STAR: A Benchmark for Astronomical Star Fields Super-Resolution

This is the official codebase for the paper:
STAR: A Benchmark for Astronomical Star Fields Super-Resolution

License

This project is licensed under the MIT License.

Highlight

• Sep 19, 2025: STAR is selected as one of the 56 Spotlight by NeurIPS'25 Datasets and Benchmark Tracks! 🎉🎉🎉
• Jul 22, 2025: STAR has been released.

Introduction

The STAR (Super-Resolution for Astronomical Star Fields) dataset is a large-scale benchmark for developing field-level super-resolution models in astronomy. It contains 54,738 flux-consistent image pairs derived from Hubble Space Telescope (HST) high-resolution observations and physically faithful low-resolution counterparts. The dataset addresses three key challenges in astronomical super-resolution:

• Flux Inconsistency: Ensures consistent flux using a flux-preserving data generation pipeline.
• Object-Crop Configuration: Strategically samples patches across diverse celestial regions.
• Data Diversity: Covers dense star clusters, sparse galactic fields, and regions with varying background noise.

The dataset includes x2 and x4 scaling pairs in .npy format, suitable for training and evaluating super-resolution models.

Dataset Structure and Download

• Full Data (Hugging Face):

  • x2 Dataset: data/x2/x2.tar.gz
    • Folders: train_hr_patch/ (HR training patches), train_lr_patch/ (LR training patches), eval_hr_patch/ (HR validation patches), eval_lr_patch/ (LR validation patches), dataload_filename/ (.txt files with HR/LR pair mappings).
  • x4 Dataset: data/x4/x4.tar.gz
    • Same structure as x2, for x4 scaling.
  • Download: Hugging Face - KUOCHENG/STAR

• Sample Data (Hugging Face, for testing):

  • x2 Sample: sampled_data/x2/
    • Contains 500 HR/LR pairs in train_hr_patch/ and train_lr_patch/, 100 pairs in eval_hr_patch/ and eval_lr_patch/ (total ~1200 .npy files).

  Quick Start:

  from datasets import load_dataset
  import numpy as np
  dataset = load_dataset("KUOCHENG/STAR")
  sample = dataset['train'][0]
  hr_path = sample['hr_path']  # Path to HR .npy file
  lr_path = sample['lr_path']  # Path to LR .npy file
  
  hr_data = np.load(hr_path, allow_pickle=True).item()
  lr_data = np.load(lr_path, allow_pickle=True).item()

• Source Data (Optional):

  • Raw HST images used to generate patches.
  • Download: Google Drive

  We will update the content later (how to use this source data)

Download

1. Full Dataset

Download the complete datasets for x2 and x4 scaling from Hugging Face:

• x2: data/x2/x2.tar.gz
• x4: data/x4/x4.tar.gz

Usage:

1. Download the .tar.gz file(s).

2. Extract to your project directory under dataset/:

   tar -xzf x2.tar.gz -C dataset/  # or x4.tar.gz

   The extracted structure will be (x2 for example):

   dataset/x2/
   ├── train_hr_patch/
   ├── train_lr_patch/
   ├── eval_hr_patch/
   ├── eval_lr_patch/
   ├── dataload_filename/
   │   ├── train_dataloader.txt
   │   ├── eval_dataloader.txt
   
   

Getting Start

1. Environment config

We recommend using Conda for environment management.

# 1. Create and activate the conda environment
conda create -n star python=3.10 -y
conda activate star

# 2. Install dependencies
pip install -r requirements.txt

# 3. Prepare the Dataset
# Download the dataset from Hugging Face (or your source)
# and unzip it into the `dataset/` directory.

2. Training

All model configurations are located in the configs/models directory.

Single-GPU Training

To train on a single GPU (e.g., GPU 0):

CUDA_VISIBLE_DEVICES=0 bash tools/dist_trainval.sh configs/models/FISR.py --log_dir log/

Multi-GPU Training

To train on multiple GPUs (e.g., GPUs 0, 1, 2, 3):

CUDA_VISIBLE_DEVICES=0,1,2,3 bash tools/dist_trainval.sh configs/models/FISR.py --log_dir log/

Custom Training Options

You can append arguments to the training command to customize the run.

• --use_loss [L1/L2]: Specify the loss function to use (L1 or L2).
• --use_attention: Enable the flux loss

3. Testing & Evaluation

Evaluate your model using a saved checkpoint.

• -e or --evaluate: Switches the script to evaluation mode.
• --resume 'checkpoint_path': Specifies the path to the checkpoint file to load.
• -v or --visualize: (Optional) Enables visualization output during testing. The output path can be configured in the model's config file.

Example:

# Run evaluation and save visualization results
CUDA_VISIBLE_DEVICES=0 bash tools/dist_trainval.sh configs/models/FISR.py \
    -e \
    --resume 'path/to/your/checkpoint.pth' \
    -v

Citation

If you use the STAR Dataset, please cite:

@article{wu2025star,
  title={STAR: A Benchmark for Astronomical Star Fields Super-Resolution},
  author={Wu, Kuo-Cheng and Zhuang, Guohang and Huang, Jinyang and Zhang, Xiang and Ouyang, Wanli and Lu, Yan},
  journal={arXiv preprint arXiv:2507.16385},
  year={2025},
  url={https://arxiv.org/abs/2507.16385}
}

Contact

For issues or questions, open a GitHub issue or send me an email [12guocheng@gmail.com] for free.