Breast!

Breast! LoRA is here to restore proportions as the cornerstone for the beautiful and the sacred.

Specification:

• Base Model: SDXL 1.0

• Type: LoRA

• Training Steps: 900 steps, 71 images total, masked and captioned individually by me.

• File Size: 200 MB

Description

We all went through the"bigger is better" era, and we were far from being wrong. The fact is that i felt that we had not yet created a series of LoRAs that would somehow regard proportions between the breast and the feminine body.

Technical

Images are preprocessed using a pose estimation model or a semantic segmentation network to extract keypoints/masks for torso, waist, and breast regions. For each image ( I_i ), computed isolate perceptions for Torso-to-waist, Breast-to-Torso and Waist-to-Breast ratios, that were then represented in a low dimensional folder structure.

The whole idea was to make sure that the U-Net’s cross-attention layers are conditioned on text prompts in such a way that we effectively disentangle anatomical traits from non-useful textual semantics.

This is my first LoRA, and the methodology tries and explore the sparsity in anatomical variation to create interactive and compartmentalized LoRA adapters, effectively acting as a form of mixture-of-experts fine-tuning for diffusion models.

For breast size specifically, I've used "large breasts:yy" tag to teach the diffusion net about sizes, and they pair with a denotative tag about the size, that ranges from large to gigantic:

large (yy=0.5 to 0.7), huge (0.8 to 1.2), massive (1.2 to 1.5), gigantic (1.5 and forth).

Example: ... ((massive breasts, large breast:1.3)) ...

Each one of them having their own softness and size parametric behavior.

What Makes It Special:

• Decoupled Aging: Breasts remain anatomically correct without affecting facial age.

• Face Preservation: Minimal facial distortion even at higher weights (tested up to 1.9).

• Shape Control: Precise slider-like behavior through trigger words.

• Clothing Interaction: Realistic fabric deformation and layering.

• Multi-Style Compatible: Works with illustration, anime, and realistic photography.

Trigger Words System

**Size Control:**
[small, medium,] large (0.5 to 0.7), huge (0.8 to 1.2), massive (1.2 to 1.5), gigantic (1.5 and forth).

**Shape & Form:**
round, full, natural, sagging, tear

**Detail Enhancement:**
backboob, underboob, sideboob,

**Pose/Clothing:**
grabbing, squeezing, hiding

Training Dataset

• Total Images: 71 curated, tagged and masked images.

• Resolution: 1024x1024

• Base model: SDXL 1.0

• AI Training Tool

• Home trained with a RTX 5070.

Usage Instructions

Recommended Settings:

• Weight Range: 0.5-0.8 (optimal), up to 1.2 for extreme results

• CLIP Skip: 2 (for SDXL base models)

• Sampler: Euler A or DPM++ 2M SDE Karras

• HiRes Fix: Compatible at 0.4-0.6 denoising strength

Common Pitfalls:

• Weights above 1.5 cause anatomical distortions

• Avoid combining with [specific incompatible LoRAs]

• For face preservation, use weight ≤0.8

Support:

• ⭐⭐⭐⭐⭐ ratings help development

• Please do not distribute outside CivitAI

• Hit me in DMs for comments or to have a chat :)

• Report issues via CivitAI comments

• Follow for update notifications

Version History

v.1 - 27.12.2025