BreastBox - ADetailer breast detection

BreastBox Detection (YOLOv8s)

Description

BreastBox is a custom-trained object detection model YOLOv8s (yolo small) up to v2 designed specifically to improve the workflow of Stable Diffusion users who use ADetailer.

While default models often fail to detect breasts correctly, side views, or with specific clothing, BreastBox aims to provide a tighter, more anatomically correct bounding box, even for fantasy characters. V1 is far from perfect, especially for specific poses and small breasts under clothes, but future versions will be a significant upgrade.

Why use this model?

• ✅ High Precision: Trained on a manually curated dataset of ~200 labeled images (v1) and ~400 (v2).

• ✅ Detection: Capable of detecting breasts in challenging angles: side-boob, lying down, from behind, or partially obscured.

• ✅ Style Agnostic: Primarily trained on 3DCG, but should generalizes a bit to realistic/2D generations.

How to use (ADetailer)

1. Download the .pt file.

2. Place it in your ADetailer models folder: \stable-diffusion-webui\models\ADetailer.

3. In the ADetailer tab, select this model from the dropdown list.

4. Recommended Settings:

   • Detection Confidence: 0.3 to 0.5 (Adjust if it misses).

   • Mask Merge Mode: Merge

Version History

• v1.0 (Experimental) : Initial release. Very solid performance on standard poses and most 2DCG styles. Might struggle with extremely small chests or "lying, on back, upside down" poses where gravity deforms the shape significantly (~200 data samples).

• v2.0 : More versatile with more data for more poses, breast sizes, and camera angles (~400 data samples).

• v3.0 (in progress) : More data on obscured breasts because of hand and smaller breasts (~650 data samples).

Feedback

This is a passion project and my first detection training. If you find edge cases where detection fails, please share your feedback or example images in the comments. It helps me improve the dataset for the next version!