We introduce a structure-based generative model for sampling protein conformations efficiently and eliminate the need for pre-training and evolutionary information.