README.md

---
license: mit
---

# **FaceNormalSeg-ControlNet Dataset**  🚀 🚀

This is the training dataset for the ControlNet used in the **AnimPortrait3D** pipeline.  

For details about this ControlNet and access to the pretrained models, please visit:  
- **[Project Page](https://onethousandwu.com/animportrait3d.github.io/)**  
- **[Hugging Face Page](https://huggingface.co/onethousand/AnimPortrait3D_controlnet)**  

### **RGB Images**  
For facial RGB images, we use the **FFHQ dataset**. You can download it from **[here](https://github.com/NVlabs/ffhq-dataset)**.  
🔹 *Note: This dataset only provides annotated face normal maps and face segmentation maps; face RGB images are not included.*  


### Details

For **face** data, we utilize the [FFHQ](https://github.com/NVlabs/ffhq-dataset) and [LPFF](https://github.com/oneThousand1000/LPFF-dataset) (a large-pose variant of FFHQ) datasets. The text prompt for each image is extracted by [BLIP](https://huggingface.co/docs/transformers/main/en/model_doc/blip-2).
Using the **3D face reconstruction** method, we estimate normal maps as geometric conditional signals. 
We then apply [Face Parsing](https://github.com/hukenovs/easyportrait) to segment teeth and eye regions. Additionally, [MediaPipe](https://github.com/google-ai-edge/mediapipe) is used to track iris positions, providing further precision in gaze localization. 
 
 
For **eye** data, we first crop the eye regions from the face dataset. To augment the dataset with closed-eye variations, which are rare in in-the-wild portraits, we use [LivePortrait](https://github.com/KwaiVGI/LivePortrait), a portrait animation method, to generate closed-eye variations from the FFHQ dataset. These closed-eye face images are then processed using a similar methodology to extract conditions, and the eye regions are cropped and added to the eye dataset.

 
To construct the **mouth** dataset, we begin by cropping the mouth regions from the face dataset. To augment this dataset with a broader range of open-mouth variations, we incorporate additional images featuring open-mouth expressions sourced from the [NeRSemble](https://tobias-kirschstein.github.io/nersemble/) dataset. These open-mouth face images are processed using a similar methodology to extract conditions, after which their mouth regions are cropped and integrated into the mouth dataset.



### Download

```
huggingface-cli download onethousand/FaceNormalSeg-ControlNet-dataset   --local-dir ./FaceNormalSeg-ControlNet-dataset --repo-type dataset
```


## **📦 Dataset Overview**  

| Region  | Image Count |
|---------|------------|
| Face    | 107,209    |
| Mouth   | 131,758    |
| Eye (left + right) | 214,418    |
| **Total**  | **453,385** |