README.md

ControlNet training code for Stable UnCLIP

Adding Conditional Control to Text-to-Image Diffusion Models by Lvmin Zhang and Maneesh Agrawala.

This example is based on the training example in the original ControlNet repository.

Installing the dependencies

• Python>=3.9 and Pytorch>=1.13.1
• xformers 0.0.17
• Other packages in requirements.txt

COCO2017 dataset

COCO2017 dataset with depth and openpose conditions is used for training. Download from the website.

Depth Condition

Generating captions with BLIP2-OPT-2.7b and generating depth with Midas:

python data_preprocess/blip_inference.py

OpenPose Condition

First, obtain the image names with person inside from COCO annotation files person_keypoints_{split}.json. The file name list is stored in person_list_{split}.txt:

python data_preprocess/coco_openpose/list_from_anno.py train2017

Second, extract the open pose image condition. The conditions are saved in {split}_openposefull:

python data_preprocess/coco_openpose/condition_extraction.py train2017

Finally, refine the list, removing the images that cannot detected by openpose. The new file name list is stored in person_list_{split}_new.txt:

python data_preprocess/coco_openpose/list_refine.py train2017

Training

Training the ControlNet for Stable Diffusion V2 UnCLIP model, conditioning on image embedding for the stable diffusion.

• Depth Condition:

accelerate launch train_controlnet_unclip_depth.py --config configs/controlnet-coco-unclip-small-depth.yaml

• OpenPose Condition:

accelerate launch train_controlnet_unclip_pose.py --config configs/controlnet-coco-unclip-small-openposefull.yaml

Citation

If you make use of our work, please cite ControlNet and our paper.

@article{zhang2023adding,
  title={Adding Conditional Control to Text-to-Image Diffusion Models}, 
  author={Lvmin Zhang and Maneesh Agrawala},
  journal={arXiv preprint arXiv:2302.05543},
  year={2023}
}

@article{zhao2023makeaprotagonist,
    title={Make-A-Protagonist: Generic Video Editing with An Ensemble of Experts},
    author={Zhao, Yuyang and Xie, Enze and Hong, Lanqing and Li, Zhenguo and Lee, Gim Hee},
    journal={arXiv preprint arXiv:2305.08850},
    year={2023}
}