This repository demonstrates the application of the collisional optimal transport method (COT) published in
- Sadr, Mohsen, and Hossein Gorji. "Collision-based Dynamics for Multi-Marginal Optimal Transport.", 2025, arXiv preprint at arXiv:2412.16385.
for calibrating diffusion models used in Generative AI.
Here, we compare our approach against the well-known Denoising Diffusion Probabilistic Models (DDPM)
- Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." Advances in neural information processing systems 33 (2020): 6840-6851, arXiv.2006.11239.
in generating images of commonly known datasets.
As a shocase, let us consider both models only with 10 epochs over a subset of the datasets (10k data points). The trained model is then tested to generate new images post-training. In the case of the Food101:
and CIFAR10:
For training, please see examples in the training/ directory.
For testing the trained model, please see examples in testing/ directory.