Official implementation of the paper Improving Consistency Models with Generator-Induced Coupling (Thibaut Issenhuth, Ludovic Dos Santos, Jean-Yves Franceschi, Alain Rakotomamonjy).
Python libraries: See requirements.txt.
To train a base consistency model on CIFAR-10, run:
python train_consistency.py --cfg cifar10_Base --device 0 --eval_fid 1 --eval_freq 5000
To train a consistency model with 50% generator-induced trajectories on CIFAR-10, run:
python train_consistency.py --cfg cifar10_GenInduced_ema_mix50 --device 0 --eval_fid 1 --eval_freq 5000
The mix_gen_induced_traj parameter in config file handles the percentage of generator-induced trajectories per batch.
If you find the code useful for your research, please consider citing:
@article{issenhuth2024improving,
title={Improving Consistency Models with Generator-Induced Coupling},
author={Issenhuth, Thibaut and Santos, Ludovic Dos and Franceschi, Jean-Yves and Rakotomamonjy, Alain},
journal={arXiv preprint arXiv:2406.09570},
year={2024}
}