This repository contains the Classifier-Free Guidance Diffusion project, developed as part of the Generative Models course taught by Alain Durmus at École Polytechnique. This project explores and implements concepts from the paper "Classifier-Free Diffusion Guidance", which introduces innovative methods to guide diffusion models effectively during sampling.
The primary objective of this project is to deeply understand the Classifier-Free Guidance mechanism to simplify diffusion pipelines and improve the quality of generated results. This approach guides diffusion models without requiring an external classifier, making the process more flexible and efficient.
-
Paper Review:
- Summary and explanation of key concepts.
- Analysis of the theoretical implications for generative modeling.
-
Implementation:
- Development of a guidance mechanism for diffusion models without a classifier.
- Training and evaluation scripts for the model.
-
Experiments:
- Tests to evaluate the quality of samples generated using the guidance mechanism.
- Comparison with other diffusion guidance techniques.
-
Results:
- Visualizations and metrics demonstrating the impact of Classifier-Free Guidance.
- Discussion of the advantages and limitations of this approach.
- Python 3.8+
- PyTorch
- Additional dependencies listed in
requirements.txt
git clone https://github.com/MoutetMaxime/classifier-free-guidance-diffusion.git
cd classifier-free-guidance-diffusion
pip install -r requirements.txt-
To train the model, use the command:
python train.py
-
To generate samples with guidance, use the command:
python sample.py --guidance-scale <value>
- Original paper: Classifier-Free Diffusion Guidance Ho et al., 2021.
Special thanks to Alain Durmus for his guidance throughout the Generative Models course.
Feel free to explore this repository and reach out for any questions or collaborations! 🚀