CivArchive
    MiniMax H3: IMAGE – Polyhedron: Better Hair - v1.0
    Preview 149540343
    Preview 149540342
    Preview 149540347
    Preview 149540349

    TL;DR: The new MiniMax H3 Image model looks very real, but skin, hair and eyes often come out too harsh. A trick from kijai helps on some motifs: the model sees an empty picture during the first sampling steps – black in his node, and my Polyhedron Suite can do it too, with grey. My first LoRA for it is "Better Hair" (ha8ir) – workflows and LoRAs on Patreon.

    -----

    Okay, folks,

    so here it is, the new MiniMax H3 Image model. What a complicated sh… thing.

    In my opinion it sits at exactly the other end of the rainbow: too realistic, too harsh in its textures – skin, hair, eyes – even though the overall output already offers a great deal of realism. A woman drinking her tea at a café window: looks good, but once you zoom in, it's completely overdone.

    And I'm not the only one who noticed. So did kijai, the grey eminence behind everything that's good and important … His KJNodes include an experimental node, "MiniMax H3 Prompt Weight", with a mode called "text with empty image": for the first part of the sampling, the prompt is fed together with an empty picture – black in his node, plain mouse grey in my workflows at the moment. With the default setting that covers the first two of the model's six native steps. Why? The first steps lay down the main shapes, and I think that's where this distilled six-step image model overreacts the most … On some motifs this gives you a noticeably more balanced final image – on others it makes things harsher, so it's worth an A/B test. Even without LoRAs.

    https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack

    Over the past few days I've trained my usual LoRAs for this image model, and here are the results. :) First up, as before, I'm sharing "Better Hair" (ha8ir), because it interferes the least with the face or the subject while noticeably improving the hair structure. To be clear: the pictures here were made with my full stack – ha8ir together with five other LoRAs, one of them a third-party LoRA – so not everything you see comes from ha8ir alone.

    My workflows and the companion LoRAs, as always, on my Patreon page:

    https://www.patreon.com/c/polyhedron_ai

    Peace! ☮

    Frank

    Description

    V1.o

    FAQ

    LORA
    MiniMax H3 Image

    Details

    Downloads
    187
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/19/2026
    Updated
    11/20/2026
    Deleted
    -
    Trigger Words:
    ha8ir
    perfect hair
    thick hair
    airy volume
    natural waves
    soft flyaways
    hair subsurface scattering

    Files

    polyhedron_minimax_h3_img_image_lora__hair_ha8ir.safetensors