000th step or is it an earlier one? I'm just surprised by the step count for such a small dataset

    jandordoe
    Author
    Aug 21, 2026· 1 reaction

    It's the 9,000 step checkpoint. With both contrastive guidance and the training adapter, training is slowed down quite a bit. 3k steps would have been plenty for a passable likeness, beyond that the gains were very small (and diminishing), but still measurable.

    LORA
    MiniMax H3

    Details

    Downloads
    226
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/20/2026
    Updated
    9/5/2026
    Deleted
    -
    Trigger Words:
    Elke

    Files

    minimaxh3-elke-v01.safetensors

    Mirrors

    HuggingFace (1 mirrors)

    minimaxh3-elke-v01.yaml

    Mirrors

    CivitAI (1 mirrors)