We propose a scalable and sample-efficient framework for training diffusion samplers by integrating classical sampling methods, suitable for practical applications like molecular conformer generation.