Probabilistic Foundation Model Predicts Galaxy Spectra from Broadband Images, Mimicking IFU Capabilities Without IFU Data
This work introduces a multi-modal, probabilistic foundation model that leverages a masked autoencoder architecture to predict high-resolution spectra from broadband images. By incorporating fiber positional and redshift-aware wavelength encodings, the model achieves IFU-like capabilities without direct IFU training data. Trained on DESI survey data, it demonstrates performance comparable to supervised methods when validated against MaNGA IFU observations.