StarDiffusion arXiv Hugging Face

Offical coda and dataset for StarDiffusion: Learning to Kindle the Starlight

Structure of this project

├─checkpoint  // the weight files generated during the training process
├─weights  // the weights we provide for quick inference 
├─test_result // Storage of test results generated during testing
├─args_file.py  // Set the parameters needed for training
├─requirements.txt // The packages needed for this project
├─inference.py       // Inference Scripts
└─train.py      // Scripts for  training

Dataset

we construct the first Star Field Image Enhancement Benchmark (SFIEB) that contains 355 real-shot and 854 semi-synthetic star field images, all having the corresponding reference images. You can download the dataset from Hugging Face. Each image has a resolution of 640*640.

Usage

Before using this project, be sure to review the project structure above

Train

First, open args_file.py to set the parameters needed for training, the run train.py The weight files generated during training are saved in the checkpoint file, and the tensorboard files are saved in the logs folder

Inference

Run inference.py. We have provided multi pre-trained checkpoints in the weights folder.

If the weight file downloaded from Github corrupted, please directly download it from Google Drive

Reference

@ARTICLE{Yu2022,
  author={Yuan, Yu and Wu, Jiaqi and Wang, Lindong and Jing, Zhongliang ang Leung, Henry and Pan, Han},
  journal={arXiv}, 
  title={Learning to Kindle the Starlight}, 
  year={2022},
}