stable-diffusion-pytorch

Personal implementation of Stable Diffusion in PyTorch from scratch. Educational and research purposes only.

prompt = "A cat under the snow with blue eyes, covered by snow, highly detailed, realistic, ultra sharp, cinematic, 100mm lens, 8k resolution."

Text to Image

Image to Image

Getting started

1. Create a Conda environment:

   conda create -n sd-env python=3.11
   conda activate sd-env

2. Clone the repository:

   git clone https://github.com/martintmv-git/stable-diffusion-pytorch.git
   cd stable-diffusion-pytorch

3. Install the required packages:

   pip install -r requirements.txt

4. Download weights:

   Download v1-5-pruned-emaonly.ckpt (4.27 GB) from RunwayML on Hugging Face and save it in the /data folder. Feel free to download any other weights you like.

5. Run the demo:

   You can run the demo via demo.py or demo.ipynb notebook by your preference. Both demo scripts are configured to use a MPS GPU device, which can be easily adjusted at the start of the code file.

   # To run the demo script
   python demo.py
   
   # Or open and run the Jupyter notebook
   jupyter notebook demo.ipynb

Dependencies

• PyTorch
• Numpy
• tqdm
• Pillow
• transformers
• lightning
• ipykernel

Happy experimenting! 🚀