Proposing equivariant neural networks for direct application to measurements in inverse problems, overcoming limitations of classical reconstruction methods.