[astrange]

You can do better than a general image-to-text model reading memes, because they all use the same fonts - so you want something trained off synthetic data made with that font.