CivArchive
    Whispering - MM-H3
    NSFW

    Trigger: He/She whisper "Insert dialogue here."

    Minimax H3

    Recommended Lora Strength: 0.7

    Description

    Trained for Minimax H3

    FAQ

    Comments (15)

    brnfd24434343dAug 4, 2026
    CivitAI

    MY N.

    First loras I've seen. Does MM train well!?

    kermitfrog1202
    Author
    Aug 4, 2026

    So far so good... except for genitals for some weird reason. There does not seem to be much, if any, base knowledge there though, so it may just take a lot more steps.

    redcreamsodaAug 5, 2026

    @kermitfrog1202 First, thanks so much for these LoRAs! Second, in the meantime before genital LoRAs get there, FYI that you can actually feed a picture of a penis into the ref model's node and refer to it in the prompt and it knows enough to put it where it needs to go and voila; perfect looking penis with no training. Anyway, cheers!

    kermitfrog1202
    Author
    Aug 5, 2026· 1 reaction

    @redcreamsoda Yes... but you should not have to...

    BuzimuAug 4, 2026· 7 reactions
    CivitAI

    i know its like literally the first but what the hecky 600 megs is soo much for something simple :( I hope these shrink!

    jm112368767Aug 9, 2026· 1 reaction

    Are you using floppy disks? ;)

    zym0xAug 5, 2026
    CivitAI

    Wonder if it's possible to only train audio blocks / layers or something? Would be very useful

    12498657Aug 5, 2026
    CivitAI

    Which LoRA Loader node do you use for this? The ComfyUI i2v and ref2v workflows didn't have anything for injecting this.

    kermitfrog1202
    Author
    Aug 5, 2026

    power lora loader

    FromWoodAug 5, 2026
    CivitAI

    Nice, how did you do the lora on mh3? I tried ai toolkit from ostris but when I try on a mh3 official workflow it is like I used the wrong model

    kermitfrog1202
    Author
    Aug 6, 2026· 1 reaction

    I used ai-toolkit and it took several attempts... but ai-toolkit has been updated since the upload of this model. It works better now.

    FromWoodAug 6, 2026

    Sampling inside ai toolkit wasn't broken after first checkpoint ?

    hboxgames132Aug 10, 2026

    @kermitfrog1202  Hey same question but different. I'm training on aitoolkit just now and it looks like working fine, but I would like to know how to make it focus on audio only for example. Did you have any good settings for that ?

    kermitfrog1202
    Author
    Aug 11, 2026

    @hboxgames132 caption EVERYTHING about the video. You caption what you want to be able to change... What you don't want baked into the model. I've thought about training on blacked out video and captioning the black screen to try to only train audio but I have not tried it yet.

    hboxgames132Aug 11, 2026

    @kermitfrog1202  I'm afraid it trains the black screen in to the lora if you do that xD But I see ! The model is super vercatile so if you actually explain what you want or not it might be applied... makes sense !

    LORA
    MiniMax H3

    Details

    Downloads
    1,834
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/4/2026
    Updated
    8/21/2026
    Deleted
    -