🪐 GalaxyEmb: Galaxy Embedding Model using SimCLR (GZ SDSS)

GalaxyEmb is a self-supervised learning model trained on the GalaxyZoo SDSS dataset using the SimCLR framework. It maps galaxy images into a compact and meaningful latent space for use in similarity detection, retrieval, and evaluation of generative models.

🔍 Model Details

• Architecture: SimCLR
• Dataset: Galaxy Zoo - SDSS
• Input resolution: 424x424
• Output: N-dimensional normalized embedding vector
• Framework: PyTorch

💡 Intended Uses

This model is designed for:

• Similarity detection between galaxy morphologies
• Image retrieval based on morphological similarity
• Evaluation of conditional generative galaxy models, based on:
  • Consistency (alignment with input condition)
  • Diversity (intra-class variability)
  • Realism (visual and statistical plausibility)
  • The three metrics was proposed by Astolfi et al. (2024)

💻 How to Use

You can load the model by download checkpoint_0050.pth.tar and extract embeddings in get_galaxy_emb.ipynb using your own galaxy images. You should modify the path as your downloaded one and your galaxy image folder:

checkpoint_path = "/checkpoint_0050.pth.tar"

# Your image folder path
image_folder = "../images" 
# Your output folder path
output_folder = "../images_features"

📄 Citation

This embedding tool is release with our work below, used for calculating evaluation metrics for generated images. If you use or reproduce based on that, please cite our work.

@misc{ma2025aidreamunseengalaxies,
      title={Can AI Dream of Unseen Galaxies? Conditional Diffusion Model for Galaxy Morphology Augmentation}, 
      author={Chenrui Ma and Zechang Sun and Tao Jing and Zheng Cai and Yuan-Sen Ting and Song Huang and Mingyu Li},
      year={2025},
      eprint={2506.16233},
      archivePrefix={arXiv},
      primaryClass={astro-ph.GA},
      url={https://arxiv.org/abs/2506.16233}, 
}