Representation Alignment

A training method whose core assumption is that a pretrained visual encoder already contains rich semantic structure. Intermediate features of the diffusion backbone are made to approximate the encoder's features so as to borrow that semantics and accelerate convergence.