CivArchive
    hyperbreasts - v6
    NSFW
    Preview 1
    Preview 2
    Preview 3
    Preview 4
    Preview 5
    Preview 6
    Preview 7
    Preview 8
    Preview 9
    Preview 10

    A model for large breasted waifus or semi-realistic characters.

    V6 Changelog 2023/06/03:

    • Considering this was my first and most popular LoRA, I figured it was time to improve on it. Probably the last version too. Hyperfusion is more or less the successor to this.

    • Less harmful to the overall style, but slightly smaller sizes overall, feels like a good balance

    • Wider variety of breast shapes in the dataset

      • Some breast shape tags were added, but hard to say how helpful they are

    • Uses LoRA LoCon, but does not require any additional extensions to run

    • Changed "large breasts" to "big breasts" to match my hyperfusion model tags

    • The model prefers resolutions around 640, but can do up to 768 to a degree

    • Training tags attached under model download section

    V5 Changelog:

    • v5 now uses LoRA!

    • What changed since v4? I trained a breast size classifier to auto label the dataset which resulted in better control over breast size, and made it easier to generate larger sizes.

    • Note that the trigger words in v5 are ordered in the following from smallest size to largest:
      breasts, large breasts, huge breasts, gigantic breasts

    V4 Info:

    the keyword for v4 remains hyperbreasts, if you want better control over the size use v5

    Notes:

    I used this to train my image tagging classifiers for breast sizes
    https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification

    Description

    • Added some more variety in breast shapes to the dataset

    • Less harmful to the overall style than v5

    Training Details:

    • 2k images

    • 16 epocs

    • learning rate 2e-4

    • text encoder LR 5e-5

    • base model Av3

    • clip skip 2

    • random flip

    • tag drop chance 0.1

    • bucketing at 768

    • dim 16

    • alpha 8

    • 150 tokens

    • cosine with restarts

    • training with tags, tags attached next to model download

    • Kohya LoRA LoCon, does not require any additional extensions to use