Project workflow: 13 000 star observations were simulated with the WaveDiff PSF simulator. Each star observation is located at a particular FOV position (u_i, v_i), has a random noise level S /N and belongs to a spectral class C_i. 10000 stars are directly used for training the pixel-only classification models, and 1000 stars to test them. 2000 stars are used to optimise the approximate PSF models required to calculate similarity features. The quality of the PSF models is evaluated with the 1000 test stars. Once the train and test similarity features have been computed, the PSF-aware stellar classifier is trained with 10 000 stars and tested with 1000 stars.