Diffusion models have been wildly successful for image and text generation, but they rely on regression-based objectives and lack explicit representation regularization. As a result, their internal representations are underutilized or poorly structured. This paper introduces

→ View original post on X — @jiqizhixin