Euclid Masked Autoencoder Demo
Masked Autoencoders (MAEs) are self-supervised learning model trained to reconstruct missing parts of input data. The MAE used here is trained on 3 million images of galaxies from Euclid RR2.
Select an image, upload your own image, or provide an image URL, and see how well the MAE can reconstruct the missing parts of the image.
Adjust the masking ratio to see how it affects the reconstruction quality. The MAE can often do well at 90% masking or more!
The model is available on Hugging Face along with code to apply it locally.
For more details, see the workshop paper: Re-envisioning Euclid Galaxy Morphology: Identifying and Interpreting Features with Sparse Autoencoders, Wu & Walmsley, 2025, NeurIPS ML4Phys workshop.
Click an image to update
The model was trained to work well on images from Euclid, with Galaxy Zoo style preprocessing. These images are shared at euclid.streamlit.app and on Hugging Face.
The model was trained to work well on images from Euclid, with Galaxy Zoo style preprocessing. These images are shared at euclid.streamlit.app. You can get URLs by right-clicking each image.
Walmsley trained the model and Wu ran the sparsity analysis. Additional thanks to Inigo Val Slijepcevic, Micah Bowles, Devina Mohan, Anna Scaife, and Joshua Speagle, for their help and advice. We are grateful to the Euclid Consortium and the European Space Agency for making the data available.