CivArchive
    Furry cumshot & cum physics for Minimax H3 (REF2VA) - v1.5 (ref2va)
    NSFW

    I created this lora to improve the physics of cum with Minimax H3. Currently, when prompting for cum Minimax H3 will produce white milk-like liquid instead of semen, even with careful prompting.

    This lora, which was trained for the ref2va model of Minimax H3, improves the texture, physics, and overall appearance of cum.

    Suggested LoRA Strength: 0.7

    Like many H3 LoRAs at this early stage, higher strengths may result in some degradation in video quality; this is due to the CFG-distilled nature of H3, as well as my limited dataset size. I took measures to mitigate this, however, including active base preservation and guidance distillation when training the lora. I recommend a strength of 0.5 - 0.85. At strengths higher than 0.7 you may begin to see some visual artifacts with motion.

    Recommend prompting: (see below)

    Minimax H3 is a well-trained model, and has a powerful text encoder, so this lora does not require any special or unusual prompting: just refer to it as cum. However, I do recommend proper prompting techniques to achieve the color, texture, thickness, and volume of the fluid desired. H3 has a tendency to default to a large amount of cum unless otherwise prompted, so you may want to specify more precisely how much cum should be produced. For example, you could try referring to the cum as "a small amount of translucent thin white cum fluid." See the attached videos (upload pending) for some other examples.

    This lora was trained specifically for the ref2va model. You may be able to make it work with flf2va, but I have not yet tested that.


    This is an experiment, and was trained on an RTX 5090 for 2600 steps on a small dataset of furry e621 videos, which I specifically selected for their realistic depiction of cum (i.e. fluid simulation). I made it for myself, but found the results decent enough to share here on CivitAI. I used the AkaneTendo250 fork of musubi-tuner to train this model.. Its intended use case was for improving cumshot scenes for anthro furry characters, but feel free to experiment and try it for other types of videos as well. Your mileage may vary ;)

    Description

    • Expanded dataset

    • Trained for additional 2900 steps
      Like before, I recommend running at 0.7-0.8 strength.

    FAQ

    Comments (1)

    Ponder_StibbonsAug 20, 2026· 6 reactions
    CivitAI

    Hold up. Wait, now just a second. What the hell do you think you're doing? Posting actual training information on an H3 LoRA?? The actual model it was trained on! The method, dataset, all sorts of details? I don't know who you think you are, but that's not how it's done here at the Slop Depot. Please confine your LoRA descriptions to a blank page, 'H3" absolute maximum. Can't have people going around knowing stuff. That's a slippery slope. 3 thumbs down.

    LORA
    MiniMax H3

    Details

    Downloads
    877
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/18/2026
    Updated
    8/28/2026
    Deleted
    -

    Files

    h3_ref2va_furrycumshot_v1-5_5900.safetensors