Update README.md

README.md

@@ -3,102 +3,110 @@ license: apache-2.0
 ---
 
 <div align="center">
-
 <img src="images/header.webp" width="800px" />
 
-Custom-trained models for face detection, landmark estimation, and segmentation across realistic, anime, and NSFW content.   
-Made for soon to be released **Forbidden Vision** ComfyUI custom nodes  
-<a href="https://github.com/luxdelux7/ComfyUI-Forbidden-Vision">GitHub Repository</a>  
+Custom-trained models for face detection, landmark estimation, and segmentation across realistic, anime, and NSFW content.
+
+Made for the **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
-
-| Model | Purpose | Architecture |
-|-------|---------|--------------|
-| **Face Detection** | Locate faces in images | YOLOv11-small |
-| **Landmark Detection** | Facial orientation & alignment | MobileNetV4-Conv-Small |
-| **Segmentation** | Precise face masks | GhostNetV2-100 |
+## Dataset
 
-## What Makes These Different?
+All three models share a core mixed-domain dataset specifically curated for diffusion-generated content:
 
-**Trained on what you actually generate.**
+- images from CivitAI diffusion outputs (SDXL, SD1.5, Pony, Illustrious)
+- curated images from Danbooru (mixed anime styles)
+- real photographs from various sources
+- NSFW images from each domain without filtering
 
-Unlike general-purpose face models trained only on real photographs, these models were specifically trained on:
+**Total dataset size:** ~11k manually annotated images
 
-- **Diffusion outputs** from CivitAI (realistic and anime styles)
-- **Danbooru** anime dataset
-- **Real photography** for ground truth accuracy
-- **NSFW content** without artificial restrictions
-
-This mixed training approach ensures reliable performance across:
+This mixed approach ensures the models work reliably across:
 - ✓ Both realistic and anime art styles
 - ✓ Difficult angles, occlusions, and expressions  
 - ✓ Low-quality generations and artifacts
 - ✓ SFW and NSFW content equally
 - ✓ Edge cases that break traditional face detectors
 
+---
+
 ## Model Details
 
 ### Face Detection (YOLOv11-Small)
 
-Primary face detection and bounding box localization.
+**Purpose:** Primary face detection and bounding box localization
 
-**Training Data:**
-- 50k+ images from CivitAI diffusion outputs (SDXL, SD1.5, Pony)
-- 30k+ Danbooru anime images
-- 20k+ real photographs
+**Training Approach:**
+- [TRAINING DETAILS PLACEHOLDER]
+- Trained at 640px resolution (inference should use same resolution)
+- [AUGMENTATION DETAILS PLACEHOLDER]
 
-**Features:**
-- Handles extreme angles and rotations
-- Detects partially occluded faces
-- Works with artistic distortions and stylization
-- Reliable on both photorealistic and anime styles
+**Why YOLOv11-Small instead of nano?**  
+More reliable detection across mixed realistic/anime domains with acceptable speed tradeoff.
+
+---
 
 ### Landmark Detection (MobileNetV4-Conv-Small)
 
-Facial orientation and alignment estimation with 68-point landmark regression.
+**Purpose:** Eyes and mouth landmark detection for face alignment
 
-**Training Data:**
-- 40k+ annotated faces from mixed sources
-- Manual annotation for anime/artistic styles
-- Augmented with rotation and perspective transforms
+**Training Approach:**
+- [TRAINING DETAILS PLACEHOLDER]
+- Focused on eyes and mouth keypoints only
+- [AUGMENTATION DETAILS PLACEHOLDER]
 
-**Features:**
-- Fast orientation detection for auto-rotation
-- Works on both realistic and anime faces
-- Handles partial occlusions and unusual expressions
+**Note:** The auto-rotation feature in the Fixer node is handled by post-processing scripts, not the model itself.
+
+---
 
 ### Segmentation (GhostNetV2-100)
 
-Precise face mask generation including hair and complex boundaries.
+**Purpose:** Precise face mask generation including hair and complex boundaries
 
-**Training Data:**
-- 35k+ faces with pixel-accurate mask annotations
-- Mixed realistic, anime, and artistic styles
-- Includes hair, accessories, and complex boundaries
+**Training Approach:**
+- [TRAINING DETAILS PLACEHOLDER]
+- Pixel-accurate mask annotations
+- [AUGMENTATION DETAILS PLACEHOLDER]
 
 **Features:**
-- Pixel-accurate boundaries including hair
-- Handles complex occlusions (hands, objects)
-- Works with artistic styles and effects
+- Includes hair, face, and neck regions
+- Handles complex occlusions (hands, objects, accessories)
 - Smooth mask edges for seamless blending
 
+---
+
 ## Usage
 
 These models are automatically downloaded and used by the **Fixer** node in ComfyUI Forbidden Vision.
 
+**Manual Usage:**
+
+```python
+# Face Detection (640px inference)
+from ultralytics import YOLO
+model = YOLO('fv_face_detect_yolo11s.pt')
+results = model(image, imgsz=640)
+```
+
+---
+
 ## License
 
-GNU General Public License v3.0
+Apache 2.0
+
+The ComfyUI node pack is GPL-3.0 due to Ultralytics dependencies, but these model weights are released under Apache 2.0.
+
+---
 
 ## Contact
 
-- GitHub Issues: [ComfyUI-Forbidden-Vision](https://github.com/luxdelux7/ComfyUI-Forbidden-Vision/issues)
+- GitHub: [ComfyUI-Forbidden-Vision](https://github.com/luxdelux7/ComfyUI-Forbidden-Vision)
+- Issues: [GitHub Issues](https://github.com/luxdelux7/ComfyUI-Forbidden-Vision/issues)
 - Support: [Ko-fi](https://ko-fi.com/luxdelux)