XavierJiezou/mmcelebahq

Train ControlNet in Diffusers on MM-CelebA-HQ Dataset

https://github.com/huggingface/diffusers/tree/main/examples/controlnet

Replace "fusing/fill50k" with "XavierJiezou/mmcelebahq":

accelerate launch train_controlnet.py \
 --pretrained_model_name_or_path="stable-diffusion-v1-5/stable-diffusion-v1-5" \
 --output_dir="path to save model" \
 --dataset_name=XavierJiezou/mmcelebahq \ 
 --resolution=512 \
 --learning_rate=1e-5 \
 --train_batch_size=4

Note: Following the same settings with Collaborative Diffusion, the first 27,000 pairs are used for training, and the last 3,000 pairs are used for validation.

Citiation

Download the dataset from official repo: https://github.com/IIGROUP/MM-CelebA-HQ-Dataset.

@inproceedings{xia2021tedigan,
  title={TediGAN: Text-Guided Diverse Face Image Generation and Manipulation},
  author={Xia, Weihao and Yang, Yujiu and Xue, Jing-Hao and Wu, Baoyuan},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}