Dream Ops

• Dream Ops
  • Quick Start
  • Configurations
  • TODOs
  • References

dream-ops is a containerized photo pipeline to customize Stable Difussion Models. It automatically crops faces in the photos.

dream-ops requires a NVIDIA GPU to run, tested on RTX A5000.

Great thanks to my friend Xinyuan for allowing me to use his photos, you look stunning!

Quick Start

After cloning the repo, you can run the following commands in ./ directory:

1. Build the image and get inside the container:

   • Note: you can mount Stable Diffusion Models to /dreambooth/models in the container, eg. -v "$(dirname "$(pwd)")/models:/models" \, and the checkpoints will be moved there after training is successful.

   # build image
   docker build -t dream-ops:cu11.8 .
   
   # get inside the container.
   docker run --rm -it \
       --hostname `hostname` \
       --gpus all --ipc=host \
       -v ./dreambooth:/dreambooth \
       dream-ops:cu11.8 bash

2. Start pipeline: python3 pipeline.py and follow the instructions to put your images in the right place.

Configurations

In ./config.yml, todo.

TODOs

• Reproduce tutorial results and choose approch
• Customize input image and get acceptable results
• Local training and dockerize
• Dreambooth on any model? -> huggingface
• Face auto crop & auto resize
• [ ]

References

1. https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth
2. https://colab.research.google.com/github/sagiodev/stablediffusion_webui/blob/master/DreamBooth_Stable_Diffusion_SDA.ipynb#scrollTo=K6xoHWSsbcS3