Conditional Diffusion Models for Generating Images of SDSS-Like Galaxies

We present a novel application of conditional diffusion models for generating synthetic images of galaxies based on their physical properties. Our model, trained on data from the Sloan Digital Sky Survey (SDSS), generates galaxy images conditioned on redshift, stellar mass, star formation rate, and gas-phase metallicity. Notably, the model captures expected astrophysical trends, such as the relationship between metallicity and galaxy color or morphology. However, the generated images disagree with SDSS validation images as measured by Gini coefficients, $M_{20}$ coefficients, and Concentration-Asymmetry-Smoothness statistics, which is consistent with systematic underprediction of diffuse flux. While modern generative models are capable of producing realistic images, applying these models to astrophysics may still prove challenging.