CivArchive
    Preview 286135
    Preview 286134
    Preview 286133
    Preview 286132
    Preview 286131

    Donate Coffee for Gtonero

    Not recommended for realistic models

    Main tag : Milking machine,
    Secondary tag : nipples, lactation,
    Pose tag : leaning forward,

    norm retention: 98.14% | std: 0.010
    Training Powershell info :

    • Image Train : 160

    • DIM/Alpha : 12/6

    • Train batch size : 2

    • resolution : 768

    • mix/save_precision = BF16

    • optimizer_type = AdamW

    • GPU : RTX 4060 16GB 160W [Training Load 110W]

    • Torch 2.5 + cu12.1 + Xformers + gradient_checkpointing

    • Vram used : 10.1GB

    • Speed : 1.58s/it

    Description

    Add parameter

    • increase weight decay

    • increase text-encoder

    • increase step

    • reduce restarts to 2

    • reduce DIM 16 / 8 Alpha

    FAQ

    Comments (10)

    psieye481Mar 20, 2023
    CivitAI

    Just updated to v1.3 and using grapefruitHentaiModel_grapefruitv41. While the milking machine always shows up, it's rare for it and the nipples to look good. Have you considered using a standardised milking machine in your training images? As in, taking one milking machine asset and manually pasting it on all the nipples of your training set. Maybe do this twice: once for opaque metal and another for glass with milk.

    I've not trained LoRAs myself but I can assist with preparing such images (and lactation beam projectiles).

    Gtonero
    Author
    Mar 20, 2023

    There are some models that are not compatible with LoRA, but grapefruit has been reported to make LoRA perform poorly

    psieye481Mar 20, 2023

    @Gtonero I've tried bloodorangemixNSFW and anything-v3 now. Neither are an improvement for this LoRA. What model do you recommend?

    Gtonero
    Author
    Mar 20, 2023

    @psieye481 kotosmix or AbyssOrangeMix2 or abyss 7th layer

    Gtonero
    Author
    Mar 20, 2023

    @psieye481 block merged and weighted
    Optimal inputs and outputs for LoRA
    I'm not sure what % of input and output values ​​used by Anything v3
    cause LoRA to be incompatible If not training on a specific block merged

    psieye481Mar 20, 2023

    @Gtonero "Block merged and weighted" - are those extensions or command line args for Automatic1111's webui? I don't see any options or settings for this. Without them, AbyssOrangeMix2_NSFW or AbyssOrangeMix3AOM3 don't do it either (though they were so slow to iterate I didn't try many runs).

    I've gotten other LoRAs to work with grapefruit but I suppose they had less detailed concepts to convey. Your lactation LoRA has a higher success rate - the distorted nipples are easier to manually fix.

    btcnumeric111Mar 21, 2023· 4 reactions
    CivitAI

    Hi nice work, you should make a version trained on "hucows dot com" images

    DuanshuiliuApr 5, 2023
    CivitAI

    I could not find upscaler named lollypop😔

    DuanshuiliuApr 15, 2023

    @onepiecefan Thank you!!