Home

Awesome

Cleaned Face Datasets

TL;DR: The "cleaned" version of two popular face datasets, CelebAHQ and FFHQ, made by removing instances with extreme poses, occlusions, blurriness, or the presence of multiple individuals in the frame.

The CelebA-HQ and Flickr-Faces-HQ (FFHQ) datasets are popular face datasets, each consisting of 30k and 70k high-quality face images respectively. These images display a significant variation in terms of age, ethnicity, facial shape, expressions, poses, and lighting conditions. However, these datasets also contain instances with extreme poses, occlusions, blurriness, or the presence of multiple individuals in the frame. The removal of such images can contribute to enhancing the quality of the generated images. Below shows some examples of the those images.

<p align="center"> <img src="docs/fail_data_celeba.png"> </p> <p align="center"> <img src="docs/fail_data_ffhq.png"> </p>

We have utilized a deep 3D face reconstruction model, Deep3DFaceRecon, to identify these challenging cases. Certain images, where the model fitting failed to produce coefficients, were considered as hard cases. These images were subsequently excluded from the dataset, leaving us with 29,437 images from CelebA-HQ and 69,100 images from FFHQ. The list of images that were excluded can be found within the docs folder of this repository.

We will provide widely used models of GANs and diffusion models, trained on the two "cleaned" datasets, accompanied by a quantitative comparison, to facillate further research.

3DMM Coefficients and Rendered images

<p align="center"> <img src="docs/data_sample_ffhq.png"> </p>

For the required 3DMM coefficients and rendered images of each image in the CelebA-HQ and FFHQ datasets, please refer to the following resources. You can download the corresponding images from the original sources of CelebA-HQ and FFHQ datasets.

PathSizeFilesFormatDescription
CelebA-HQMain folder
├  coeff115MB29,437MAT3DMM coefficients of each image in CelebA-HQ
├  render834MB29,437PNGRendered image of each image in CelebA-HQ of size 256×256
FFHQMain folder
├  coeff270MB69,100MAT3DMM coefficients of each image in FFHQ
├  render1.9GB69,100PNGRendered image of each image in FFHQ of size 256×256

The parameters and rendered images are obtained using Deep3DFaceRecon. Data samples are provided here. For more details about the data format, please refer to their repository. The mesh files are too large to be uploaded. We provide a script to obtain meshes and rendered images from the provided coefficients.

The obtained datasets can be utilized for various purposes, such as 3D-aware face image synthesis using the 3DMM parameters as inputs, or training a ControlNet with rendered images as inputs.

Credits: references.

Citation

If you use this processed dataset and the pretrained models for your research, please cite our paper:

@inproceedings{xia2023retrofitting,
  title={Retrofitting 2D Latent Diffusion for Controllable Face Image Generation},
  author={Xia, Weihao and Öztireli, Cengiz and Xue, Jing-Hao},
  booktitle={NeurIPS Workshop},
  year={2023},
}