SynCast: Synergizing Contradictions in Precipitation Nowcasting via Diffusion Sequential Preference Optimization

This repository contains the official PyTorch implementation for the paper "SynCast: Synergizing Contradictions in Precipitation Nowcasting via Diffusion Sequential Preference Optimization".

Installation

1. Clone the Repository:

   git clone https://github.com/Dtdtxuky/SynCast.git
   cd SynCast

2. Create and Activate a Conda Environment: We recommend using Python 3.10 or higher.

   conda create -n syncast python=3.8
   conda activate syncast

Training Workflow

The complete training process is divided into three distinct stages. The configuration files (.yaml) for each stage are located in the ./configs directory.

The three stages are:

1. Base Model Pre-training: Training the foundational diffusion model from scratch.
2. DPO Fine-tuning for FAR: Fine-tuning the base model with Direct Preference Optimization (DPO) to reduce the False Alarm Rate (FAR).
3. SPO Fine-tuning for CSI: Further fine-tuning the model from the DPO stage with Sequential Preference Optimization (SPO) to improve the Critical Success Index (CSI).

How to Run Training

The primary training script is train.py. We have provided a convenience script, train.sh, to streamline the execution of each stage.

To run a specific training stage, simply execute train.sh and pass the path to the corresponding configuration file as an argument.

Citation

If you find our work useful in your research, we would appreciate a citation to our paper.

@article{xu2025syncast,
  title={SynCast: Synergizing Contradictions in Precipitation Nowcasting via Diffusion Sequential Preference Optimization},
  author={Xu, Kaiyi and Gong, Junchao and Zhang, Wenlong and Fei, Ben and Bai, Lei and Ouyang, Wanli},
  journal={arXiv preprint arXiv:2510.21847},
  year={2025}
}