DragDiffusion

This is an unofficial code for DragDiffusion.

We show the DragDiffusion in a proof-of-concept way where we present the clean structured code of per-image optimization.

We hope the implementation of the principles helps.

The performances are not comparable with the paper's, and considering the performances, we do not include the GUI version yet.

<img src="assets/demo_case.jpg" width="500" alt="Demo case of Our Implementation"/>

Environment

conda env create -f environment.yml
conda activate diff

How-to

1. Put the image file in the ./finetune_data/ and finetune the SD-v1.5 with LoRA.

   python dreambooth_lora.py --pretrained_model_name_or_path 'runwayml/stable-diffusion-v1-5' --instance_data_dir './finetune_data/' --instance_prompt 'xxy5syt00' --num_train_epochs 200 --checkpointing_steps 200 --output_dir 'lora-200'
   

2. Latent optimization.

   python run_drag.py
   

Acknowledgement

• Developed based on official version of DragGAN, unofficial version of DragGAN, and DIFT.