AdjointDPM: Adjoint Sensitivity Method for Gradient Backpropagation of Diffusion Probabilistic Models

arXiv:2307.10711v3 Announce Type: replace-cross
Abstract: Existing customization methods require access to multiple reference examples to align pre-trained diffusion probabilistic models (DPMs) with user-provided concepts. This paper aims to address the challenge of DPM customization when the only available supervision is a differentiable metric defined on the generated contents. Since the sampling procedure of DPMs involves recursive calls to the denoising UNet, na\"ive gradient backpropagation requires storing the intermediate states of all iterations, resulting in extremely high memory consumption. To overcome …

abstract arxiv backpropagation challenge concepts contents cs.ai cs.cv cs.lg customization differentiable diffusion examples generated gradient multiple paper reference sensitivity supervision type