Diffusing Differentiable Representations

Yash Savani, Marc Finzi, J. Zico Kolter

We introduce a novel, training-free method for sampling differentiable representations (diffreps) using pretrained diffusion models. Rather than merely mode-seeking, our method ach…