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

TL;DR

AdjointDPM introduces an efficient adjoint sensitivity method for gradient backpropagation in diffusion probabilistic models, enabling customization with minimal supervision and reducing memory usage during training.

Contribution

The paper presents AdjointDPM, a novel approach that uses the adjoint sensitivity method and probability-flow ODE reparameterization for memory-efficient gradient computation in DPMs.

Findings

Effective in converting visual effects into text embeddings

Enables fine-tuning for stylization tasks

Facilitates initial noise optimization for adversarial sample generation

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 this issue, we propose a novel method AdjointDPM, which first generates new samples from diffusion models by solving the corresponding probability-flow ODEs. It then uses the adjoint sensitivity method to backpropagate the gradients of the loss to the models' parameters (including conditioning signals, network weights, and initial noises) by solving another augmented ODE. To reduce numerical errors in both the forward generation and gradient backpropagation processes, we further reparameterize the probability-flow ODE and augmented ODE as simple non-stiff ODEs using exponential integration. Finally, we demonstrate the effectiveness of AdjointDPM on three interesting tasks: converting visual effects into identification text embeddings, finetuning DPMs for specific types of stylization, and optimizing initial noise to generate adversarial samples for security auditing.