🖌️ Stylus: Automatic Adapter Selection for Diffusion Models

🌎 Overview

Stylus automatically retrieves and composes relevant adapters based on prompts' keywords, generating beautiful and creative images that are tailor-made for each user.

🔧 Setup

We recommend using Python version >=3.10. To install Stylus, run:

git clone --recursive https://github.com/stylus-diffusion/stylus.git
cd stylus
pip install -r requirements.txt
pip install -e .

To download the neccessary checkpoints, embeddings for Stable Diffusion WebUI and StylusDocs (our dataset), run:

./setup/download.sh

Finally, Stylus requires API keys from OpenAI (link), Civit AI (link). Place the keys in configs/keys.yaml and Stylus will automatically manage the env variables for you.

Alternatively, such env variables can also be added to ~/.bashrc or ~/.zshrc.

echo 'export CIVIT_API_KEY="[CIVIT_KEY]"' >> ~/.bashrc
echo 'export OPENAI_API_KEY="[OPENAI_KEY]"' >> ~/.bashrc

NOTE: Stylus also supports Google's Gemini 1.5 for Composer component and Pinecone DB for the Retriever component. Login to your GCP cloud account (gcloud init) to access Gemini 1.5 and add the Pinecone key to configs/key.yaml for such options to work.

🚀 Launch Stylus

Stylus can be launched in two simple steps. The first step is launching Stable Diffusion Web UI (link).

First, to launch SD Web UI, run:

# NOTE: We highly recommend running this in a separate Python environment, due to conflicting dependencies with Stylus.
conda create -n sd python=3.10
conda activate sd
# Launches N replicas of SD Web UI on ports 7860, ..., 7860 + N-1
./setup/launch_sd.sh [N]

Second, Stylus can be run in two ways:

# For a single prompt. See `configs/default_config.yaml` for default options.
python single_prompt.py --prompt "[PROMPT]" --config [CONFIG_PATH]

# For a batch of prompts, we require them to be in a CSV file (see `datasets/`)
python batch_prompts.py --config [CONFIG_PATH]

Both programs will output images generated by Stylus (lora) and the base checkpoint (normal) in the directory: [OUTPUT_FOLDER]/cfg_[CFG]/{lora/normal}/[PROMPT]/... See examples of Stylus generated images in stylus_examples/.

📝 Miscellaneous

Creating StylusDocs (Refiner)

We provide a step by step tutorial for re-creating StyusDocs in stylus/refiner/README.md.

Experiments

See stylus/eval/README.md to run and plot our experiments in the paper. The plotting code and corresponding figures are located in plots/.

🎯 Citation

@misc{luo2024stylus,
      title={Stylus: Automatic Adapter Selection for Diffusion Models}, 
      author={Michael Luo and Justin Wong and Brandon Trabucco and Yanping Huang and Joseph E. Gonzalez and Zhifeng Chen and Ruslan Salakhutdinov and Ion Stoica},
      year={2024},
      eprint={2404.18928},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}