Jian Yang

This paper enhances contrastive learning by introducing the variable similarity of augmented views.

We present FreeControl, a training-free method for structural control in diffusion models that achieves strong semantic alignment with less than 2\% additional inference cost.