We introduce a multi-concept feature description framework which can identify and score polysemantic features.

We propose a training-free framework for sampling by exploiting multiscale dynamics and general diffusion paths.