Xudong Mao

Treating pretrained vision embeddings as temporary scaffolding to inject and internalize—rather than passive alignment targets—slashes diffusion transformer training budgets by over 43×.