Abstract
Light emission from galaxies exhibit diverse brightness profiles, influencedby factors such as galaxy type, structural features and interactions with othergalaxies. Elliptical galaxies feature more uniform light distributions, whilespiral and irregular galaxies have complex, varied light profiles due to theirstructural heterogeneity and star-forming activity. In addition, galaxies withan active galactic nucleus (AGN) feature intense, concentrated emission fromgas accretion around supermassive black holes, superimposed on regular galacticlight, while quasi-stellar objects (QSO) are the extreme case of the AGNemission dominating the galaxy. The challenge of identifying AGN and QSO hasbeen discussed many times in the literature, often requiring multi-wavelengthobservations. This paper introduces a novel approach to identify AGN and QSOfrom a single image. Diffusion models have been recently developed in themachine-learning literature to generate realistic-looking images of everydayobjects. Utilising the spatial resolving power of the Euclid VIS images, wecreated a diffusion model trained on one million sources, without using anysource pre-selection or labels. The model learns to reconstruct lightdistributions of normal galaxies, since the population is dominated by them. Wecondition the prediction of the central light distribution by masking thecentral few pixels of each source and reconstruct the light according to thediffusion model. We further use this prediction to identify sources thatdeviate from this profile by examining the reconstruction error of the fewcentral pixels regenerated in each source's core. Our approach, solely usingVIS imaging, features high completeness compared to traditional methods of AGNand QSO selection, including optical, near-infrared, mid-infrared, and X-rays.[abridged]