metadata
license: mit
pipeline_tag: unconditional-image-generation
MixFlow: Mixed Source Distributions Improve Rectified Flows
This repository contains the weights for the model presented in the paper MixFlow: Mixed Source Distributions Improve Rectified Flows.
MixFlow is a training strategy for rectified flows that reduces the generative path curvatures and improves sampling efficiency. It trains a flow model on linear mixtures of a fixed unconditional distribution and a distribution conditioned on an arbitrary signal (called $\kappa$-FC), which aligns the source distribution better with the data distribution.
- Paper: https://arxiv.org/abs/2604.09181
- Repository: https://github.com/NazirNayal8/MixFlow
Usage
Inference can be performed using the scripts provided in the official repository. After setting up the environment, you can run inference from a checkpoint as follows:
bash scripts/run_inference.sh cifar10 /path/to/model.ckpt outputs/inference/cifar10
For higher-resolution experiments such as FFHQ or AFHQv2 64x64:
bash scripts/run_inference.sh ffhq_64x64 /path/to/model.ckpt outputs/inference/ffhq \
test.num_samples=10000 \
test.num_inference_timesteps=64
Citation
@inproceedings{
nayal2026mixflow,
title={MixFlow: Mixed Source Distributions Improve Rectified Flows},
author={Nazir Nayal and Christopher Wewer and Jan Eric Lenssen},
booktitle={ICLR 2026 2nd Workshop on Deep Generative Model in Machine Learning: Theory, Principle and Efficacy},
year={2026},
url={https://openreview.net/forum?id=uWktyU3OIJ}
}