No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models