---
license: gpl-3.0
---

<div align="center">

<img src="images/header.webp" width="800px" />

Custom trained models for the soon to be released **Forbidden Vision** ComfyUI custom nodes.  
<a href="https://github.com/luxdelux7/ComfyUI-Forbidden-Vision">GitHub Repository</a>  

<a href="https://ko-fi.com/luxdelux" target="_blank">
  <img src="https://ko-fi.com/img/githubbutton_sm.svg" alt="Support me on Ko-fi">
</a>

</div>

---

## 📦 Models

All models were trained on a custom dataset of **7k+ images** (a mix of SFW/NSFW from Civitai and my own curated images).  
The goal was to cover style diversity across anime and realistic domains.

### 🟢 Face Detection (YOLO fine-tunes)
- **Base:** YOLO face detection models (GPL-3.0).  
- **Variants:**  
  - `ForbiddenVision_facedetect_yolov11s.pt` → YOLO Small fine-tune -> generally more robust and reliable
  - `ForbiddenVision_facedetect_yolov11n.pt` → YOLO Nano fine-tune -> if you need maximum speed and lower vram usage
- **Training:** Fine-tuned on the <a href="https://github.com/YapaLab/yolo-face">YOLOv11 Face base</a>.  
- **Usage:** Lightweight, accurate bounding box detection for faces.  

---

### 🔵 Face Segmentation (EfficientNet-B0 advprop)
- **Base:** timm EfficientNet-B0 (advprop, Apache-2.0).  
- **Training:** Besides manually annotated civitai and custom images, the model used an equal split of <a href="https://github.com/switchablenorms/CelebAMask-HQ">CelebAMask-HQ</a>
- **Usage:** Provides clean masks for downstream processing.  

---

### 🟣 Landmark Rotation (MobileNetV3-Small)
- **Base:** MobileNetV3-Small (Apache-2.0).  
- **Training:** Trained for landmark-based rotation correction.  
- **Usage:** Aligns faces to level eyes for consistent preprocessing.  

---

⚖️ **Licensing Note**  
This repository is distributed under **GPL-3.0** due to the YOLO fine-tunes.  
The segmentation and rotation models are originally Apache-2.0 based, but included here under GPL for consistency.