Iteratively improve Stable Diffusion outputs using aesthetic scores

This script uses the A1111 webui API, so ensure you have --api in the launch flags.

# clone with submodules
git clone --recurse-submodules https://github.com/feffy380/sd-aesthetic-optimizer
# install dependencies (you may want to do this in a venv)
pip install -r requirements.txt

Copy config.json.template and fill in the fields:

• seed_search_patience: number of txt2img seeds to sample. Best result will be used as the starting point for img2img.
• img2img_patience: after generating this many images with img2img with no score improvement, reduce denoising strength.
• initial_denoising_strength: img2img start with this denoising strength.
• options: passed to /sdapi/v1/options. See webui's /docs page.
• parameters: passed to /sdapi/v1/txt2img and /sdapi/v1/img2img. See webui's /docs page.

# see --help for available options

# default values. same as running with no arguments
# search seeds with txt2img and iterate on the best result with img2img. save current best to "outputs" folder
python aesthetic_optimizer.py --config="config.json" --outdir="outputs"

# skip txt2img phase by providing a starting image
python aesthetic_optimizer.py --init_image="path/to/init_image.png"