Streamlined Bulk Diffusion Generation

What is this?

In my research, I often found myself re-writing a new notebook every time I wanted to draw a bunch of samples from a new diffusion model.

This is just a little tool I made to streamline the process. Given a prompts file, This repo will take your chosen model and generate a corresponding image for each prompt.

Nothing too special, I just wanted something easy (for me) to use. Hopefully it might be useful for someone else out there.

I've also included a script to easily view your images in a web page.

Setup

Please run the following to create a conda environment and install dependencies:

conda create -n easy-diffusion
conda activate easy-diffusion
conda install pip
pip install torch torchvision
pip install -r requirements.txt

You can then run get_prompts.py to download several popular prompt collections.

Generation

You can generate images and save them in ./generated_images with the following command

python gen.py --model=$MODEL --prompts=$PROMPTS --save_dir=./generated_images

For appropriate values of $MODEL and $PROMPTS, please run

python list_models.py

and

python list_prompts.py

respectively.

Adding --skip_existing will skip over prompts for which there already exists an image in the output directory.

Adding --det_seed will use a deterministic seed (equal to the index of the prompt) for reproducibility.

Web Viewer

To create a simple HTML page to view your images, please run

python web_gen.py --root_dir=./generated_images
python -m http.server -d=./generated_images 1111

The HTTP server will then be viewable on port 1111.

If you are on a local machine, you can navigate to localhost:1111.